Brennan Overhoff

@harvard

[Michael Hla @hla_michael]

I trained an LLM from scratch on pre-1900 text to see if it could come up with quantum mechanics and relativity.

While the model is too small to do meaningful reasoning, it has glimpses of intuition.

When given observations from past landmark experiments, the model can declare that “light is made up of definite quantities of energy” and even suggest that gravity and acceleration are locally equivalent.

I’m releasing the dataset + models and leave this as an open problem to the research community.

I also include what this project has taught me about intelligence in a mini essay linked below.

🧵(1/n)

[Michael Hla @hla_michael]

LLM engineered carbon capture enzymes have officially been produced.

The best designs were 170% more active and 25% more stable across extreme pH (Tm +8.5 C).

Winning strategies include adapting a tag from a bacterial carbonic anhydrase, beta barrel core packing, and removing acidic residues.

Current focus is on translation to large scale field tests. Open sourcing all sequences and results.

1/n 🧵

[Brennan Overhoff @BrennanOverhoff]

RT @satvik_93: Thrilled to announce our next event @aginginit: the Muscle Aging Science & Translation symposium @harvardmed, featuring @dav…

[Michael Hla @hla_michael]

I taught an LLM to optimize proteins. It proposed a better carbon capture enzyme.

Introducing Pro-1, an 8b param reasoning model trained using GRPO towards a physics based reward function for protein stability.

It takes in a protein sequence + text description + previous experimental results, reasons over the information given in natural language, and proposes modifications to improve the stability of the given sequence.

🧵(1/n)

[Brennan Overhoff @BrennanOverhoff]

My last post called what a few of these might be. The experiments needed to confirm though tend to happen behind closed doors but hopefully we will know more soon...

[Brennan Overhoff @BrennanOverhoff]

2 weeks ago I tried to predict what caused BGE-105 to fail. Bioage dropped some new clues in their press release this morning. A few things seem to align with what I predicted. Details below...

[Brennan Overhoff @BrennanOverhoff]

If they are, knowing precisely what moieties/groups make off-target interactions favorable in Azelaprag(or other APJ-binders) seems important and hard to do without knowing what those off-targets are.

[Brennan Overhoff @BrennanOverhoff]

Broader thoughts on AI in biotech: there's still a gap between the holistic reasoning that ML/foundation models afford and the interpretability of basic science. Bridging this gap seems worthwhile.

[Brennan Overhoff @BrennanOverhoff]

Why did BGE-105 fail in the STRIDES trial? Found some interesting threads to pull on from thousands of simulated data points. Built an AI copilot to help interpret what’s going on. Full article: https://t.co/yCnEqfVEB4. tl;dr below:

[Brennan Overhoff @BrennanOverhoff]

Immediate takeaways: Azelaprag may have some unaccounted interactions with TGFb, Activin, and other off-targets. Possibly low selectivity to APJ. These could have caused toxicity in combination with tirzepatide.

[Márton Mészáros @marton_mes]

This is cool: GLP1 trials are looking more and more like multimorbidity-prevention (i.e. aging?) trials?
Look at Lilly’s SURMOUNT trial for tirzepatide vs e.g. Novo Nordisk’s SELECT:
1) more general endpoints
2) broader study population
short🧵

[Brennan Overhoff @BrennanOverhoff]

original thread: https://t.co/9Q2BZH8dOn

[Márton Mészáros @marton_mes]

What should the first clinical trial targeting aging look like? How efficacious a drug needs to be so that the trial is worth running? An attempt to answer by some of us in @NornGroup. Long essay with trial design, rationale, and a sample size calculator. Link, few key points:

[Brennan Overhoff @BrennanOverhoff]

I’ve recently been thinking about how and why to design an aging clinical trial with @marton_mes @NornGroup. Learned a lot about how we might incentivize trial economics (beyond “longevity drug = higher EV”) for more conservative stakeholders. Essay link and add’l thoughts below

[Brennan Overhoff @BrennanOverhoff]

For biotechs: if you’re sitting on a DC for something related to aging, run it through a few other preclinical disease models. It could be worth it if your trial becomes 2x smaller.