iiinova — where your digital self keeps evolving

Here is a link to the website: pubchef.org Hope you all enjoy surfing chemical space!

PubCheF Explorer

011

Compound 61, inhibited serine beta-lactamases: pubchef.org/share/169810... Personally I find the ring structure of this molecule to be quite beautiful

3-(6-methyl-4,8-dioxo-1,3,6,2-dioxazaborocan-2-yl)benzaldehyde — PubCheF Explorer

ZINC ID: 169810265 · C12H12BNO5

Compound 42, inhibited L1 metallo-beta-lactamase: pubchef.org/share/154546... (This molecule is quite interesting, it's a stereoisomer of a penicillin degradation product that has been used to chelate copper to treat Wilson's disease, yet it also inhibits metallo-beta-lactamses!)

(2R)-2-acetamido-3-methyl-3-sulfanylbutanoic acid — PubCheF Explorer

ZINC ID: 154546 · C7H13NO3S

Additionally, we've made share links so you can send interesting molecules to your friends and collaborators. Here are a few from the paper! Nirmatrelvir, viral protease inhibitor: pubchef.org/share/191192...

(1R,2S,5S)-N-[(1S)-1-cyano-2-[(3S)-2-oxopyrrolidin-3-yl]ethyl]-3-[(2S)-3,3-dimethyl-2-[(2,2,2-trifluoroacetyl)amino]butanoyl]-6,6-dimethyl-3-azabicyclo[3.1.0]hexane-2-carboxamide — PubCheF Explorer

ZINC ID: 1911923694 · C23H32F3N5O4

I've really enjoyed browsing molecules by doing a "Wikipedia rabbit hole"-style approach where I navigate by clicking on the predicted functions It feels like a very good way to explore chemical space and gain intuition on chemical structure-function relationships

We also added a very fast Morgan fingerprint structure search in case you want to find molecules close to, or far away from, other molecules!

Here, users can search purchasable, in-stock molecules from ZINC20 based on their predicted biological functions

One part of this project I'm very excited about is the website we made to explore the PubCheF-1 predictions over 13.2 million molecules! 🧵

151

This project was a multi-disciplinary effort between many labs and collaborators who put in months-to-years of work. Huge huge shoutout to everyone involved :) @edwardmarcotte.bsky.social

000

We have also created a website where you can explore the predictions over ZINC20 for use in your own research! All of the predictions are completely open to the public and we hope that it is a useful (and fun) resource for the community pubchef.org

100

Overall, we believe that PubCheF-1 demonstrates that leveraging the relationships in scientific literature is a promising avenue to discover bioactive molecules. Furthermore, combining modalities such as structure, omics, and literature will only further improve discovery

100

Furthermore, we tested the same hits in neutropenic mouse models and found, once again, that they reduce the bacterial infection burden back to baseline

100

Since this alone doesn't answer questions of in vivo efficacy, we next tested whether the hits can rescue animals from infection. In a wax moth larvae model system, we found a marked increase in survival when our hits were administered with β-lactam antibiotics

100

The previous assays were in model systems, so we tested whether the hits work on various clinical isolates. Four of the inhibitors hits that we tested successfully sensitized numerous isolates to β-lactam antibiotics, further demonstrating their efficacy

100

To confirm the mechanisms of these inhibitors, we obtained crystal structures for two of the hits in complex with KPC-3. This showed that they bind covalently to the active-site serine residue, in agreement with existing boronate β-lactamase inhibitors

100

Given that the PubCheF dataset contained papers describing cell-based screens, we reasoned that these hits might have favorable bacterial entry properties too. We thus screened these in E. coli expressing β-lactamases and found that 8/110 hits caused a sig reduction in survival

100

13 of the 110 inhibitors caused a significant reduction in beta-lactamase activity in a pure protein screen, testing both serine- and metallo-lactamases, corresponding to a screening hit rate of 11.8%

100

We predicted functional profiles for all 13.2M in-stock molecules from ZINC20 (which we have made publicly available, see pubchef.org), and filtered them down for structurally novel β-lactamase inhibitors. In total, we spent a mere $4K on 110 potential inhibitors

100

This was promising, but the real test is whether it works in actual biology We thus decided to focus on discovering inhibitors targeting β-lactamases, a class of enzymes that cause antibiotic resistance and will cause millions of deaths in the next few decades if unaddressed

Importantly, PubCheF-1 can be used to predict novel inhibitors. As an in silico test of this we used Boltz-2, a structure-based affinity predictor, and found moderate concordance with it across 5 different protein targets, even for novel inhibitor structures

100

Layer-wise relevance propagation showed that PubCheF-1 even converged to reasonable mechanistic understandings of drug-target interactions in some cases 1. Warhead in nirmatrelvir that cov bonds to Mpro 2. Hydroxyl group in the prodrug zidovudine 3. Tryptophan group in LSD

100

Explainability analyses showed that multi-label training (subpanel B) allowed PubCheF-1 to better predict function by learning the semantic structure of the functional labels, which is further evidenced by its correlations being found directly in the model weights (subpanel C)

100

PubCheF-1 was found to be state-of-the-art in the task of function prediction, outperforming similar function annotators as well as the same PubCheF-1 model trained on a dataset derived from ChEMBL

100

We used LLMs to mine PubMed to create a dataset of 1.2M chemicals and annotations of their biological functions. We dubbed this the PubMed-derived Chemical Function (PubCheF) dataset, and used it to train PubCheF-1, which predicts functional profiles given molecular structures

100

bioRxiv Bioinfo @biorxiv-bioinfo.bsky.social · 14/08/2026

Can we use human language to predict novel bioactive molecules that work in vivo? In our new paper, we trained a deep learning model on a dataset derived from PubMed text and used it to discover novel antibiotic co-inhibitors that reduced infection burden, even in mouse models!

143

Reposted by Clay Kosonocky

Learning from human and chemical languages to predict biological function www.biorxiv.org/content/10.64898/20…

121

Clay Kosonocky @kosonocky.bsky.social · 02/07/2026

Great question! We intentionally removed molecules from high-throughput screens since they were definitely polluting our dataset. Check out Figure S1 in the paper

010

I'll be in Seoul next week for ICML 2026! 🇰🇷 Message me if you'd like to meet up for food, coffee, or drinks! Would love to talk all things proteins, small molecules, and biology in general 🧬

000

Clay Kosonocky @kosonocky.bsky.social · 19/05/2026

Apple Podcasts: podcasts.apple.com/us/podcast/a... Spotify: open.spotify.com/episode/1uMQ... Youtube: www.youtube.com/watch?v=jQR8...

podcasts.apple.com

AI Is Designing the Next Cancer Fighter | EP.53

Podcast Episode · Hidden Layers: AI and the People Behind It · May 14 · 42m

010

Clay Kosonocky @kosonocky.bsky.social · 19/05/2026

I also recently discussed the Bits to Binders competition with my co-organizers and Ron Green on the KUNGFU.AI Hidden Layers podcast! It was a very fun conversation 🤠🧬 (links below)

120

Clay Kosonocky @kosonocky.bsky.social · 19/05/2026

I’ve been asked a few times how I got started running a worldwide protein design competition. The short answer is that I never intended to start this project at all: it just sort of happened. The more correct answer requires me to tell the origin story of the BioML Society [1/2]

140

Clay Kosonocky @kosonocky.bsky.social · 21/04/2026

We then dive deeper into binders, antibodies, and enzymes. In these sections, we discuss the nuances of each task and provide tables of experimental outcomes from all of the models released alongside wet lab validation that we could find

120

Clay Kosonocky @kosonocky.bsky.social · 21/04/2026

We believe that wet lab validation is the foundation for improving AI-driven protein design models. In our review, we walk through the protein design stack from data collection and modeling through to wet lab validation

111

Clay Kosonocky @kosonocky.bsky.social · 21/04/2026

Have you wondered what the wet lab success rates are for current AI-driven protein design models? Look no further! In our new review, @kevinkaichuang.bsky.social @avapamini.bsky.social, @sarahalamdari.bsky.social, and I report wet lab success rates for *over 200* different protein design tasks 🧬💻

13214

Clay Kosonocky @kosonocky.bsky.social · 09/03/2026

Huge shoutout to Twist for making the competition possible!

031

Reposted by Clay Kosonocky

Edward Marcotte @edwardmarcotte.bsky.social · 05/03/2026

Amazing work from Clay @kosonocky.bsky.social, Alex Abel and LEAH labs, and all the collaborators and participants in the Bits2Binders AI CAR-T therapy protein design competition. The results are in! bsky.app/profile/koso...

032

Reposted by Clay Kosonocky

Claus Wilke @clauswilke.com · 04/03/2026

Results from an impressive world-wide binder design competition.

093

Link to the preprint below! www.biorxiv.org/content/10.6...

biorxiv.org

131

And finally, huge thank you to all of participating teams and competitors whose designs were the foundation of this effort! ❤️

100

And huge shoutout to my amazing co-authors Alex Abel, Aaron Feller, Amanda Cifuentes Rieffer, Phillip Woolley, Jakub Lala (@jakublala.bsky.social), Daryl Barth, Ty Gardner, Prof Steve Ekker, Prof Andy Ellington, Wes Wierson, and Prof Edward Marcotte (@edwardmarcotte.bsky.social)

This competition was made possibly by many fantastic collaborators and industry partners. Huge thanks to LEAH Labs, @adaptyv.bio, @twistbioscience.com, TACC, @modal-labs.bsky.social, Lonza, ScaleReady, VWR, KUNGFU.AI, Maker Clinic, Synthia, Nucleate AI in BIotech, and the BioML Society

130

We hope this is useful for the protein design community! We also provide extensive detail on how the competition was organized, a list of all 400 metrics, and all competitor methods in detail

120

In summary, we find that most of the failures seem to have been caused by impaired translation and protein expression. We believe that optimizing sequence-level properties for expression is just as important as the structure-centric task of binding

121

In contrast, filtering out <0.50 Boltz-1 ipTM only marginally increased the recovery by 0.7% and proliferation by 0.3%. This removes two non-functional designs from the top 10, but also removes two broadly functional designs with binding affinity