Sebastian Ibarraran

Machine learning has improved transferable coarse-grained models, but the best architectures are nearly impossible to train due to noisy objectives. A simple strategy called mean force matching dramatically reduces data costs (↓87%) and enables scaling learned representations.

Efficient, Few-shot Directed Evolution with Energy Rank Alignment 1. A new method called Energy Rank Alignment (ERA) enables highly efficient protein engineering by adapting large pre-trained protein language models using minimal experimental data. 2. Unlike previous approaches that rely on simple models due to sparse data constraints, ERA leverages the strong inductive biases of ESM3-1.4B, a 1.4 billion parameter protein language model, to navigate complex fitness landscapes. 3. The key innovation is using quantitative experimental rankings rather than just binary preferences, allowing the model to preserve relative fitness magnitudes while learning from small batches of just 96 sequences per round. 4. ERA outperforms existing methods including MLDE, ALDE, EVOLVEpro, and Direct Preference Optimization across five diverse combinatorial fitness landscapes involving antibiotic resistance, antibody binding, and enzymatic activity. 5. The method achieves state-of-the-art performance with only 384 total samples across four rounds, successfully finding global optima even in landscapes with strong epistatic effects and rugged topography. 6. Surprisingly, adding structural conditioning or thermostability pre-training did not improve performance, suggesting that pure sequence-based adaptation is sufficient for effective directed evolution. 7. The adapted models maintain sequence diversity while shifting probability mass toward high-fitness regions by several orders of magnitude, making them interpretable and useful for understanding biophysical requirements. 8. This work establishes a compelling interface between foundation models and experimental design, demonstrating how post-training algorithms from statistical physics can solve real biological optimization problems. 💻Code: github.com/rotskoff-group… 📜Paper: biorxiv.org/content/10.648… #ProteinEngineering #DirectedEvolution #MachineLearning #Bioinformatics #ProteinDesign #ESM3 #ComputationalBiology #FewShotLearning

Big thanks to my co-authors: @shriramc1 , @FrankWho1050502 , and @grantrotskoff . Would love feedback or discussion from folks working on protein design, directed evolution, protein language models, or preference optimization.

Across various fitness landscapes, ERA • Achieves state-of-the-art or competitive performance with small amounts of experimental data • Outperforms active-learning and regression-based ML-assisted DE baselines • Avoids mode collapse; remains interpretable at the residue level

Excited to share our preprint on using energy rank alignment (ERA), our physics-inspired preference optimization algorithm, to guide protein language models for directed evolution! Check out our project page with links to code, data, and the paper: rotskoff-group.github.io/era-directed-e…

Excited to share that our most recent preprint on automating NMR structure elucidation is now available on arXiv! check it out now at arxiv.org/abs/2512.18531!

I'm hiring a postdoc with a flexible start, sooner is better! Come work with us at the interface of machine learning, biophysics, and nonequilibrium physics. Interested? Send me a CV and a short summary of why you think you'd be a good fit. statmech.stanford.edu

Had a great time presenting our work on chemical language model alignment at NeurIPS 2025! It was a wonderful week of science and fascinating conversations, and if you'd like to learn more about what we're up to, please reach out!

Had an awesome time presenting our chemical and protein language model alignment work at NeurIPS! Really appreciated all the insightful feedback and new connections, and I’m looking forward to continuing our efforts in this and other directions for molecular design

The internship posting is now live! The role will focus on topics related to our ongoing effort to post-train LLMs to be used in active small-molecule drug discovery campaigns. roche.wd3.myworkdayjobs.com/ROG-A2O-GENE/j…

Separately, I'll be presenting some of my PhD work with @s_ibarraran and @FrankWho1050502 on Wednesday. Please stop by if you'd like to learn more about aligning chemical language models! neurips.cc/virtual/2025/l…

I'll be at NeurIPS this week! Our team at Prescient Design/Genentech is working on post-training LLMs for small-molecule drug discovery applications. I'll also be opening an internship for next summer related to this effort. If you're interested in learning more, I'd love to chat

@eliaswehbe @thinkymachines This is still a bit unclear, but we think this is likely related to variability in the complexity of compounds across different batches.

@s_ibarraran @thinkymachines Why are there accuracy peaks at specific steps across all models? What causes that behavior?

While reinforcement learning has been demonstrated to improve LLM performance on mathematical reasoning tasks, currently, there is far less evidence of performant scientific reasoning models. Using Tinker by @thinkymachines, we were able to rapidly train a variety of models on -