how i think AI, especially general purpose models like LLMs, *might* solve (one-shot design human viable therapeutics) bio in 10 years:
1. improvements in efficient physical simulations and understanding => bottoms up priors. For example, dynamics priors would (likely) be extremely helpful for tasks like binding affinity and functional potency prediction. The best computational method for binding affinity, FEP, is reliant on these dynamics.
2. conditioning on massive amounts of semi-structured high level data. examples include all experimental conditions for biochemical and in vivo assays including ambient conditions and protocol deviations; free-text patient records over time including symptoms and outcomes; lots of wearables data; videos of procedures being performed. These all affect the outcomes we care about and I think it's necessary to factor them in when predicting outcomes.
3. calibration on high level outcomes like human efficacy and free-text symptoms that include nuances of disease progression/regression; raw animal behavior; cell state changes. collapsing outcomes also takes away information to learn from and obscures partial wins that are still important.
4. Massive gains on sample efficiency; a classical example is that a chemist can get more information out of a failed assay on a designed molecule (e.g. "this functional group is important because replacing it kills potency") while it is not clear a gradient update can do the same.
The idea is that i think there is much signal in the information that is not being used in all attempts to model biology. I am not sure if this is sufficient, but i am quite convinced it is necessary to "solve" biology. i suspect an upcoming trend will be figuring out how to obtain and pair and mix all this data together. or perhaps it will be a task for post-RSI AI to figure out. with the exponential takeoff of AI, i do not think it is impossible for therapeutics to be one-shot with such high confidence that people will take the therapeutic with the certainty it will pass clinical trials before it does
@megha_byte ohh yeah i wonder if more examples, not necessarily with poses just with Kd/IC50/binary binding information, can help. perhaps it might help more as the model gets better? super interesting
woahh i wonder if you could get improvements by in-context conditioning on affinity measurements for other ligands on the same target
this will be super useful for agents natively reasoning about biomolecules and becoming more sample efficient
insanely cool work
Which predicted protein-ligand pose is closer to the crystallographic binding mode? 🧬 Modern cofolding models let us sample infinite candidate poses, but can we reliably *rank* them for drug discovery?
Check out Pyrite, a pose ranking benchmark and AI agent framework w/ @genesismolai!
@datavorous_ make massive biochemical assay dataset (binding affinities, stability, ADMET properties, etc) with all experimental conditions and oddities included and see if LLMs can condition on those and get better predictions. if so, LLMs might be universal model
Some cool (and overdue) results from playing with Bioemu2
1: Maltose binding protein
MBP captures maltose in the periplasm of bacteria and transports them across membranes. Maltotriose is its native ligand and can displace α-gluco-2-fluoromaltose (a reporter for NMR) from MBP while cellobiose cannot detectably.
Impression: simulations appear to show both triggering a conformational change, while the apo form stays mostly where it starts.
(1/3)
If there are more experiments I should be running or any tweaks I should include please let me know (or run it yourself and show!)
shoutout @ycombinator for AWS compute credits
3. Benzene unbinding from T4 lysozyme
I wanted to see how calibrated/uncalibrated the model is for Koff estimation.
I simulated two variants: F104A and L99A.
According to https://t.co/mJqlusDbFz
F104A has Koff >1.00×10^4 ~ mean residence time of 100us
L99A has Koff 9.50×10^2 ~ mean residence time of 1ms
I find it quite cool that double residue mutation results in a really large change in binding kinetics, AND that is picked up by Bioemu2!
hello hi everyone i am saying my first words 😃😃
my name is shayaan and i’m an undergrad studying cs at udub! i really like proteins and work in david baker’s lab using RL to post-train protein design models so they behave like a good boy
im really interested in post-training, especially for flow-matching models. i also love mech interp. i think there’s so much low hanging fruit when applying this to protein design
my end goal in life is to solve disease and as a first step i prolly need to solve my sleep schedule
hmu if you’re working on ANYTHING cool i need to do more OOD things 🫡
hello hi everyone i am saying my first words 😃😃
my name is shayaan and i’m an undergrad studying cs at udub! i really like proteins and work in david baker’s lab using RL to post-train protein design models so they behave like a good boy
im really interested in post-training, especially for flow-matching models. i also love mech interp. i think there’s so much low hanging fruit when applying this to protein design
my end goal in life is to solve disease and as a first step i prolly need to solve my sleep schedule
hmu if you’re working on ANYTHING cool i need to do more OOD things 🫡
@bearablylight I'm curious how they determined this! I would imagine it passed in vitro/biochemical screening; I wonder if failure to engage was actually something to do with the environment the target is in under therapeutic settings (something that's difficult to condition in those models)
:ooo
i was thinking this could be a super interesting objective to optimize against to make new enzyme active sites
would especially be so cool to do PFAS degraders and lots of exotic or industrially-relevant synthetic pathways
maybe even binding specificity?? synthetic data for tricky binding affinity prediction would be amazing
Holy moly
Bioemu2 https://t.co/0kQPRHtIhu
and
Vitrify https://t.co/WGunMw1aIY
are essentially what I was thinking about when I described making rules for bio. I'm certain there are many shortcomings in the models that we will come to get familiar with, but I think this is the correct next direction for the field. I think a lot of it will center on knowing what level of accuracy is sufficient (we technically know all of the physics needed to describe biology it's just far too complex to fully compute), and figuring out how to achieve that level across all possible edge cases robustly enough to train generative models against them
My prediction is that someone will figure out how to do Free Energy Perturbation much more efficiently while retaining good accuracy, probably through some clever deep learning methods, and that'll be the next breakthrough. after all LLMs are really good at optimizing something quantifiable
also if anyone has some spare compute for me that would be swell I have many things I would love to simulate and poke at
First X take: biology/medicine is a harder science than the fields classically considered harder like math and physics
Unlike those fields, there’s not a consistent, predictive, manipulable language to describe biological systems and prove corollaries with. Without it, we have to rely completely on empirical trends and fuzzy language. I don’t believe it’s impossible; it’s a result of us not being able to fully define all the individual processes and their interactions that make up a biological system
I think building that language is the holy grail of computational biology
I also think this is a somewhat LLM-legible task: establishing a set of principles that recapitulate known experimental phenomena and then show it generalizes. The principles might be super complex, perhaps each being an Alphafold-shaped model, but as long as it’s computable it’ll still be amazing
It could perhaps even have some uncertainty baked in
(Semantically I’m defining biology as the study of biological systems, and medicine as the study of human health. I also realize I’m kinda reframing it as a physics problem)
Hello X nice to meet you
My name is Douglas. My goal in life is to live more life. The dream is to cure every disease, including aging. Eventually, I hope to enable arbitrary design of biological systems, because the beauty of life is too sublime to be limited to what evolution has sampled.
In this upcoming era of superintelligence my dreams seem to have become more possible than ever. However, my role in the future has become less clear and so that is top of my mind most days.
I will post thoughts of varying refinement here about my learnings. My goal is to make 7th grade me widen his eyes in wonder and poop his pants in excitement at what he will experience in the future
Please talk to me for any reason i really like learning about different fields or expounding on what i find cool