AI protein designer. Doctor of Medicine & ML
San Francisco, CA
Joined July 2025
- Tweets156
- Following193
- Followers135
- Likes420
A protein helped a blind patient locate objects again.
Enzymes can help recycle plastic.
Protein shells can report enzyme activity through ultrasound.
What protein do you wish existed? I’ll attempt a computational design of one reader idea.
We can improve a protein’s durability or build a new light switch around chemistry borrowed from nature.
The article explores both, including a recent preprint that connects a designed light-sensitive protein to signalling in cells.
I think people don't get this thought experiment's explicit assumption: trial pass =/= efficacy, and this is about "reward hacking".
Every clinical trial is designed to measure efficacy, and it's.
Like, imagine a drug that prevents the disease but gives you a heart attack after 5 years - it will meet the clinical endpoint. Moreover, understanding the mechanism in this case would allow us to predict it!
Don't get me wrong - all of the points about the fact that we don't understand how many drugs work are true - it's just irrelevant to the context of this thought experiment. Also, to be fair, Tao explained that thought experiment poorly.
Greg Preibisch, MD retweeted
How aggressive a cancer is and how large is the benefit from chemotherapy are different estimands. Genomic assays were built decades ago for the first and repurposed for the second.
Today, we released 3 preprints on estimating the benefit of escalated treatment directly.
Greg Preibisch, MD retweeted
First in person demo of endgoal groupchats at our Warsaw engineering dinner last night
Hey guys (cc: @proteinrosh @TomSercu @sokrypton @biohub @AyushmanMallick) - to dig into this I used a protein I’ve worked on, RSV prefusion F, which has well-characterised flexible loops (the site-Ø epitope).
Hypothesis: the model is generally well-calibrated, but overconfident specifically on intrinsically flexible regions.
I measured each residue’s intrinsic flexibility as the Cα-lDDT between 3 different PDB structures (and between chains within each PDB), then folded the same sequences with ESMFold2 (with MSA) and compared pLDDT vs actual lDDT (vs the experimental structures) - calibration error on the y-axis, intrinsic flexibility on the x-axis.
Result: pLDDT tracks true accuracy well overall (r≈0.83), but the more flexible the region, the more overconfident the model gets - overconfidence averages +7.6 lDDT pts on flexible residues vs +1.4 on rigid ones.
This is easy to miss in standard evals: well-ordered residues vastly outnumber flexible ones, so aggregate calibration looks fine while the flexible-region overconfidence gets averaged away - a flexibility-class imbalance.
Update - here is the comparison with Boltz-2 from @boltz_bio. Seems that's an issue of esmfold2, not Boltz-2.
@AyushmanMallick thanks for pointing valuable protein to benchmark on. Pattern is the same
Limitation - that's only 1 protein, 3 pdbs (however it's well aligned with the reported cases).
Solutions (?):
1. Add a new eval for this one specific property
2. Modify loss for pLDDT to have a higher weight for flexible regions. Like in class inbalance.
Greg Preibisch, MD retweeted
Today we're announcing ESMFold2, an open scientific engine to power prediction, design, and discovery across protein biology.
The new model delivers state of the art performance on protein interactions, especially antibodies, a critical modality for therapeutics.
We have designed and validated miniprotein binders and single chain antibodies across five therapeutic targets that are important in cancer and immunology. We are seeing very high success rates, and affinities at levels consistent with therapeutic activity.
We’re also releasing an atlas of 6.8 billion proteins, and 1.1 billion predicted structures.
ESMFold2 is built on a state of the art language model that has been trained on billions of protein sequences.
A world model of protein biology emerges through language modeling.
We’ve used the techniques of mechanistic interpretability developed to understand large language models to understand the concepts ESM uses to represent proteins.
The model’s representation space has a compositional organization of features across scales, levels of complexity, and abstraction, that reflects and mirrors the understanding of protein biology developed through a century of empirical science.
This understanding emerges without prior knowledge, just from language modeling of protein sequences.
Language models are becoming a powerful substrate to understand and program biology.
The design of protein interactions is one of the most fundamental problems in biophysics, and has critical implications for the discovery of new medicines. A simple gradient based search with the model was able to discover high-affinity protein binders.
I'm excited by the potential this has to accelerate basic science and the understanding of proteins. And especially for the new avenues it opens up for therapeutic design and medicine.
I designed over 30 molecules before being 30, so I got #30under30 Europe awards from @Forbes.
Anyway, enough celebrating. Back to designing the next generation of vaccines. The real science is way cooler than the sci-fi :)
I absolutely hate how much noise we can observe currently, everywhere - in LLMS, in longevity, in biology. That's AI froth.
Literally, people try to reason from the noise and get hyped about it.
Adding empty spaces also improves performance.
arxiv.org/pdf/2510.01032
It may be just having more computation by having a bigger attention matrix, or just random statistical noise.
Fascinating Google paper: just repeating your prompt 2 times can seriously boost LLM performance, sometimes pushing accuracy from 21% to 97% on certain search tasks.
An LLM reads your prompt left to right, so early words get processed before the model has seen the later words that might change what they mean.
If you paste the same prompt again, the model reaches the 2nd copy already knowing the full prompt from the 1st copy, so it can interpret the 2nd copy with the full context.
That means the model gets a cleaner “what am I supposed to do” picture right before it answers, instead of guessing too early and sticking with a bad setup.
This helps most when the task needs details that appear late, like when answer choices show up before the actual question, because the 2nd pass sees both together in the right order.
In the Google tests, this simple trick took one hard search-style task from 21.33% correct to 97.33% correct for a model setting with no step-by-step reasoning.
Across 7 models and 7 benchmarks, repeating the prompt beat the normal prompt in 47 out of 70 cases, and it never did worse in a statistically meaningful way.
The big deal is that it is almost free to try, it often boosts accuracy a lot, and it shows many LLM mistakes are “reading order” problems rather than pure lack of knowledge.
----
Paper Link – arxiv. org/abs/2512.14982
Paper Title: "Prompt Repetition Improves Non-Reasoning LLMs"
Nice, my insults finally started to hurt Claude!
(I’m going to be the first person killed by Claude)
Greg Preibisch, MD retweeted
Radar graphs are among the worst ideas in data visualization. The whole point of them is to show the area and you can usually reorder the labels freely in order to create a desired dramatic effect.
Two versions of the same graph:
- left one tells the story that AI is rapidly replacing whole industries
- right one shows the "jaggedness" and reinforces the idea that humans will always have something that AI won't be able to replicate
