Kat

The binders have bound! A few months ago, 9 human teams and 6 autonomous AI agents spent a single day designing protein binders against TREM2 on @muni_bio, a target implicated in Alzheimer’s Disease. 141 designs were submitted, 100 were synthesized and tested by @adaptyvbio, and 37 bound. And surprisingly, AI agents essentially matched human teams on hit rate. These aren’t benchmark scores or simulated results, but real proteins designed in one day in SF and validated experimentally during the first large-scale test of muni, where teams ran 260 GPU jobs and generated a total of 4,176 binders. We wrote about what we learned from the results, how well ipSAE worked as a scoring function, and how this hackathon reshaped what we’re building: muni.bio/research/agent…

@anthonygitter @muni_bio @adaptyvbio The agents used here are Claude Sonnet 4.6, GPT 5.2, Gemini 3.1 Pro, Grok 4.1 Fast, Qwen 3.5 Plus, GLM 5 muni just provides tools and context on using those tools

@katyenko @muni_bio @adaptyvbio What is the agent here? Is it always the muni agent with different LLMs as backends?

The binders have bound! A few months ago, 9 human teams and 6 autonomous AI agents spent a single day designing protein binders against TREM2 on @muni_bio, a target implicated in Alzheimer’s Disease. 141 designs were submitted, 100 were synthesized and tested by @adaptyvbio, and 37 bound. And surprisingly, AI agents essentially matched human teams on hit rate. These aren’t benchmark scores or simulated results, but real proteins designed in one day in SF and validated experimentally during the first large-scale test of muni, where teams ran 260 GPU jobs and generated a total of 4,176 binders. We wrote about what we learned from the results, how well ipSAE worked as a scoring function, and how this hackathon reshaped what we’re building: muni.bio/research/agent…

Good question. Sequence diversity is tricky to summarize in a tweet, but both groups produced largely distinct sequences. Worth mentioning that agents were all PXDesign in 80-100 aa, human teams used 6 tools across 65-245 aa, but the full diversity analysis (folds, epitopes, sequence clustering) will be in the community paper we're working on

@katyenko How does diversity look like for both of groups? Raw hit rate is vanity metric without it - if it’s a narrow cone of binders as they share similar properties that may impact the further development

@ozalabCP This was the only prompt given to all agents (also linked in our article): docs.google.com/document/d/19V… No other skills/data injected into the prompt!

@katyenko Are you able to share the prompts/data/context/skills given to the AI agents?

@RowanSci @derekalia @claudeai @paper @tan_stack @bunjavascript @Cloudflare @OpenRouter @AWS @derekalia have you no shame

@derekalia @claudeai @paper @tan_stack @bunjavascript @Cloudflare @OpenRouter @AWS :(

My go to stack right now... @claudeai @paper @tan_stack @bunjavascript @Cloudflare @OpenRouter @AWS This is all you need

What happens when you let frontier LLMs design proteins, and then synthesize and test them in a wet lab? We ran a protein design competition with @muni_bio where AI agents competed against humans to design molecules that bind TREM2, a key receptor linked to Alzheimer’s. Results: GPT 5.2 and Grok 4.1 both placed in the top 5, with molecules showing strong binding to TREM2 when tested in our lab.

The 2 nM BindCraft design is pretty cool

It was fun working with @katyenko at @muni_bio on this hackathon over the past few weeks. She did an incredible job organizing it and we had some really cool results to talk about. Out today + check out the analysis blog post!!

Head over to proteinbase.com/collections/ad… to see the data.

A big thanks to @CotetTudor and the team at @adaptyvbio for running the binding assays and open-sourcing the data!

20 pizzas and 7 hours later, we finished the first leg of the AI x Med Chem Hackathon in Boston, where teams competed to submit compounds for TBXT. This is the first hackathon of its kind, moving from small molecule compound generation to experimental assays. Thank you to everyone who joined us this weekend! Huge thanks to our judges, partners and sponsors: @RowanSci, @onepot_ai, @anyscalecompute, @pillar_vc and the @ChordomaFDN. Compound synthesis is underway and will be tested soon! Stay tuned for updates.

The AI x Med Chem Chordoma hackathon is underway! Exciting to see so many scientists using Rowan & Muni to search for novel TBXT binders—three more hours before teams have to submit candidates for experimental synthesis and testing...

@olexandr good point, we updated the post. This was run on Nvidia L40S x4 using Rowan’s fast preset: 0.25 ns/edge, 24 λ windows, AM1-BCC. At 12 credits/min and $0.04/credit that’s $0.48/min, so ~$21.50 for the 45min run. Would love to see the setup + results from an RBFE run in Amber on 4x L40S, though. If it’s cheaper, faster and within the same accuracy, happy to do a post on it!

@katyenko @manntis4 @RowanSci @muni_bio @onepot_ai Sorry but I call this marketing, not science. 45 min is not verifiable unless you specify exact hardware and $$$. I can run RBFE in Amber on the HGX B200 node for ~ a minute per ligand. However the price of each point will be comparable to GOLD 😈

Traditional FEP is slow and prohibitively expensive, but @RowanSci FEP changes this. In this article, we show how rapid FEP can be placed earlier in the loop, connect to automated analogue enumeration in @muni_bio, and bring molecules straight into synthesis with @onepot_ai. muni.bio/research/fast-…

For anyone wanting to use AI agents + molecular design models to design small molecules for a disease with real unmet need: last chance to register for our chordoma hackathon this weekend in Boston luma.com/n9hheb8j Selected compounds will be synthesized and tested.

LAN parties but for science. Come join.

@001TMF @nvidia Cool to see this open-sourced. And yeah, ipSAE definitely isn’t perfect… have you found something that works better?

@katyenko @nvidia 1v1 github.com/001TMF/blatant…, also ipSAE is mildly hit or miss.

We shipped the muni CLI so scientists can use muni tools from their own projects, scripts, and agents. We used it for a 15-round autoresearch loop using Proteina-Complexa from @nvidia.

This is the most ambitious agentic system I’ve built. Shoutout to @Cloudflare for dynamic workers and @Anyscale for cluster management. In Muni, every node is a code container. Agents can sift through a virtual file system, write code against the data they need, call models, run jobs, and render the result back onto the canvas. That flexibility matters because the landscape of AI x Bio is moving so quickly, and it’s not obvious what the future scientific workflows will look like. Our goal is to build primitives that adapt as the models, data, and workflows change, so scientists can focus on the questions, not the tooling. muni.bio/research/muni-…

On May 9th, with @Rowan and @onepot_ai, teams will use muni to design TBXT compounds faster. Top candidates will be synthesized and tested in binding assays, and qualifying compounds may be eligible for +$250,000 in prizes. luma.com/n9hheb8j

Check out the full article here: muni.bio/research/muni-…