
PhD candidate, @USC_Physics and @qcb_usc, University of Southern California | ML in chemistry and biology | AI for drug discovery
Los Angeles, CA
Joined September 2022
- Tweets230
- Following895
- Followers532
- Likes1.8K
Pinned Tweet
I’ll be presenting our #GenerativeAI drug design method DrugHIVE at #ACS tomorrow. Come check it out! @JCIM_JCTC
talk: Automated drug design and optimization with structure-based generative deep learning
paper: pubs.acs.org/doi/10.1021/acs…
code: github.com/jssweller/DrugHIV…
Jesse Weller retweeted
📣 MLCB + MLSB are joining forces under MLBio.org ! 🧬🤖
The first workshop will be held 2027, Join the MLBio mailing list to stay updated on calls for papers, deadlines and more info on the first event!
groups.google.com/g/mlbio
Jesse Weller retweeted
We are so excited to share the formation of MLBio! 🚀🚀
Looking forward to seeing what comes of merging two very vibrant scientific communities 🧬🔬
What a fantastic effort to do this alongside @ebetica @pkoo562 and @david_a_knowles
Sign up or follow @workshopmlbio to stay up to date!
📣 MLCB + MLSB are joining forces under MLBio.org ! 🧬🤖
The first workshop will be held 2027, Join the MLBio mailing list to stay updated on calls for papers, deadlines and more info on the first event!
groups.google.com/g/mlbio
Jesse Weller retweeted
Excellent work led by @butcher_jasper and @Chensss_J . Seamless design infrastructure has the potential to impact the field of protein design more than incremental progresses in diffusion tech, exemplified by zero shot wet lab successes presented here.
doi.org/10.64898/2026.10.19.…
Reliable and Efficient Agentic Protein Design with Toolset Harnessing @RosettaCommons
1 Chaperone is introduced as a lightweight “agent harness” that standardizes protein-design tool orchestration: it wraps RFdiffusion3 generation with downstream sequence design, structure prediction, and scoring (e.g., ProteinMPNN, AlphaFold3) via a single DataFrame-to-DataFrame API, enabling rapid iteration without bespoke glue code.
2 The core abstraction is the “block”: each tool is implemented as a small Python class that takes a DataFrame of designs and returns an extended DataFrame, appending new structure records plus metrics while preserving full lineage (history of steps, file paths, residue-level annotations). This makes generation/prediction/scoring composable in any order.
3 Chaperone focuses on interoperability without forcing tools to change: each block separates prepare/execute/compile phases, treating the underlying tool as a black box while converting inputs/outputs to a canonical schema. Intermediate results are stored as Parquet, and pipelines can resume from partially completed chunks.
4 The execution engine scales pipelines on local multiprocessing or HPC schedulers (via Submitit), automatically chunking DataFrames across GPUs, tracking lineage during merges, and supporting failure recovery. It also supports “bundled” multi-step jobs (e.g., RFD3→MPNN→AF3 in one job) to reduce queueing and straggler overhead.
5 Agent-oriented design choices include: (i) standardized block docs and examples so an agent can plan workflows before running them, (ii) a build–test loop with a test mode that runs lightweight end-to-end checks, and (iii) skills for monitoring jobs, debugging, and recovering runs—aimed at reducing non-scientific orchestration work.
6 In a 9-task autonomous evaluation suite (12-hour runs; agents must submit 96 diverse sequences; held-out scoring pipelines), GPT-5.6-sol agents with Chaperone used 86% fewer tokens and 64% fewer GPU-hours than baseline agents that had to assemble pipelines themselves, while increasing mean computational pass rate from 6% to 12%.
7 The paper argues that as “reasoning effort” increases, baseline agents spend disproportionately more tokens/actions on workflow implementation and job management; with Chaperone, token use stays far lower even at high effort, consistent with reduced repeated polling and less scaffolding regeneration.
8 Wet-lab validation: a two-week end-to-end cortisol small-molecule binder campaign used Chaperone to explore many workflow variants (COM/length sweeps in RFD3; LigandMPNN composition biases; multiple oracle combinations; iterative cycling and filtering). From 27 tested candidates, the best binder achieved KD = 1.32 nM “zero-shot” (no affinity maturation).
9 Additional experimental demonstrations show versatility beyond standard generate→predict pipelines: (i) a tetrahedral 12-subunit nanoparticle workflow integrating uncommon software (RPXdock) plus MPNN/AF/Rosetta, yielding 1/15 successful assemblies (6.7%); (ii) rigid linker/fusion design between C2- and D3-symmetric complexes using RFD3 fusion + pilot sweeps for linker length/MPNN variant + CaGE cycling + AF3 monomer/hexamer evaluation, producing crystal observations by brightfield microscopy.
💻Code: github.com/RosettaCommons/ch…
📜Paper: biorxiv.org/content/10.64898…
#ProteinDesign #ComputationalBiology #AI4Science #RFdiffusion #ProteinMPNN #AlphaFold3 #HPC #Bioinformatics #SyntheticBiology #AgenticAI
Jesse Weller retweeted
Public preview of BioEmu2! @MSFTResearch AI for Science’s new foundation model for biomolecular structure, dynamics and function. Open weights, MIT license.
github.com/microsoft/bioemu2
Movie: BioEmu2-simulated binding of intrinsically disordered peptide PUMA to MCL-1 receptor. 🧵⬇️
Jesse Weller retweeted
BindCraft2 v1.0.4 has been released: with significantly improved VHH, scFv, ARP, and induced fit targeting. WE also improved speed and code stability. Try your difficult campaigns again. Great work by Erik Mäeots, Leonardo Tredese and Elizaveta Chernova.
Jesse Weller retweeted
Out today in @ScienceMagazine: RFdiffusion3, with capabilities beyond both the preprint and contemporary models. science.org/doi/10.1126/scie…
I’m also excited to share that I’ve joined Flex Therapeutics @flextx_AI, building next-gen biomolecular design models for oral therapeutics.
Jesse Weller retweeted
Today we're sharing an update on our diffusion method from last year, RFdiffusion3, in @ScienceMagazine. Notably, we show significant improvement in design of sequence specific proteases for amyloid sequences and metallohydrolases for pesticide degradation.
Jesse Weller retweeted
In Science, researchers present RFdiffusion3, a new generative AI system that designs proteins and nucleic acids atom by atom, which could enable the ability to engineer novel biomolecules with precise, complex functions.
Learn more: scim.ag/4hOQVIF
Jesse Weller retweeted
This week we're releasing a preprint exploring state space models for long-sequence genomic prediction, alongside Cerberus, a set of open PyTorch models built on this work. biorxiv.org/content/10.64898…
Jesse Weller retweeted
Biosecurity is too important for security theater.
In a quick test with Raygun, we found that DeepMind's SynthIDBio watermark signal could be washed out while preserving the predicted structure.
We don’t think protein sequence watermarks buy much biosecurity.
Jesse Weller retweeted
ʙɪɴᴅᴄʀᴀꜰᴛ2 is out, and we're not waiting for the paper. The full code drops today, free for academic and industry use.
We're releasing it early so you can start designing right now, and bring its full power to the current Adaptyv competition.
github.com/PacesaLab/BindCra…
Impressive work! But it highlights a critical challenge: synthesizability is not a fundamental chemical property. It is a function of our technology.
We propose a more general solution to this problem in a recent paper (Model-Guided Virtual Screening) doi.org/10.1021/acs.jcim.6c0…
Impressive work! But it highlights a critical challenge: synthesizability is not a fundamental chemical property. It is a function of our technology.
We propose a more general solution to this problem in a recent paper (Model-Guided Virtual Screening) doi.org/10.1021/acs.jcim.6c0…
Large Drug Discovery Model is out! I am exited to introduce our new generative SBDD framework. We experimentally validated LDDM across 5 targets and successfully designed novel, validated hits with structural accuracy confirmed by X-ray.
Preprint: biorxiv.org/content/10.64898…
🧵 1/7
Jesse Weller retweeted
With AlphaFold we mapped the protein universe - now with AlphaGenome Atlas we’re charting the human genome. It can predict the impact of all 9 billion possible single-letter DNA variants, helping scientists better understand disease. Freely available for academic research: alphagenome.google/atlas
Jesse Weller retweeted
Since the 1960s, the genetic code has been used to predict protein sequences from DNA and mRNA sequences. Our @Nature article demonstrates that these predictions miss thousands of protein sequences present in human tissues.
Across >1,000 human samples, we identified numerous abundant proteins whose amino acid sequences differ from those predicted by the genetic code.
These proteins are not rare translation byproducts. They accumulate to thousands of copies per cell. Some are more abundant than the proteins predicted by the genetic code from the same transcripts.
Their abundance reflects a combination of alternate RNA decoding mechanisms — including codon-anticodon mismatches, tRNA abundance, and RNA modifications — and selective stabilization of the resulting proteins. The last factor – protein stability – emerges as a major determinant of protein abundance across proteins, proteoforms and cell types: slavovlab.net/research.htm#P…
Alternate RNA decoding is pervasive across functional groups of proteins, healthy and diseased tissues. It affects proteins playing key roles in neurodegeneration, and some alternately decoded proteins show strong enrichment in tumors compared to their surrounding tissues.
This discovery has been a long and exhilarating journey with Shira Tsour and the @slavovLab team. It started in 2019 and proceeded through many challenges and thrilling highs. A journey that has opened new perspectives that we long to explore!
1/
We report many proteins not predicted by the genetic code.
They are stable & abundant O( 10³ ) copies / cell.
Generative mechanisms include codon-anticodon mismatches & RNA modifications.
Their abundance depends on codon frequency & protein stability.
biorxiv.org/content/10.1101/…
Jesse Weller retweeted
Big news from Boltz - our biggest update yet! 🚀
Today we’re releasing two new state-of-the-art models for protein and small molecule design with extensive wet lab validation and a new API to run all of our models on scalable GPUs wherever you (or your agents) work! 🔥
Jesse Weller retweeted
🍹Weekend project: I (with help from claude - budget $40), was able to get OpenFold3 weights running inside AlphaFold3 codebase (jax). Works for proteins, ligands, rna/dna (1/3)!
AF3 source code is now open source (aka Apache 2.0)! 😎
Though not the weights... 🙃
github.com/google-deepmind/a…
Jesse Weller retweeted
We also found a scaling law that explains (p=0.03) Ab-Ag accuracy across model lineages - from Boltz to AF3 and beyond. 1.8x-ing AF3 performance (to match IsoDDE) would seem to require e.g. 10x more distillation and 16x compute. (7/8)
Jesse Weller retweeted
Replying to @mihirbafna14
@mihirbafna14 and I are excited to introduce Promera, a co-folding and design model with
• best-in-class binder filtering
• nanobody design with in-silico success rates matching hallucination
• case studies on hantavirus epitope targeting and GPCR agonism (1/8)
Jesse Weller retweeted
AF3 source code is now open source (aka Apache 2.0)! 😎
Though not the weights... 🙃
github.com/google-deepmind/a…



























