Moving sand and measuring effort, Earth Mover’s Distance

EMD takes me back to the early 2010s when I used it as a color perceptual similarity proxy for apparel images. Say you have a tray with a pile of sand shaped like a hill on the left. Next to it, a photo of how you want the sand to look, maybe a smaller hill on the right, or two bumps instead of one.

You hire a tiny construction crew with toy shovels. Their job is to move the sand until the real pile matches the picture. Every time they move a bit of sand, they pay a cost that depends on how far they moved it.

• Move a lot of sand a long way: expensive.

• Move a little sand a short way: cheap.

The Earth Mover’s Distance is simply the minimum total effort they need to reshape pile A into pile B. How hard is it to rearrange this into that. For my apparel project we moved pixels in RGB space.

Mathematicians reuse this idea for probability distributions. Instead of sand grains, you think of “mass” that says how likely something is, for example where people live in a city, or which pixels are bright in an image.

From sand to drifting fireflies, Schrödinger Bridge

Swap sand for glowing fireflies, introducing chaotic movement. You look up at a field at night. In the first long exposure photo, most fireflies are on the left. In the second photo, taken later, most are on the right. You know that in between they mostly float around randomly in a gentle breeze.

So you ask a new question: given that fireflies usually wander in a simple random way, what is the most likely story of how they moved between photo 1 and photo 2?

The Schrödinger Bridge asks for the most natural and likely random motion that could turn distribution A into distribution B while staying as close as possible to normal drifting. You do not get a single “move this grain there” plan, you get a whole movie of how a cloud can realistically morph over time.

Teaching the computer how to model wind, Flow Matching

Flow matching is a new technique, powerfully enhancing modern generative models. Imagine the sky full of tiny droplets floating everywhere at random. You know that it’s possible for wind and water to organize into a cloud that’s shaped like an elephant. To do this, we teach a model the right way for droplets to move. At every point in the sky and at every moment, there is a tiny arrow telling a droplet which way to drift. We learn these arrows by picking a random droplet and a real elephant droplet, drawing a simple path between them, and telling the model what the motion along that path should look like.

Here we are doing something similar to EMD. After seeing many such examples, the model learns how the whole wind should behave. When we start with new random droplets and let them follow the learned wind from start to finish, they drift into place and form a fresh elephant cloud.

Notes

E. Schrödinger, 1931. On the Reversal of the Laws of Nature

L. N. Vaserstein, 1969. Markov processes over denumerable products of spaces, describing large systems of automata

Y. Lipman et al, 2022. Flow Matching for Generative Modeling.

• The Infinity Machine: Demis Hassabis, DeepMind and the Quest for Superintelligence by Sebastian Mallaby

Chapter 15: Fermat for Biology

Two quotes from me about the AlphaFold project:

“We figured, if something is structured like a game, DeepMind can apply reinforcement learning and make progress” (p. 261).

The shift from contact map to a distogram was “like going from a black and white to a full color TV” (p. 268).

• Conversation with Jack Nikodem an (ex-DeepMind) fractional CAIO about agentic coding

• Hugo’s post on simulation and our early work on Alphafold

• Discussions of 2024 Nobel Prize work

• Further mentions of NCBR Scientific Council

• National AI strategy

• Scaling AI in Enterprise

• RL at Scale at IndabaX

• InstaDeep’s impact on African AI

• Image recognition patent

• On emotional component of facial recognition

• New Scientist discusses facial recognition in TV shows.

• Doctoral thesis

About me

I am the Executive Director of Machine Learning Engineering at Generate:Biomedicines, leading teams that combine generative AI, high-performance computing, and experimental automation to design new medicines. Previously, I led ML Engineering at Isomorphic Labs, reporting to the CTO and managing three tech leads across drug discovery, core ML research, and large-scale cloud deployment.

I helped initiate DeepMind’s AlphaFold, whose breakthroughs in protein structure prediction earned the 2024 Nobel Prize in Chemistry for its scientific leads. That project showed how machine learning could close a fifty-year gap in biology and set the path for my career by demonstrating how to translate frontier AI into real biomedical impact.

At GSK.AI, I built the company’s first large-scale knowledge graph learning systems to help scientists identify relationships across biomedical literature. As Head of Machine Learning at InstaDeep, I partnered with BioNTech to design ML systems for personalized mRNA cancer vaccines and led distributed engineering teams across Europe and Africa.

I have been an early employee and key contributor in three AI-first startups:
2014 DeepMind -> Google (400–650M USD)
2019 Cortexica -> Zebra Technologies (undisclosed)
2021 InstaDeep -> BioNTech (500M EUR)

These experiences taught me how to build organizations where deep research and engineering discipline coexist, where abstract models become production systems with measurable scientific and commercial outcomes. My focus is on scalable infrastructure for scientific discovery and on turning conceptual advances into platforms that scientists and partners can actually use.

I have also served on the Scientific Board of IDEAS NCBR, advising Poland’s National Centre for Research and Development on AI strategy and helping emerging ecosystems avoid the barriers I once faced starting from the scientific periphery.

My academic path began in pre-EU Poland, where I earned an M.Sc. in Physics from the University of Warsaw, researching biomedical signal processing and spending time at Vrije Universiteit Amsterdam in medical physics. I later completed a Ph.D. in Neuroscience at Ruhr University Bochum in Germany, working with early deep learning models such as LSTMs and manifold learning systems for memory and face recognition.

That formative journey continues to shape how I think about innovation, policy, and talent. I believe the next breakthroughs will arise from combining human ingenuity, experimental feedback, and AI at scale.

Introduction

It is becoming increasingly common for the articles we read to carry the fingerprints of large language models. Unlike humans, whose accountability is tracked in names, affiliations, and reputations, the accountability of an LLM is more opaque. It is measured less by truthfulness than by continued usage, user satisfaction, and persistent A/B testing on the servers of companies where the model resides.

During development, these models are tested against factual benchmarks. But once they are released and exposed to the variability of human prompting, their responses bend toward user’s preferences. That flexibility can make them useful, but it also risks reinforcing biases, producing confirmatory bias in the interaction itself.

The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it - Alberto Brandolini, 2013. With LLMs, the cost of manufacturing has dropped close to zero. Which makes the question pressing: how do we make verification practical?

Guardrails

One idea is to let LLMs watch one another with a specialized focus. For example, they could check for cognitive biases. When we use language we are imprecise and through - often unintentional - habits of omission, equivocation, exaggeration, presupposition, and metaphor we can skirt by through a chain of reasoning that wouldn't compile under strict scrutiny.

Squirrel! Diversion (a working example)

To test this idea, I used a recent BBC science article on salt consumption and asked an LLM to flag cognitive biases from the Estimation category, as defined in the list of biases on Wikipedia. It identified eight. When pressed to be strict, it reduced the list to three “clear textbook” cases:

Textbook-clear matches

1. “One meta-analysis found a 17% greater risk of cardiovascular disease from consuming an extra 5g of salt per day.”
   ➡️ B2 (Base rate fallacy) – Reports a relative increase without anchoring it in baseline population risk, which can distort perception of the result.

2. “In fact, some scientists are now arguing that a low-salt diet is just as much of a risk factor for developing high blood pressure as high salt consumption.”
   ➡️ B5 (Hard–easy effect) – Frames a highly complex medical debate as a simple equivalence, overestimating our ability to resolve hard problems.

3. “The finding of a sweet spot in the middle is consistent with what you would expect for any essential nutrient… where at high levels you have toxicity and at low levels you have deficiency.”
   ➡️ A2 (Attribute substitution) – Replaces a complex epidemiological relationship with a simplified analogy, making the reasoning easier but less precise.

For transparency, here is the bias-spotting prompt:

Read the estimation section of cognitive biases. For each sentence in the pasted article, highlight any bias you detect. Label it using shorthand (A7, I1, O4, SP5, etc.), explain your reasoning if ambiguous, and remain neutral toward the article itself.

The choice of the article was deliberate: a reputable source, a non-sensitive, non-political and benign topic. Imagine what this would do to a controversial social media thread.

Independent Evaluation

In classical machine learning, a common practice is to hold out a fraction of the training data for validation. These holdout sets test whether a model can generalize to examples it has not seen.

Rather than evaluating whether a generative model produces simple a self-consistent and in-distribution piece of text, we can establish a reusable battery of validation criteria: verifiable facts, widely accepted scientific principles, domain-specific reference data, and carefully curated holdout statements. The text can then be tested against this battery for accuracy and coherence, rather than by direct article evaluation.

Through this indirection the reviewer can ask the language model not to assess the text as a whole, but instead to verify whether each criterion is upheld within it. As with the salt article, I asked the LLM to check whether the piece aligned with a series of “tripwires”: explicit criteria designed to test its consistency against well-known facts. This provides an independent axis of evaluation for the strength of the thesis under review.

A refresher: Achilles gives the tortoise a head-start. To catch it he must first run to the tortoise’s starting point; by then the tortoise has inched ahead, so Achilles must next run to that new point; and so on ad infinitum. Zeno concludes that because there are infinitely many legs of the chase, Achilles can never overtake the tortoise, even though he is faster. There’s a slowing down at the core of this setup. Each of these legs represents a shorter duration - and distances - than the previous one.

Suppose I send you a short video clip in which Achilles overtakes the tortoise. And tell you that it happens sixteen seconds in. As eight seconds pass you decide to tap the magic button: playback 0.5×. So the next four seconds of video take eight in real time. You click playback 0.25×, then 0.125x and so on and so on, every 8 seconds. Never getting to time in the video where Achilles overtakes the tortoise. Welcome to Zeno’s Problem Exists Between Chair and Keyboard. There is no paradox, only user error. If you keep slowing down the clip, you won’t get to the good part.

Radiohead already wrote the footnote: you do it to yourself, you do, and that’s what really hurts.

Human minds struggle with grasping very low probabilities. We’re asked to act on statements like “less than 1 in 100,000,” and our language reaches for metaphors: needle in a haystack, grain of sand on a beach, trying to visualize the improbable.

The foveated bias of visual acuity

Human perception doesn’t work like a camera sensor. From the eye to the cortex, our vision is deeply uneven. We stitch together the visual scene by rapidly moving our eyes, collecting sharp detail from a tiny central patch: the fovea, which spans just 1 to 2 degrees of visual angle. That’s less than one-thousandth of the space in front of us.

Despite its size, that foveated region claims half of our primary visual cortex.

Figure 1: Retinotopic projections into the visual cortex [wiki]

What dominates the physical eye and visual cortex likely dominates the mind’s eye too. So when someone says an event has a one-in-a-million chance and compares it to “hitting a tissue on a football field,” our visual imagination immediately zooms in on the tissue. We fail to perceive the proper ratio. The metaphor breaks.

Reframing the risks

Let’s take a few examples of highly unlikely events:

• Winning a 6/49 lottery: 1 in 14 million

• Winning Powerball: 1 in 292 million

• Being dealt a royal flush on the first hand: 1 in 650,000

• Dying today as a 40-year-old man in Cambridge (either one): 1 in 130,000

• Serious adverse drug reaction: “less than 1 in 500”

Now, let’s ground those numbers with something physical: coin tosses. The logarithm base 2 of a number tells us how many heads in a row would equal that probability. Google log2(14 million), and you get 24. That means your chance of winning the lottery is equivalent to flipping 24 heads in a row.

Here’s a quick table:

By converting abstract numbers into a simple physical experiment we build intuition. We feel the rarity.

I call it the grounding coin toss.

First principle mechanics of aging have always been a topic that I poorly understood and given enormous progress in understanding reflected in scientific literature, there’s little excuse not to try to grok the issue at a molecular level. Over the next three posts I will explore causes of damage:

• primary hallmarks

• body’s response to damage: antagonistic hallmarks

• culmination of damage: integrative hallmarks.

Most of the information here will be reflections and interpretations of a 2023 Cell review paper Hallmarks of aging: An expanding universe - Carlos López-Otín, Maria A. Blasco, Linda Partridge, Manuel Serrano, Guido Kroemer.

Introduction

It never ceases to amaze me how far the modern scientific method has gotten us. Diving into a review paper on a topic one is not an expert on already, typically results in multiple lateral deep-dives into topics explainign a new term or a fascinating side quest.

This time I ended up jotting down specific single atom interactions as they pertain to DNA oxidation, got side-tracked into history of a Sicilian cathedral and did a double take when I found out that blood cells are over 80% of total cell count in a human body.

Cells

A distributed system of microservices

We - humans - are not just vague cellular goo. There’s order to our organism. Neighbouring cells are likely to be the same tissue, form the same organ. Cells are spatially compartmentalised.

Each microservice (organ or cell) has a distinct role, operating independently yet contributing to the system's functionality. Strong API boundaries prevent unintended interactions. Biological barriers maintain order and prevent harmful crosstalk, such as inflammation. When a microservice overreaches or leaks data, it mirrors a failing biological barrier, causing system wide errors or chronic disease.

A mosaic

We’re a mosaic. A region of our body may have a set of cells that replicate from the same point mutation, creating a distinct tessera in a huge 3D mosaic of the human body.

Primary Hallmarks

What does hallmark mean in this aging scenario? It’s an indicator, perhaps a biomarker, perhaps a reflection of a biological process. And its change (or presence) is correlated with aging. It’s likely causal. But aging can also causally influence an indicator. An easy to grasp hallmark of aging is VO2 max for physiological tests.

Review paper lists groups of damage categories and I will briefly go through them before focusing on a more holistic view of cell repair mechanisms. They are genomic instability, telomere attrition and epigenetic alterations.

Genomic Instability

Numbers game

The numbers game of DNA replication inside the cell’s nucleus makes me incredibly uncomfortable. The amount of mutations taking place on a regular basis that are caught by a series of mitigating processes makes aging feel dreadfully inevitable. The length of human DNA is about 3 billion base pairs. A one in a million mutation rate would lead to 3000 mutations in the cell.

For tissues that undergo a lot of environmental stress: esophagus tissue, for twenty year olds a few hundred mutations per cell are observed, going up to 2000 for 75 year olds. So even a tissue strongly exposed to the environment accumulated fewer than one in a million mutations over a human lifespan. Somatic mutant clones colonize the human esophagus with age, Science 2018.

Common mutations

• Point mutation: C→T

• Insertion: C→CCC

• Deletion: CCC→C

• Inversion: GATTACAGATTACA → GATTAAGACTTACA

• Translocation: when sequences on two distinct chromosomes end up on a different chromosome during cell division.

Atomic level damage: oxygenation

A very common type of damage to a DNA base is when it interacts with an oxygenating factors like -OH, O2, H2O2. These are byproducts of cellular metabolism. They are found mostly in mitochondria as that’s the cell’s powerhouse and some nasty reactive oxygen species are produced and emitted. A picture of a power plant spewing toxins into the air that land in a residential area. This particular process often leads to mutations as an oxygenated guanine can mispair with adenine (G:C→T:A) during replication.

Atomic level damage: deamination

Guanine’s pair: adenine in turn can undergo deamination, which becomes uracil (a ribose that pairs with adenine again) leading to another G:C→T:A mutation. It can occur simply in contact with water and is more likely in acidic environments and higher temperatures also make the NH2 less stable.

Telomere Attrition

Why aren’t telomeres fully copied during division and replication? DNA polymerases - enzymes synthesising new DNA strands - need a specific starting point: the 3′-hydroxyl group. An -OH group attached to the third carbon atom of a sugar molecule. The 3' hydroxyl group acts as a reactive site, allowing new nucleotides to be added to a growing nucleic acid chain through the formation of a phosphodiester bond.

During DNA replication, the lagging strand cannot be fully synthesized at the chromosome ends due to the absence of a 3'-hydroxyl group for DNA polymerase to extend, leading to progressive chromosome shortening.

During DNA replication, the ends of chromosomes can't be fully copied because the machinery that builds DNA works like a zipper that needs a starter piece of fabric to get going. At the very end, there's no place to attach this starter piece, so a tiny bit of the DNA is left unzipped and lost each time. This problem is solved by an enzyme called telomerase which extends the parental strand with repetitive sequences using its RNA template, allowing the lagging strand to be completed.

See also below in mechanisms of DNA repair 7. TERT.

Epigenetic alterations

Methylation: Introduction

We already saw how individual atoms can cause a guanine and cytosine to be altered and be paired with thymine during replication. This is yet another mechanism by which cytosine is affected. However, cytosine methylation on its own is not mutagenic. A methylated cytosine is less likely to be expressed. So having the ability to silence gene expression via this single chemical process can be desirable.

As a hallmark of aging, as organisms age the overall levels of DNA methylation across the genome decrease, this decrease is called global hypomethylation. As methylation is desirable in non-coding region, awakening them has negative consequences. Reduced methylation in promoter regions of DNA allows RNA polymerases easier access to DNA leading to increased gene expression.

Epigenetic Drift
One of the most intriguing mechanisms influencing the aging process is DNA methylation—a chemical modification where methyl groups attach to DNA and help regulate which genes are turned on or off. As we grow older, the carefully balanced patterns of methylation across our genome begin to shift. Scientists refer to this phenomenon as “epigenetic drift,” and it has profound consequences for cellular function, tissue maintenance, and our overall health. Understanding these changes provides valuable insight into why our bodies become more susceptible to diseases like cancer and neurodegeneration with age.

Global Hypomethylation
A key feature of aging is a decrease in global DNA methylation levels, often termed global hypomethylation. This drop in methylation primarily occurs in repetitive DNA sequences—such as LINEs, SINEs, and satellite DNA—along with various non-coding and intergenic regions of the genome. When these regions lose methylation, two major concerns arise. First, repetitive elements can become activated, potentially causing genomic instability and DNA damage. Second, genes that should remain silent may inadvertently switch on, fueling processes like unchecked cell growth and contributing to diseases such as cancer.

Targeted Hypermethylation
However, not all methylation changes in aging cells are decreases; some regions become more heavily methylated. Two important examples are the hypermethylation of Polycomb group (PcG) targets and tumour suppressor genes. PcG proteins typically control genes critical for development and cell differentiation, so when these regions become overly methylated, cells lose their ability to adapt and regenerate effectively. Meanwhile, hypermethylation of tumour suppressor genes like P16, BRCA1, and MLH1 silences these protective pathways. This leaves cells vulnerable to unregulated growth and mutation accumulation—factors that elevate the risk of cancer and other age-related diseases.

Aging Signature
Collectively, the decline in methylation across the genome, alongside localized increases in methylation at essential regulatory genes, forms a hallmark of aging. Global hypomethylation undermines genomic stability, while localised hypermethylation of critical genes compromises vital protective and regenerative pathways. Together, these alterations contribute to the broader epigenetic drift that disrupts normal gene regulation, ultimately driving many of the degenerative changes we associate with getting older.

Chromatin remodeling

What is chromatin? Chromatin is the complex of DNA and proteins that tightly winds and packs DNA strands. It’s not as tightly packed as when DNA is in a “chromosome” shape. The chromosome shape protects the insides of the DNA strands from transcription. The looser chromatin state makes DNA accessible for transcription, replication and repair. Visualization of DNA level processes will never cease to mesmerize.

Video 1. "Emerging Evidence of Chromosome Folding by Loop Extrusion". Fudenberg G, Abdennur N, Imakaev M, Goloborodko A, Mirny LA (2017). doi:10.1101/sqb.2017.82.034710 PMC 6512960. PMID 29728444.

So what’s with remodelling? Chromatin in the nucleus is supported by nuclear lamina - the inner wall of the nucleus. Defects in the lamina affect chromatin and by extension genomic stability.

Specific aging diseases, such as Hutchinson-Gilford Progeria Syndrome (HGPS) and Nestor-Guillermo Progeria Syndrome, are caused by mutations in two key proteins related to the nuclear lamina: LMNA and BANF1.

Mechanisms of DNA repair

In the paper a number of DNA repair processes are mentioned that affect genomic stability. Let’s go through them one by one.

1. BER (Base Excision Repair)

ELI5: A missing base on a double strand DNA. It has an abasic site. An enzyme sees what base is missing and fills it in.

Base Excision Repair (BER) is a precise mechanism that fixes small, non-helix-distorting damage to DNA bases, often caused by oxidation, alkylation, or deamination. It begins with a DNA glycosylase enzyme recognizing and removing the damaged base, leaving behind an abasic site. An enzyme called AP endonuclease then cleaves the DNA backbone at the abasic site. Next, DNA polymerase fills in the missing nucleotide using the undamaged complementary strand as a template, and DNA ligase seals the break to complete the repair. BER is crucial because these types of small damage events happen frequently due to normal cellular metabolism and environmental factors. Without BER, these small errors could accumulate, leading to mutations or genome instability. BER ensures that the DNA sequence is restored to its original, undamaged state, maintaining the fidelity of the genetic code and preventing diseases like cancer.

2. HR (Homologous Recombination)

ELI5: When there’s a break in both strands then we need a template to see which pair is missing. During cell division when all chromosomes are duplicated, it’s easy to do that HR based recombination.

Homologous Recombination (HR) is a high-fidelity DNA repair pathway that fixes double-strand breaks (DSBs) using an undamaged sister chromatid as a template. It is most active in the S and G2 phases of the cell cycle when a sister chromatid is available. The process begins with the recognition of the DSB and trimming of the broken ends by nucleases to create single-stranded DNA (ssDNA). These ssDNA regions are coated with RAD51, which facilitates the search for a homologous sequence on the sister chromatid. Once the homologous region is located, the damaged DNA strand uses it as a template to accurately repair the break through DNA synthesis and ligation. HR is crucial for maintaining genome stability because it ensures error-free repair. Defects in HR, such as in the BRCA1/2 genes, can lead to cancer, particularly breast and ovarian cancers.

3. NER (Nucleotide Excision Repair)

ELI5: Sometimes DNA “clumps”. Something as common as UV radiation can hit two adjacent T bases and cause them to form a covalent bond. The regular bond between thymine and adenine is a hydrogen bond. Hydrogen bonds work best along straight lines and are generally 10-20x weaker than covalent bonds. This weakness is a good thing. You don’t want your DNA to be difficult to unwind.

Read a detailed study of UV caused mutations. Formation of cyclobutane pyrimidine dimers at dipyrimidines containing 5-hydroxymethylcytosine Photochem Photobiol Sci, 2014.

Nucleotide Excision Repair (NER) removes bulky DNA lesions that distort the double helix, such as those caused by UV radiation (e.g., thymine dimers) or chemical damage. NER works by excising a short segment of the damaged DNA strand, followed by repair synthesis. The process starts when specific proteins, like XPC, recognize the distortion. Next, a multi-protein complex, including XPB and XPD helicases, unwinds the DNA around the lesion. Endonucleases like XPF and XPG cut the damaged strand on both sides of the lesion, removing a fragment of 24–32 nucleotides. DNA polymerase fills the gap using the undamaged strand as a template, and DNA ligase seals it. NER is essential for preventing mutations and diseases like xeroderma pigmentosum (XP), where individuals are hypersensitive to UV light due to defective NER, leading to high rates of skin cancer.

4. NHEJ (Non-Homologous End Joining)

ELI5: When there’s no template but both strands are severed, the unequal dangly bits are chopped off and DNA is simply glued together. It’s error prone of course. But presumably a win over a total seizure of replication.

Non-Homologous End Joining (NHEJ) is a rapid, error-prone repair mechanism that fixes double-strand breaks (DSBs) without requiring a homologous template. It operates throughout the cell cycle but is especially critical during the G1 phase, when a sister chromatid is unavailable. In NHEJ, the broken DNA ends are recognized and bound by the Ku70/Ku80 protein complex, which recruits a DNA-PKcs kinase to stabilize the ends. If the ends are not perfectly compatible, nucleases may trim them, and polymerases can add random nucleotides to allow ligation. Finally, DNA ligase IV, along with its cofactor XRCC4, seals the break. While NHEJ is efficient, its error-prone nature can lead to small insertions or deletions (indels) at the repair site. Despite this, NHEJ is essential for maintaining genome integrity in non-dividing cells and plays a key role in immune system development through V(D)J recombination.

5. MMR (Mismatch Repair)

ELI5: Post replication stage repair of errors of insertion or deletion. Again, based on the paired base (C-G, A-T).

Mismatch Repair (MMR) is a post-replication repair system that corrects base-pair mismatches and small insertions or deletions introduced during DNA replication. These errors occur when DNA polymerase incorporates the wrong nucleotide or slips on repetitive sequences. The process begins with MutS proteins recognizing the mismatch and binding to it. Then, MutL scans the DNA to locate a nick on the newly synthesized strand, marking it as the one to repair. Exonucleases remove the mismatched section, and DNA polymerase fills the gap using the parent strand as a template. DNA ligase seals the repair. MMR is critical for maintaining replication fidelity, and defects in MMR proteins, such as MLH1 and MSH2, are linked to Lynch syndrome, a hereditary cancer predisposition syndrome that increases the risk of colorectal and other cancers.

7. TERT (Telomerase Reverse Transcriptase)

ELI5: TERT is rebuilding and extending the lengths of telomers following division and replication.

Telomerase Reverse Transcriptase (TERT) is an enzyme that maintains the length of telomeres, repetitive DNA sequences at the ends of chromosomes. Telomeres protect chromosome ends from being recognized as DNA breaks. During replication, telomeres shorten because DNA polymerase cannot fully replicate the ends of linear DNA (the "end replication problem"). TERT, the catalytic subunit of telomerase, extends the telomeres by adding repeats (e.g., TTAGGG in humans) using an RNA template within the telomerase complex. Telomerase is active in stem cells, germ cells, and cancer cells but is generally inactive in somatic cells. Telomere shortening contributes to cellular senescence and aging. Reactivating telomerase could theoretically delay aging, but excessive activity is a hallmark of many cancers, highlighting the need for balance in its regulation.

8. TLS (Trans-Lesion Synthesis)

ELI5: TLS is the last line of defense for error correction during copying: when there’s a missing base or the two TTs which created a covalent bond are found, there are additional helpers that inject a random base. There’s no error correction and so it’s just a solution that’s “better” than completely stopping replication.

Trans-Lesion Synthesis (TLS) is a DNA damage tolerance mechanism that allows cells to bypass lesions that block normal DNA replication, such as thymine dimers or abasic sites. When the replication fork stalls at a lesion, specialized low-fidelity polymerases, like Pol eta (η) or Pol kappa (κ), are recruited to synthesize DNA across the damage. These polymerases are error-prone because they lack proofreading activity, so TLS often introduces mutations at the lesion site. However, TLS prevents replication fork collapse, which could lead to double-strand breaks or genome instability. TLS is particularly important for survival under conditions of high DNA damage, such as during UV exposure. Mutations in TLS components, like Pol η, cause diseases such as xeroderma pigmentosum variant (XP-V), which leads to UV sensitivity and skin cancer susceptibility.

RNA therapies

Emerging RNA therapies such as antisense oligonucleotides (ASOs) offer promising treatments of lamina and chromatin remodeling.

In HGPS, the aberrant protein progerin, resulting from a splicing defect in the LMNA gene, disrupts nuclear lamina integrity and accelerates cellular aging. ASO therapies target the defective mRNA responsible for progerin production. By blocking the abnormal splicing or degrading the faulty mRNA, ASOs can reduce progerin levels and alleviate disease symptoms, offering a hopeful strategy for treating HGPS.

Although NGPS results from mutations in BANF1, similar RNA-based therapeutic approaches could be developed to address the underlying molecular defects. For example, ASOs or other RNA therapies might correct splicing errors, restore normal protein function, or mitigate downstream effects of the mutation.

Antisense oligonucleotides (ASOs) are designed to specifically bind to the mRNA responsible for producing progerin through complementary base pairing, similar to the way the two strands of DNA interact.

By attaching to the target mRNA, ASOs prevent the ribosome from reading it, effectively halting the translation of progerin protein.

This approach is classified as RNA-based therapy or RNA-targeted therapy, as it works at the RNA level and is temporary in nature. It is not considered "gene therapy," which involves permanent changes to the DNA.

The second, and I want to make a case that, strictly more complex, is the cultural shift required within the organisation. This aspect deals with people and how they adapt to the new technology. While technical changes might feel routine, culture change represents the real challenge. These two dimensions—technical implementation and cultural adaptation—are interconnected, like two sides of the same coin. On one side (the "heads"), you have digital transformation; on the other side (the "tails"), you have change management. It’s where the greatest obstacles often lie, or, as the saying goes, here be dragons (’ tails).

The “teaching a monkey how to juggle” metaphor comes to mind. To engineers it feels familiar to jump into system design, abstractions and grand plans of “the big refactor”. But that is only, really, ever - the pedestal building. And ironically, by building the pedestal first - we shy away from the painful learning curve of transformation.

Let’s go through five stages of employee grief as “yet another”, “big T” transformation idea takes shape. Below are lessons from “Sooner Safer Happier”, a 2020 book about building productive, innovative teams.

Five stages

Shock and denial

An all-too-frequent pattern is for change to be dictated from the top. Instead of emerging organically, these changes are imposed on employees by those who are convinced they “know better.” This leads to the inevitable chorus of, “Here we go again!” about yet another fleeting initiative. Even worse, employees may perceive these transformations as potential threats to their comfort zones, or worse yet, their job security. And fear is the mind killer.

Anger

A shocked employee, may blame the latest fads, “out of touch” leaders or even exhibit ill will towards colleagues who consider this change to be potentially a good thing. It could lead them to become difficult to work with. Both morale and productivity is likely to take a hit as they withdraw towards their comfort zone.

Bargaining

When the spectre of change is not going away a common pattern is to appeal to company’s uniqueness. “We’re special! Sure it’s a good idea for others, but not for us!” Some aspects of change are beginning to take hold, with plenty of ring-fencing of old - comforting - ways of working. Some plans to sabotage or “work to rule” ideas may surface.

Depression

The most painful stage to many. From a happy employee who was in their stride before the “big T” we may begin to observe a husk of former self. Someone who’s performing their duties in a state of almost “learned helplessness”. I knew how to do things my way, and now with this change I feel like I need to ask for help each time. The uncomfortable state of needing to relearn mastery.

One can only take so much depression. The bigger the transformation the longer each individual stage. This is the biggest risk of “biting off more than one can chew”. It’s in this dark night of depleted morale that the battle is lost.

It’s here where empathetic and supportive leaders need to invest the most energy to remind the team that what’s awaiting on the other side is a place where bigger, better things are possible. It’s always darkest before the dawn. This is the time for all the optimism and courage the leaders can muster.

Acceptance

“…with empowerment, support and coaching; through starting to see the green shoots of better outcomes;” hope arrives. The big impediments to success: allowing learned helplessness to take hold, dogmatic - cargo cult - implementation of change through meaningless rituals, and crucially lack of psychological safety, of trust in the leaders and their vision.

Know yourself and your teams

Quiet conversations

Have a keen, empathetic eye, for your team and understand where each of the members is on the journey. Being able to quietly assess, and privately support them with an individually tailored approach will do wonders for the overall team morale and minimise the dips.

Don’t Teach Everyone at Once

Just as individual team members will react differently to change, some teams, on a larger scale, will be more enthusiastic. These are the "quick wins" that can be fully capitalised upon. Given the limited bandwidth to train teams to become fully self-sufficient, this selective approach is beneficial. Don’t move on to the next team until you’re confident a lasting change has been achieved. It pays to be diligent and ensure that everyone on the team is comfortable before progressing. This serves as the beginning of word-of-mouth marketing, acting as social proof that encourages the next group to take a courageous step toward transformation.

Do Keep Safe

Create a metaphorical staging environment where those who are learning can trip over their skis without the whole world watching. Invite those willing to experiment on-board, and ski alongside them. Make mistakes together. These shared experiences will build trust and pay off in unexpected ways well into the future.

#forces

Last week I tried to understand the mechanisms behind a therapy involving antibodies designed to stop tumour cells from deactivating T-cells via the PD-L1 <> PD-1 handshake. There are many such protein pairs, but this one is the most researched and targeted.

Today I want to better understand mechanisms of occasional immune system’s resistance to those therapies, which is discussed in the second half of the paper Cancer immunotherapy with immune checkpoint inhibitors (ICIs): potential, mechanisms of resistance, and strategies for reinvigorating T cell responsiveness when resistance is acquired.

Some patients exhibit acquired resistance, meaning they initially respond to therapy but eventually, after months or sometimes years, the disease comes back. Others never react to treatment in the first place: primary resistance.

Important term: gene pathway—inside a cell, different proteins interact to affect each other's behaviour or trigger the production of more proteins. A pathway can be turned on or off depending on the cell's needs. The “on” signals are often controlled by proteins called kinases, which turn on signals by adding a phosphate group (P) from ATP. The “off” signals are managed by phosphatases, which remove the phosphate group from proteins to stop the signal. Kinase giveth, phosphatase taketh away.

Battling primary resistance

ICIs here are often used in combination with chemotherapies. A more personalised approach may mean looking at a patient's individual biomarkers to see whether and which type of ICI treatment may be successful:

• Level of expression of—the perennially interesting—PD-L1 is assessed.

• Analysis of the patient's entire exome and screening for the amount of mutations, the so-called tumour mutational burden. The exome is the small 1–2% portion of DNA that encodes actual protein sequences.

• Whether there has already been a successful infiltration of the tumour by the patient’s immune system, specifically lymphocytes. This requires a biopsy of the tumour to search for their presence inside the tumour. Interestingly, staining of different immune cell types is achieved via infusing the sample with dyed antibodies designed specifically to bind to them.

We already learned how protein mutations - associated with cancer - found inside the cell can alert the immune system. Proteins inside the cell are routinely chopped up, and their fragments, peptides 8–11 amino acids long, are exposed on the cell surface within a structure resembling a vice or a grip called MHC: major histocompatibility complex. MHC plus peptide is a pMHC. These peptides are sniffed by the immune system.

The B2M gene (and therefore its protein) is associated with the stability of MHC. B2M mutations result in non-functioning MHC complexes, which prevent T-cells from triggering an alarm. A review paper cites a melanoma study where patients exhibited these B2M mutations, and that patients with this mutation also did not respond to therapy targeting another therapeutically popular inhibitory receptor CTLA-4.

Important term: epigenetic suppression—silencing of gene expression through environmental (non-mutation) factors. This is reversible and can help the tumour avoid immune detection by reducing the expression of proteins that would otherwise trigger an immune response.

Important term: immunological evasion—tumours can achieve this by down regulating immune-related molecules (like MHC building proteins) or by expressing checkpoint molecules (like PD-L1) that inhibit T-cells action.

Interferon gamma IFNG

Next, let’s better understand the role of IFNG in its default setting and imagine how cancer may cause its dysregulation. IFNG is a small protein (~150 amino acids long) which serves as a signalling molecule produced by many immune cells with the goal of activating and regulating an immune response. It has been found to be involved in a vast number of gene expression pathways.

• MHC expression: As mentioned above, the B2M protein is responsible for the stability of the MHC—the tiny vice presenting 8–11 amino acid chains of proteins expressed inside the cell. IFNG stimulates cells to produce more MHC on their surface.

• Directly induces cells to increase production of antiviral proteins.

• Increases the chance that a mutating cell is killed:

  • by CD8+ T cells: by stimulating expression of “kill me” signals on the sick cell’s surface

  • by macrophages, by binding directly to receptors on their surface

IFNG signals via the JAK-STAT pathway (JAK1, JAK2, STAT1 proteins) to build more MHC complexes on the cell surface. If any of those genes undergo a mutation, it disrupts MHC presentation and hence makes cells less visible to the immune system. One can imagine a cell being filled with IFNG, triggering more and more JAK-STAT pathway steps but still no new MHC appearing on the cell surface. Since IFNG is involved in many pathways, it can also lead to upregulation of PD-L1, the protein which “shakes hands” with T-cell’s PD-1, deactivating its attack. Finally, the overabundance of IFNG will lead to T-cell exhaustion, where T-cells don't know what to attack anymore.

Tumour suppressor PTEN

Earlier I talked about kinases and phosphatases as “on” and “off” switches for gene pathways. PTEN is heavily involved in modulating and depressing the cell division rate. When cells need to stop multiplying, PTEN is the one that puts on the brakes. It is also involved in DNA repair by protecting cells from “replication stress”. This replication stress, making inaccurate copies, is of course a source of cancerous mutations. The paper references patients whose genomic profiling found deleterious PTEN mutations coincided with acquired resistance to a PD-1 inhibitor therapy.

Are we there yet?

We’re not there yet with understanding, control and interventions in gene pathways for specific tumours. Attacking something as omnipresent as PD-1 cannot possibly be the limit of our capabilities as researchers and doctors.

Improved analytical tools and experimental techniques can provide clearer insights, leading to better predictive models and personalised therapies. By shining a more informative light on these problems, we move closer to overcoming resistance mechanisms and enhancing patient outcomes in the fight against cancer.

#atoms

Next time we’re looking at something even more complex than immunology. Continual improvement of ways of working for tech teams.

Today I'm exploring a 2023 review paper on cancer immunotherapy with immune checkpoint inhibitors. We're going to examine:

• What immune checkpoint inhibitors (ICIs) are and how they work

• Their effects on reversing the course of metastasised cancers - those where cancerous tissue has spread to secondary locations via the bloodstream

• Innovations in delivering treatments for brain cancers despite the presence of the blood-brain barrier (BBB)

Essential Terms and Concepts

Antibody:

• A protein produced by:

  • B-cells residing in lymph nodes

  • Mucosal tissues in the gut and respiratory tract

• Size: Antibodies are similar in size to enzymes like lactase, which breaks down lactose, and are approximately 2.5 times larger than haemoglobin, the molecule that can carry up to eight atoms of oxygen. Side note on haemoglobin: it's composed of roughly 15,000 atoms to carry eight oxygen atoms.

• Function: Antibodies bind to a specific protein expressed on a cell surface. By binding to it, they change their conformation to expose a receptor that immune cells recognise as a tag to engulf the antibody-tagged cell.

Antigen (mnemonic: "antibody generator"):

• Anything that may trigger an antibody response. In the context of this article, cancerous tissues will present on their surface mutated proteins that are "non-self" - neo antigens. Sometimes they will present "self" proteins that are not mutated but over expressed, meaning there’s more of them on the cell surface as well. This also may trigger an immune response.

Mechanisms of tumour detection

All cells present on their surface small 8 to 11 amino acid long sequences - peptides - for immune surveillance. They come from recycled proteins previously inside the cell’s body. A regular cell will only present sequences that are known to the immune system. Cells that begin to express mutated proteins inside them, or cells infected with foreign viruses causing them to produce viral proteins, will "smell funny" to the immune system by expressing unknown short peptides. Quick maths: there are 10 trillion combinations of 10 long amino acid sequences.

T-cells which patrol the body also express proteins on their surface. Full proteins, not tiny peptides. One highly investigated protein on the T-cell's surface is called PD-1. When a complementary protein, PD-L1 (L for ligand), binds to PD-1, it deactivates the T-cell. In a normal scenario, PD-L1 will be increasingly expressed on healthy cell surfaces following an increase in T-cell populations to downregulate their activity and prevent autoimmunity. That's the inhibitory checkpoint to keep the immune system in balance.

Many cancers express PD-L1 on their cell surface to bind to PD-1 on T-cells, thereby deactivating them. Although PD-1 is one of many proteins that have this function, and cancers may have complementary counterparts, it is sufficient for us to focus on PD-1 to explore antibodies and their function as immune checkpoint inhibitors - sounding the bugle call for another charge against cells that express PD-L1. Some of these happen to be cancer cells.

Fighting metastasis

Localised treatment of tumours is no longer possible once the cancer begins to spread via the bloodstream or other avenues: nerves, lymphatics, neighbouring tissue, and others. Chemotherapy and ICIs are then therapies of choice. Different checkpoint proteins silence T-cells for different cancers. For some, even after metastasis, it is possible to win the war through correct identification and preparation of a matching cocktail of monoclonal antibodies, which enable the immune system to wipe out tumorous tissue.

Blood-Brain Barrier

I must admit that I always thought the blood-brain barrier (BBB) only allowed very small molecules like oxygen, glucose, and certain hormones to pass through. It did not occur to me that it's possible to create large molecules that can pass through the BBB with the aid of helper proteins.

The review article cites two methods of transmitting treatment cargo through it.

• Even with hormones, I was wrong. They are too large to simply diffuse through the membrane. Insulin, leptin, and other substances are transferred via their respective receptors in a process called transcytosis—which literally means "across the cell." In this Trojan horse technique  [link to the paper that the review cites], a drug is coupled to a monoclonal antibody that is directed against one of the BBB's transcytosis receptors allowing passage.

• Apart from hormone passage receptors, transferrin - which serves the iron-hungry brain - can be targeted with a matching antibody delivered inside a small fatty vesicle, a liposome.

Ultrasonic blasts to BBB

In patients with brain cancers (glioblastomas), caused by astrocytes - a type of glial cell - turning cancerous, radical treatments disrupt the blood-brain barrier with ultrasound(!). Quote: "Low-intensity pulsed ultrasound (LIPU) was utilised in a recent Phase I/IIa clinical experiment to damage the BBB and let a medication enter the brain. During the trial, patients had the SonoCloud-1 device implanted into their skull bones so that pulsed sonication could be administered. The research revealed that patients tolerated LIPU well and that carboplatin (a chemotherapeutic) could penetrate the brain after being sonicated."

Job not done

While in some cancers ICI treatment may result in up to half of patients responding, many develop resistance to ICIs often after a few months. Sadly, the precise mechanisms of this resistance are not well understood. It will be explored further next week.

Ontologies and conceptual models 

We are swimming in an ever growing ocean of available data. For specific domains, knowledge experts, computer scientists and philosophers aim to make sense from it, by defining entity types, categories and investigating relations between them. These efforts result in creation of  ontologies: formal representations of a set of concepts and their relationships. Typically an ontology has a well defined scope, applied to a particular domain. 

For every technical domain, people develop language to describe the phenomena within it. They build conceptual models. These conceptual models require “formal semantics”. A large number of words need to have precise specific meanings. And they need to be explained through the lens of “ontological commitment” to the world. That is the relationship between the well defined words should reflect reality. Some of those statements must be “given”. For example: patient X attended hospital Y for treatment Z. Others can then be derived from them via definitions of this particular conceptual model. But facts about the world do not automatically emerge purely from the conceptual model. These facts are known as truthmakers - linking abstract definitions to actual entities in the world.

Ontological unpacking and weak truthmaking

Authors of the paper, advocate for construction of conceptual models, where it’s possible to arrive at “unpacked” ontological commitment which do not merely describe what or how something is but explain why it is. This ontological commitment reflects the alignment between the model and the actual structure of reality, specifying what entities or types must exist for the model's descriptions to hold valid truth.

After distancing themselves from the semantics and ontologies of strictly mathematical and computer science theories of logical sets, the authors begin to enter the waters of quantitative, and numerical truths - going beyond binary truth and false statements. They introduce the concepts of strong and weak truthmaking. Where the weak one is a question of degree, the way of being rather than absolute existence. Authors give an easily understandable example of someone’s height. And assert that allowing for these numerical facts to be weak truthmakers opens up the avenues of contrastivity and counterfactuality in reasoning.

Scaling the knowledge base through semantic interoperability

Authors quote a real life example of a question posed to a knowledge base that requires understanding across multiple semantic domains. 

Which organisations have a contract with a governmental institution and have donated money to the political campaign of any politician who governs that institution or who has a personal relationship with someone who does. 

As a side exercise I resorted to highly capable modern LLMs to see if it is even possible to construct a SQL or SPARQL query to attempt to answer the above question - given reasonable assumptions about the way data is structured:

SQL

FROM organizations o

JOIN contracts c ON o.organization_id = c.organization_id

JOIN institutions i ON c.institution_id = i.institution_id

JOIN politicians p ON i.politician_id = p.politician_id

LEFT JOIN donations d ON o.organization_id = d.organization_id

LEFT JOIN people rp ON p.related_person_id = rp.person_id

LEFT JOIN politicians rel_pol ON d.politician_id = rel_pol.politician_id

WHERE (d.politician_id = i.politician_id 

       OR d.politician_id = rel_pol.politician_id

       OR EXISTS (

           SELECT 1 

           FROM politicians p2

           WHERE p2.related_person_id = rp.person_id 

           AND d.politician_id = p2.politician_id

       )

      );

SPARQL

SELECT DISTINCT ?organizationName

WHERE {

  # Organization has a contract with an institution

  ?organization :hasContractWith ?institution .

  ?organization :name ?organizationName .

  # Politician governs the institution

  ?institution :governedBy ?governingPolitician .

  # Check if the organization has donated to the governing politician

  OPTIONAL {

    ?organization :donatedTo ?governingPolitician .

  }

  # Check if the politician has a personal relationship with another person

  OPTIONAL {

    ?governingPolitician :hasPersonalRelationshipWith ?person .

    ?relatedPolitician :hasPersonalRelationshipWith ?person .

    ?organization :donatedTo ?relatedPolitician .

  }

  # Ensure that at least one of the donation conditions is met

  FILTER(BOUND(?governingPolitician) || BOUND(?relatedPolitician))

}

Authors correctly point out that in order not merely to satisfy computer science requirements, but actually find out the answer behind the meaning of the question one needs to be aware of additional relationships between entities and “how the world works”. For example - properties of parthood are transitive. Which may not result in the graph database query. And “hasPersonalRelationshipWith” is a relationship whose definition is at the essence of the meaning behind the question.To quote in detail:

“Finally, one has to understand a plethora of temporal aspects including being able to answer in which period: (a) campaign donations are made; (b) contracts are established between (potentially donor) companies and organizational units; (b) a certain organizational unit is part of another; (d) people are assigned to play certain roles in organizational units; (e) personal relationships (whatever those mean) hold. Understanding (a-e) requires explaining what makes the relations in this scenario hold between those specific relata when they do, that is, what are the truthmakers of these relational propositions.”

The authors set out a promise of “ontological unpacking” in order to combine truth representations of different knowledge bases. They start with an example of a patient being treated by a healthcare provider.

Doctor, doctor! I’m a patient patient

In order to represent the reality of a specific patient (entity) being treated by a healthcare provider (entity), authors introduce a few relation types: 

• Comparative relations (e.g., one condition being more severe than another) are derived from intrinsic properties of the entities involved (like medical conditions).

• Material relations (e.g., "treated by") require a relator to explain the connection between entities (like treatments connecting patients and providers).

• Phases represent temporary states that individuals can move in and out of without changing their fundamental identity.

Because many patients can be treated by many healthcare providers, therefore it’s important to introduce cardinality constraints. A specific treatment may only apply to 1 patient - even when performed in total by multiple healthcare providers.

Finally the authors note that when it comes to a decision about patient A being treated ahead of patient B, this comes from a numerical value of “condition severity”, similar to earlier discussed height. And the truthmaker of treatment prioritisation is not a relationship between patients but a relative comparison of severity of their condition. Since this is transitive it allows for reasoning and decision making around healthcare triaging.

The authors discuss the use of ontology design patterns as part of the ontological unpacking process. These patterns help systematise how different relations and entities are represented, ensuring that different models can be aligned based on their underlying structures. The paper shows that by recognising these patterns, it is possible to reduce ambiguity in how models represent the world and thereby support interoperability between different systems.

Practical implications

Authors introduce a practical tool for ontological unpacking based on OntoUML, a framework for conceptual modelling that allows to visualise potential instantiations of entity relations. And discover faults in ontology logic that can be addressed by re-reviewing the conceptual model. They observe how the model represented in Figure 2 above creates an unintended but possible interpretation that “in the absence of additional constraints, it also makes clear that the same person can be patient and healthcare provider in the very same treatment which might not be an interpretation intended“.

Conclusion

This paper allowed me to gain a much deeper appreciation of how complex ontologies associated with knowledge bases can be. Perhaps combining modern advanced in LLMs for data ingestion with debugging tools like visualisations shown in paper’s Figure 6 it’s possible to bootstrap to a larger knowledge base combining different sources and conceptual models.

#bits

The Three Roles

TRANSFORMED introduces three roles with familiar names but insists on a nuanced understanding of their responsibilities. These roles function as a unified, core team with distinct ownership areas, yet all are deeply aligned with the product vision. The triad consists of the product manager, product designer, and tech lead. While there is typically one product manager and one designer, the tech lead represents a team of engineers. Ideally, the entire team is invested in the POM, but at least having the tech lead on board is crucial.

The Four Pillars of Ownership

1. Value: How likely is the proposed solution to outperform existing alternatives? Will customers prefer it, or will internal teams advocate for it over third-party options?

2. Viability: Can the solution meet privacy and compliance requirements? For AI-driven initiatives, can a suitable dataset be created, or a model trained within budget constraints?

3. Feasibility: Does the team possess the technical expertise to build the solution?

4. Usability: Is the product intuitive and user-friendly? Does it solve problems efficiently?

Product managers are responsible for value and viability. Product designers handle usability, and tech leads oversee feasibility. So far, this division seems straightforward and uncontroversial.

The Twist

However, this structure does not imply a waterfall process where product management defines features, engineering implements them, and design team adds a visual layer. Instead, the triad collaborates to discover the product together. For more on fostering a collaborative environment, refer to my previous post, “Yes and …” your solution.

Building Blocks of the Product Operating Model

Effortless Push to Production

The tech lead must foster a culture where deploying to production is routine. Robust CI/CD pipelines ensure that pushing to production happens daily. The absolute lowest acceptable frequency is to push at the end of a two-week sprint cycle, and even then the strong expectation is that integration and deployment run automatically and frictionless.

Access to User Data and Feedback

Understanding the impact of changes requires both quantitative metrics and qualitative user feedback. While product designers often focus on optimising usability, engineers also need to have access to data on how their solutions are performing in the real world. This visibility is key to building products that resonate with users. Seeing the product being “held the wrong way” is a source of many engineering epiphanies.

Design is More Than a Pretty Interface

Steve Jobs once said, “Design is not just what it looks like and feels like; design is how it works.” This ethos is central to the POM. The entire triad must focus on building something that works seamlessly. With easy deployment in place, Product designers can create prototypes at minimal cost (around 1% of the final product) and iterate based on user feedback.

The Empowered Engineer

Two quotes from TRANSFORMED underscore the importance of empowering engineers:

1. “Of all the principles in the product model, the single most important is the realisation that innovation absolutely depends on empowered engineers.”

2. “Most tech leads will tell you that spending a few minutes discussing a product idea early on can save weeks or months of damage control later.”

Building an innovative, resilient company requires moving beyond a top-down, waterfall approach where product managers dictate features. While such a model may function in highly regulated industries with substantial moats, companies that treat engineers and designers as mere “implementers” rather than core contributors will inevitably be out-innovated by startups with agile, customer-centric cultures. In the end, product-market fit is what truly matters.

Conclusion

These are just broad brushstrokes. Throughout the book, we see case studies from real companies and go into much greater detail on the behavioural expectations and competencies of each role within the product triad. There's even a dedicated section on how to preempt or resolve objections from those resistant to change. TRANSFORMED offers an inspiring framework for creating teams with clear ownership, shared purpose, and a sense of meaning.

#forces

Next week, we’ll shift gears to discuss data ontologies, exploring how they shape the landscape of modern data science.

The scientific approach is driven by the ability to identify patterns and formulate hypotheses that predict future experimental data. At the smallest scales, we employ the language of quantum physics and chemistry. When imaging ensembles of atoms that form materials, the periodic arrangement, including any defects, is of particular interest. Given the emerging capabilities of machine learning, it is reasonable to investigate whether algorithms capable of approximating complex distributions—modern deep neural networks - might potentially uncover deeper relationships within atomic assemblies.

We are beginning to see approaches that utilise machine learning to model interactions and strain within a graphene structure imaged using scanning transmission electron microscopy. This is reflected in Physics and chemistry from parsimonious representations: image analysis via invariant variational autoencoders, published on 2024-08-14. We will explore how imaging studies of spatially periodic materials with atomic-level defects are being transformed by an algorithm developed just a decade ago: Variational Autoencoders.

Before we get to computational materials a quick - equation free - refresher on how VAEs work. 

A word to the wise …is sufficient

There is something powerful about building systems that can compress or distil a vast amount of information into a small number of data points while still preserving the essence of the story. A recent example from the Olympics demonstrates this. Immense TV audience excitement was generated by observing only three evolving speed values to determine if a swimmer was slowing down or speeding up. The entire complexity of the race was condensed into a few numbers: time and position at the flip turn and distance to leader / speed during the top down view. With a strong mental model of the arena, the swimming pool, the competition, and the prize - an Olympic medal - the audience was able to fully engage in the thrill with no more than ten real-time numbers.

Vivid imagination or veracious reconstruction

We can condense the essence into a few key numbers. But what about a pixel perfect reconstruction of the race? Perhaps swimming, with all its splashes, might be too challenging. But what about Formula 1 racing? With access to all the rigid body metrics of the car: speed, position, steering wheel angle; is it beyond imagination to fully reconstruct the camera view of a race? It can be achieved in computer graphics. Why wouldn’t it be possible to train a neural network to regenerate the image from vast amounts of past examples, given camera parameters and cars data?

Breaking News: game engine generation. While not as challenging as real time video modelling of a race track with all of its weather and audience variations, Google DeepMind still achieved a mind blowing milestone. On 2024-08-27 it published Diffusion Models Are Real-Time Game Engines GameNGen, modelling the Doom game engine at 20 FPS at fidelity making it - in shortish context time windows - indistinguishable from the real thing for human observers. It shows the power of neural networks ability to deeply understand the entire distribution of input data they have seen during training and capacity for pixel perfect reconstruction.

AE

We discussed distilling a complex situation into a few words: encoding. We also talked about reconstructing a complex scene from just a few coordinates: decoding. This can be achieved by utilising a typical image processing neural architecture: a convolutional neural network, capped with a layer that produces a vector - rather than a typical categorical classification decision; hotdog - not hotdog. Consider this vector as the aforementioned racing car coordinates - a vector of latent space. The reconstruction would "simply" be carried out by an inverse architecture to the one that encodes into latent space. It would take the latent space vector as input, providing the context to reconstruct the original image. What you have there is a non-variational autoencoder.

The V in VAE

What’s the variational trick? Rather than asking the encoder network to predict the vector of length N directly we ask it to predict parameters of an N-dimensional gaussian: the mean and the variance of N dimensions. One can be tempted to whole-hog an NxN covariance matrix but that’s only for special cases. 

Now, during training we need to balance two sources of loss

1. Distribution alignment loss

We want the neural network to encode an image in such a way to predict a mean of 0 and variance of 1 for each latent dimension.

2. Reconstruction loss 

We want the parameters of the predicted Gaussians used to generate a latent vector to cause the decoder network to regenerate the input faithfully. Meaning we need expressiveness and diversity which we won’t achieve if we always predict the same mean=0 and variance=1.

This tug of war between predicting distribution parameters close to Gaussian normal and infusing the latent space with expressiveness contributes to learning a model that’s expressive, wise - through compression - and generative.

Why go through all this variational trouble?

Adding the distribution component enables the architecture to be generative and produce new, realistic samples. During the variational training, the neural network setup enforces smooth behaviour of the distribution in latent space, ensuring that the samples produced are realistic. The ability to interpolate in low-dimensional latent space between samples seen at the input allows for never-before-seen, but plausible, new structures to be produced.

Let’s talk materials and microscopes

The authors of the paper first test their VAE architectures on toy datasets: images of handwritten digits and suits of cards that underwent rotation and shear, to assess the robustness of the networks in building a latent space that places two images of the same class closer to each other than two images of different classes. After testing the architecture, they apply it to analyse the 'distilled essence' of the meaning of hexagonally arranged carbon atoms with silicon imperfections, as well as collecting atomic microscopy feedback force when it comes into contact with a thin slice of ferroelectric film to determine the walls of individual ferroelectric domains.

They typically choose very low-dimensional latent spaces. They either pick two dimensions for two latent variables or add a third dimension to introduce an explicit dimension for 2D angular rotation and two dimensions for x and y translation. They also experiment with adding an M-dimensional vector for M classes to be one-hot encoded.

The authors are able to compress the images and derive basic observations about how different elements of the latent space may encode different arrangements of carbon atoms in graphene. They also speculate that a part of the latent space that varies with the image but cannot be easily explained by inspecting them visually may encode strain. I think it’s an early attempt at applying modern ML, and the material scientific conclusions are somewhat underwhelming. That being said, the article serves as a reasonable refresher on VAE architectures.

Future of scientific collaborations

Authors of the paper observe that we live in a world where data, coding researchers and computational power influence each other. As more and more data comes online it is imperative that we move from qualitative inspections to quantitative methods. Especially when inspecting materials where emergent properties not described by typical physics computations is too expensive to be done for every single sample.

And others are not resting. Just this week Google DeepMind published a Science paper on using ML to approximate excited quantum states. But that’s a topic for another time.

#embeddings

Introduction

There are a number of shared elements in the process of generating a protein rich solution to be imaged by X-ray crystallography and CryoEM. But once we have this valuable droplet, how do we prepare it for imaging?

A three-microlitre droplet, less than 2 mm in diameter, is spread onto a grid and then blotted with filter paper to reduce the solution to a thin film of 35 nm thickness. In the previous article, we discussed the apoferritin protein used to calibrate an electron microscope. Apoferritin has a diameter of 2.5 nm, so about a dozen apoferritins stacked on top of each other would equal the thickness of the aqueous film.

The original solution is manufactured at 8 mg/mL density. If proteins were distributed uniformly in the film following the blotting, the expected number of observed proteins would have been more than ten times lower than actually observed. A spherical droplet is converted into a thinly laid-out film, dramatically increasing the air-water interface to which proteins preferentially stick. A strict mechanistic explanation is an area of ongoing investigation, but this phenomenon works in experimenters’ favour [1].

Amorphous Ice

We then proceed to flash-freeze it in liquid ethane at -180°C. It cools in a few milliseconds. This rapid cooling process creates an ice enclosure for target proteins where water molecules do not have time to orient themselves into a well-known hexagonal structure.

Why does this matter? As discussed in the previous article, electrons scattering on a spatially periodic lattice are going to generate a powerful coherent signal. Since our interest lies in imaging the atomic distances between protein atoms, not water atoms, it’s beneficial for the electron scattering against frozen water to occur in random directions, forming a uniform “background radiation.” This amorphous—vitreous—ice is technically glass, which is fitting for our microscopy endeavour.

Amorphous Carbon

Just as structured, periodic lattices for ice enclosing the proteins reduce the signal-to-noise ratio, the same is true of the material on which the frozen solution is deposited: a metal grid covered in carbon. But carbon is… graphite, carbon is diamond, carbon is graphene, carbon is fullerene—very orderly, periodic structures. How does one create carbon that’s:

• Amorphous

• Disorganised

• Three to five nanometres thick

• High purity

• Adhering to the metal mesh

Let’s examine modern manufacturing techniques involving hot plasma and evaporation-based carbon deposition [2]. This publication describes a multitude of possible techniques; however, it concludes that Filtered Cathodic Vacuum Arc (FCVA) is the most efficient method to synthesise coherent thin films. In this process, a carbon target is treated like a cathode. Between it and the anode—separated by a few centimetres—a voltage of around 100V is applied to create an ionised carbon+ plasma. By adding the filtered component, which removes the occasionally ejected pieces of cathode material, deposition onto the substrate can be of high purity and quality.

Caged neurotransmitters

…is an excellent punk band name. They are also key to observing the dynamics of the protein-molecule system. Neurotransmitters—dopamine, serotonin, glutamate, GABA, and ATP—are manufactured in a form that’s bound to another chemical group, which, when hit with a specific wavelength of UV—there’s that quantum physics again—separates and allows the small neurotransmitter to bind to its receptor, i.e., a protein binding pocket.

What does this cage look like? A common compound is a nitrobenzyl group, which efficiently absorbs UV. The molecule arrangement looks like this:

NT - CH2 - Ph - NO2

Where NT is the neurotransmitter and Ph is the hexagon-shaped phenyl group C6H5 (benzene ring minus a hydrogen). Initially, the UV light hits the Ph-NO2 bond; however, this introduces a series of quick steps involving spatial rearrangements and further bond severance, leaving NT in its free form.

From still images to movies

It’s time to introduce caged neurotransmitters into the high concentration of proteins in a thin film solution, adhering to amorphous carbon. Scientists have automated the process of hovering the sample over the liquid ethane, bombarding it with UV to release neurotransmitters from their cages, and then flash-plunging it in a programmable time window as short as 10 ms. By repeating the experiment at different time scales, they can reconstruct the behaviour of a target protein interacting with small molecules.

In one paper [4], they release individual protons (hydrogen ions) to vary the pH of the solution from 8.0 down to 6.0, which alters the resting conformation of a cASIC1a protein’s pocket. By altering the shape, one is able to reconstruct the temporal dynamics of a protein pocket altering its shape with CryoEM snapshots.

Summary

Over the course of three articles we’ve gone through: biological applications of modern CryoEM technology. Physical and modelling complexities of electron microscopy and today we dove into the engineering of freezing and temporally evolving biological samples. It’s difficult to overstate how sophisticated the modern applied sciences are. And impossible not to feel inspired about where will they take us next.

Let me know what you found surprising, informative or muddled. Feedback is a gift.

#atoms

References

[1] Better, Faster, Cheaper: Recent Advances in Cryo-Electron Microscopy. Chua EYD, Mendez JH, Rapp M, et al. - Annu Rev Biochem. 2022

[2] A review of plasma‑assisted deposition methods for amorphous carbon thin and ultrathin films with a focus on the cathodic vacuum arc technique. Anurag Roy, Shengxi Wang, Kyriakos Komvopoulos - Journal of Materials Research 2023

[3] Thermal stability of Si/SiC/ta-C composite coatings and improvement of tribological properties through high-temperature annealing. Jang, YJ., Kim, JI., Kim, Ws. et al. - Sci Rep 2022

[4] Light-coupled cryo-plunger for time-resolved cryo-EM - Yoder N, Jalali-Yazdi F, Noreng S, Houser A, Baconguis I, Gouaux E - J Struct Biol. 2020

Continuing in the vein of self-study based blog posts, let’s understand electron based imaging together.

“I think I can safely say that nobody understands quantum mechanics” - Richard Feynman

A classic physics experiment of individual photons passing through a barrier with two parallel slits, causing each photon to create an “interference pattern with itself”, is striking in its implications about the wave-particle duality of light.

The incomplete timeline is as follows:

• 1801 - Thomas Young observes the interference pattern for light. No one questions the wave nature of light.

• 1887 - Heinrich Hertz observes a photoelectric effect experimentally. Negative electric charge is emitted from a UV irradiated zinc plate.

• 1905 - Albert Einstein provides a theoretical explanation of photoelectric effect with a hypothesis about light being composed of particles (photons).

• 1909 - G.I. Taylor shows the interference pattern can be obtained even with the faintest of light - approaching individual photons travelling through the double slit barrier “interfering with themselves”.

• 1961 - Claus Jönsson shows electrons create an interference pattern in a double slit experiment.

• 1989 - Akira Tonomura demonstrates that even when sending individual electrons an interference pattern emerges.

So… if electrons behave like light, can we use them for imaging?

Wavelengths, resolutions and reconstructions

When photographing a test target with a monochromatic light, for aperture tending to 1.0, a theoretical image resolution approaches half of the light’s wavelength. For visible violet light of 380 nm that’s around 200 nm. Nanometer is one billionth of a metre.

For atom level imaging, an Angstrom: Å, tenth of a nanometer is a useful unit. For example a salt bridge would typically be about 3Å long.

For X-ray crystallography a common wave choice is 1.54Å, produced by copper anodes in X-ray tubes. Resolution of structures at this level of precision involves taking pictures from many angles, utilising the periodicity of the crystal lattice, as well as incorporating physical knowledge to reconstruct raw X-ray diffraction patterns into a 3D structure that satisfies the conditions imposed by the 2D outputs.

Direct electron detectors

Theoretically the wavelength of electrons accelerated to 300keV energy is approximately 0.02Å. A beam of electrons is sent through a sample and focused with an electromagnetic lens whose focal length varies with the strength of electric current. One may date oneself if a TV cathode ray is a helpful analogy. These lenses operate in magnification modes of around 100.000x.

Prior to development of direct electron detectors, CCD sensors were used for electron microscopy imaging, by first directing scattered electrons through a solid state crystal scintillator which absorbed electrons and emitted photons for CCD to capture. 

Currently machinery exists to capture the electron signal directly. They have two modes of operation: cumulative, where energies of multiple electrons interacting with a single pixel are added. And counting, where only the number of interactions is reported.

A modern direct electron detector like Apollo, will have the following specifications: 4096 x 4096 pixels. Each pixel is 8 micrometres wide, implying a 32mm x 32mm sensor. These pixels can report 12 bit (4096 values) deposited energy resolution. And are capable of a refresh rate of 60 frames per second. Pairing the sensor with an FPGA and utilising the time component allows the detector to quadruple its spatial resolution to 8K x 8K when in counting mode.

Phase contrast

CryoEM, which we’ll dive into next time, was recognised with a Nobel Prize in 2017. It builds upon an extraordinary feat of technical brilliance and theoretical insight: a phase contrast microscope. Recognised with a Nobel Prize in 1953 awarded to Fritz Zernike “for his demonstration of the phase contrast method, especially for his invention of the phase contrast microscope”. That microscope was manufactured by a factory founded by afore mentioned Lili Bleeker, a physicist and entrepreneur of enormous wartime bravery. 

In light microscopy, waves passing through objects of slightly different refractive index, will undergo a phase shift. When refocused with a lens, aligned waves will strengthen (constructive interference) or when out of phase dampen (destructive interference).

Defocussing trick

For any given wavelength there will be lens and objective configurations that may extinguish the wave (see an informative video on Contrast Transfer Function). It’s therefore standard practice to capture images at various defocus ranges and to combine and resolve them in post processing.

Back to school: diffraction and Fourier transform

Both of these concepts were developed in the 1910s. Bragg’s diffraction in 1913 and Radon’s transform in 1917.

Diffraction

When a moving particle, photon or high speed electron, encounters a periodic grid and interacts with it, the angle at which it changes its course of flight is strictly determined by its wavelength, periodic spatial frequency of the grid (and an integer diffraction order). With certain a priori knowledge about protein chemistry, combined with a vast number of electrons at different defocus lengths it’s possible to reconstruct the distances between atoms that caused a particular electron scattering angle. 

Radon transform based Fourier reconstruction

Operations in frequency (Fourier, reciprocal) space have always had an element of magic to them. Transforming signals or images into complex numbers with numerically delicate phases, multiplications with kernels and coming back into image space with numerically accurate convolution was an early computer science thrill.

For 3D imaging via 2D image capture we take things a step further: we reconstruct an actual 3D structure in frequency space. Imagine trying to guess a 3D pixel cube containing a protein. It can be proven that taking 2D projections (microscopy images) at various angles, computing their 2D Fourier transforms and placing them in a 3D frequency cube, when inverted back to pixel space will be equivalent to reconstructing the 3D pixel model of a protein. 

The scale of it all

It’s important to appreciate the scale of post-processing required to complete such 3D imaging. To calibrate the direct electron detector based microscope on two known protein structures, a vast amount of data was collected.

• Aldolase [PDB]: 8,682 movies, 1,824,301 particles. Achieving 2.24Å resolution

• Apoferritin [PDB], 3,804 movies and 401,410 particles. Achieving 1.66Å resolution

Aldolase 3D structure from PDB.

References

• Nobody understands quantum mechanics - a short YouTube clip of Richard Feynman.

• Ruizhi Peng, Xiaofeng Fu, Joshua H. Mendez, Peter S. Randolph, Benjamin E. Bammes, Scott M. Stagg, Characterizing the resolution and throughput of the Apollo direct electron detector, Journal of Structural Biology: X, Volume 7,

  2023.

Next time on atoms embeddings and forces

Cryogenic electron microscopy, a technique awarded the 2017 Nobel Prize in Chemistry to Jacques Dubochet, Joachim Franke and Richard Henderson, has accelerated scientific progress, by simplifying preparation and measurement of isolated protein samples compared to the previous imaging workhorse: X-ray crystallography. The crystallisation step - slow growth of a stable lattice of proteins - is replaced with “flash freezing” of randomly oriented proteins with liquid ethane at just above its melting point of -183C.

#atoms

There’s a moment from an improv class I took many years ago that sticks with me to this day. Two people are on stage, opposite each other, a few metres apart. The prompt is: give your undivided attention to your partner, and mirror them. What do you think happens?

Nothing, you're both standing still, looking at each other? Wrong! Magic happens.

With nothing planned, apart from the commitment to be mindful and treat the other person as an equal partner, carefully mirroring their mirroring, adjusting slightly, and making tiny copying errors from one blink of an eye to another, you go in absolutely unpredictable directions.

Today I want to talk about product discovery and ideas in Marty Cagan’s book TRANSFORMED.

In its early chapter we learn about three key areas of transformation that a company needs to undertake to embrace a disruptiveness as a core competency:

1. How we build

2. How we solve

3. How we choose what to solve

Let’s dive into 2: discovering a solution to a problem. The author makes a strong case for empowering the product—design—engineering triumvirate, rather than have them act in a top-down managed feature factory.

In the improv analogy, the steps to success are:

1. Rapid prototypes: tiny, organic moves that cause a user reaction

2. Mindful user observation: a tight feedback loop between actors

3. Sensing how the solution performs: listening to the audience’s reaction

The book acknowledges that it almost always takes many iterations to arrive at a solution that works, hence the underlined importance of rapid prototyping.

This brings me back to the world of comedy. Great stand-up comedians experiment, adapt, hone, and brutally rewrite their “bits,” their pieces of routine. It takes a year of showcasing to get to a strong hour long show ready for a streaming service special, but the iterations, feedback loops, and adjustments happen nightly.

There is something exhilarating about pushing the boundaries of the unknown, with no preconceptions, to discover an answer that’s surprising and beyond any one individual’s imagination. This magic comes from a fusion of people’s ideas interacting with reality. There’s art - and very often comedy - in engineering an outcome.

Improv classes have been a surprisingly powerful way to overcoming anxiety about uncertainty and becoming comfortable with The Unknown and The Unplanned. Shout out to Meave Ryan and Steve Roe from Hoopla.

#forces

The late Stephen Hawking said that each equation in a popular science book halves its sales. Today we’re looking at Context-Guided Diffusion for Out-of-Distribution Molecular and Protein Design with the additional challenge of expressing mathematical concepts in plain English.

This paper addresses an important problem in the application of AI in science: the scarcity of experimental data and the unreliable predictions of models outside their training distribution. It also discusses a key challenge in molecular design: scientists are eager to design molecules—i.e., generate samples from a generative model—that maximise a specific property for which some data exists, though typically only within the "already explored" regions of space.

Empirically, the authors observe that when samples generated are distant from the training distribution, guidance models predicting sample’s properties have a propensity to suggest high (promising) numbers with undeserved high confidence. To quote: “[for poorly understood regions of the input domain] overconfident guidance signals risk steering the generative process toward false-positive regions of chemical or protein sequence space.”

Unconditional Diffusion Model Explanation

The diffusion-based approach aims to learn how to generate a sample that resembles one from a training dataset, or more correctly, one that came from the same statistical distribution as the training data. Let's use images as an understandable example. Assume pixel values range from 0.0 to 1.0, and we accept T = ~1000 steps as a reasonable heuristic:

1. Create T images where the image at step 0 is the original, and the image at step T appears as if generated by normal noise.

2. At each step t in [0,T], for the next iteration of the image, every pixel value is drawn away from the original image towards the mean (0.5 - akin to the grey of November London sky).

3. A tiny bit of noise is added to every pixel at each of the T steps.

Once the training data is prepared, starting with N training images, we should now have N*T image pairs. We "simply" train a network to predict image pixel values going backwards in time one time step (1/1000) from pure noise to structured image. So from image[n][t] to image[n][t-1].

Guided (Conditional) Diffusion Model Explanation

The reasoning above can be extended by breaking the dataset into subclasses, where, apart from the input sample, a class label is provided. Both pieces of information can be supplied jointly to the diffusion model. Further, one can imagine not just a class label but a free text description encoded into a fixed-length pre-trained embedding vector accompanying the image. However, it's crucial to underline a significant difference in regression-based modelling, where the goal is to maximise or minimise said value for a previously unseen, generated output sample.

An important distinction is that in image generation, the text description embedder is typically trained jointly with the image generator to integration of conditioning information, and consequently descriptive user prompts. In regression models for tasks like molecular optimisation, the actual value of the target label is used as one of the objectives during training of the guidance model. The model is explicitly trained to predict and optimise for specific target values, which guides the generation process accordingly.

Standard Guided Model Training

A typical model will consist of three parts:

1. An embedding function H that converts the sample during its formation (at any point between 0 and T) into a fixed-dimensional embedding.

2. A mean-predicting function that operates on the embedding to predict the mean value, or k-dimensional values if that’s the target.

3. An equal dimensionality variance-predicting function for the predicted value.

The authors propose that the mean is well predicted via an inner product of the embeddings with a matrix of learned parameters, and that an exponent of the inner product with a second set of learned parameters is effective for variance prediction. H ideally should also be learnable, and experimentation with both warm-started pre-trained embeddings and training H from scratch is recommended. During training, regularisation of the learned parameters is performed by adding the network’s parameter L2 norm to the overall loss.

Context Aware Guided Model Training

Now the authors introduce an additional regularisation that increases the prediction uncertainty in areas with missing labels. This approach is considered suitable as it increases the likelihood of discovering more true positives. The authors specifically reference molecular design, highlighting how finding that "needle in a haystack" molecule is an attractive prospect. More positives (both true and false) and fewer false negatives (missed opportunities) are thus to be investigated. To quote: “(this) enables the conditional denoising process to focus on regions in data space that are near the training data and have the highest likelihood of containing molecules with improved properties.”

The authors work on batches of sample sizes M and analyse how embeddings of each of the M batch elements covary. They construct a square MxM matrix holding information about the similarities between all pairs from the M-sized batch, enhancing the smoothness of the input embedding space. This method helps the model avoid predicting extreme values when target values are unknown.

Combining these two objectives—the standard conditional of matching the predicted values to the labels, and the enhanced smoothness of the embedding space in regions without labels—leads to improved quality of predictions, which the authors demonstrate across a suite of benchmarks.

Summary

It’s great to see such a formally rigorous approach tackling a real problem in biological sciences and, particularly, in molecular design, given the vast unexplored chemical space.

Acknowledgments

Shout out to the paper’s author, Leo Klarner, for helpful clarifications around the motivations for particular choices of targets for unlabelled data and the effects of smoothness constraints.

#embeddings

In the final days of the 1950s(!) Richard Feynman gave an inspiring lecture at the Annual Physics Society about the future where manipulating individual atoms or even construct nanomachines would be possible. It was titled “There’s Plenty of Room at the Bottom” and each year scientists and engineers are taking steps to realise that vision.

Chip manufacturing capabilities have pushed the boundaries of physics many times over, with modern processors manufactured in a technology that runs individual metal lines 20 nanometers apart. See this transmission electron microscopy image of a metal gate. 

These extraordinary capabilities are not limited to manufacturing microprocessors and GPUs. We are beginning to understand precisely the workings of living nanomachines, collections of proteins that form sophisticated engines. And from understanding we’re onto designing structures that have never existed.

How to design a motor and reverse gear from aminoacid chains

I have been drawn to CryoEM structures reveal how the bacterial flagellum rotates and switches direction for its promise of:

• discussing movement principles of nanoscale motors

• appreciating the details of sophisticated experimental methodologies and modelling the outcomes with tools like ChimeraX

• authors augmenting their research with AlphaFold 2

• learning about strong hypotheses and detailed analysis of a fascinating engineering solution of bacterial movement.

As I finish reading the paper I can’t help thinking there’s a fractal nature to appreciation of the detail on many of the above points that an expert could elucidate. If you’re one of them please don’t hesitate to lean in. Still I’ll do my best to share the lessons I’ve learned.

I’ll talk about similarities between a DC motor and a bacterial flagellum movement. I will discuss the protein-based structure and how nature has solved the problem of switching from clockwise to counterclockwise rotation at nanoscale. I’ll describe how changes to alignment of individual protein domains lead to creation of various binding pockets. I’ll talk about experimental cryoEM results and how an imperfect alignment of numbers of protein-based staves in neighbouring parts of the engine leads to a desirable wobbling of the engine. I’ll talk about how neighbouring copies of the same protein link up together to create a circular structure. 

Authors of the paper make several references to terminology of a motor when describing the rotation of a bacterial flagellum. 

• Stator - which in a DC motor is the source of the constant magnetic field and in the paper is a MotA-MotB protein complex that anchors to the peptidoglycan layer of the bacterial cell wall and forms a proton channel providing source of an electric current.

• Bushings - Two sets of rings (Peptidoglycan and Lipid) that stabilise the torque transmission.

• Switch - the centre of attention of the article, which through conformational changes of a vast (6 megadaltons) multiprotein complex is able to move from clockwise to counterclockwise rotation and affect the direction of bacterial movement. 

• Staves - another manufacturing term, although not specifically motor related it paints a vivid picture of a structure composed of a number of identical slats that make up a barrel or a cylinder.

  

The authors make a number of detailed observations about how the physical and chemical process occurs and we will now dive deeper into it.

Properties

Torque

How is the energy provided by a direct current of protons converted into rotational movement? A much older paper elucidates it, by showing electrostatic interactions between wild type and mutated protein electrostatic potential surface. Bacterial flagellar motors rotate, obtaining power from the membrane gradient of protons.

Type of electrostatic interactions between rotor and stator in the bacterial flagellar motor can be inspected in Figure 4 in the paper. Figure shows various mutations (R->E, D->K) that hindered or increased torque transfer.

Stability 

Another piece of interesting structural behaviour I picked up on is the concept of domain swapping. At first I thought this means that proteins detach their domains and swap them out(!) though that made little sense. In fact the meaning behind this is that they slot between each other to strengthen the structure of the overall engine.

Direction switching

I was blown away by the details of the structural switch mechanism. A single Che-Y bound to a cleft in the FliM part in just one of 30+ staves causes a conformational change where the torque coil (FliG-D5) reorients itself [paper’s Figure 4] and all the other staves follow the reconfiguration like dominoes.

There’s a lot more detail in the paper that I absorbed cursorily, and it leaves me wanting to dig deeper: the impact of a three amino acid deletion mutation PAA on CW/CCW propensity, discussion around the full integer ratio vs other than 3:1 proportion of staves (34 FliM to 102 FliN) of different part of the structure on its low energy minima, perhaps preventing the seizing up of the motor and much more about modelling of the CryoEM measurements. It will have to wait for next time.

Conclusion

We live in extraordinary times. Cutting edge research is available open access and thanks to search engines and AI assistants it’s never been easier to grok the details of an ostensibly dense academic publication. What inspires me most is Prash Singh, the author himself, being incredibly generous with his time to answer a few of my clarifying questions. What an exciting, interconnected world we live in.

#atoms

July 28th addendum

Smarter Every Day just posted a 30 minute video on this topic. It’s a reflection of how many people are seeing the consequences of this research.

Join me on a journey to explore biology: #atoms, machine learning: #embeddings, and the messy business of people: #forces. The best way to truly understand something is to try to explain it to yourself and to others. Science is always advancing, and being part of it is like running a Red Queen's race. I aim to stay humble and eager for feedback from diverse viewpoints and dissenting insights.