I'm a Founding Research Engineer at Topos Bio, where I joined as employee #2 and the first ML hire. I build generative models that learn distributions, not single answers, from compute-heavy molecular simulations and experimental data.
Research
Topos-2 & Topos-Bind2026
Ensemble generation across the order-disorder spectrum, extended to small-molecule binding.
Tomas Salgado, Andre Graubner, Malhar Kute, and the Topos Bio team
Ensembles generated by Topos-2 and Topos-Bind
Topos-2 generates the many shapes a protein can take rather than a single structure, now across the order-disorder spectrum, from well-folded to fully disordered. It sets a new state of the art on the independent benchmark PeptoneBench, outperforming leading protein models from Google DeepMind, Microsoft, and Meta. Topos-Bind is, to our knowledge, the first generative model of disordered proteins bound to small molecules, trained in part on Topos-DB, the largest all-atom protein simulation dataset by number of distinct systems (100,000+). I led Topos-Bind, contributed core engineering to Topos-2, and wrote the technical blog post.
An all-atom foundation model for intrinsically disordered proteins.
Tomas Salgado*, Andre Graubner*, Scott Leishman*, Malhar Kute, Yiran He, Benjamin Rousseau, Connor Bybee, Anthony Giannetti, Xiaotong Lu, Anne Pipathsouk, Jake Drummond, Robert Galemmo, Fernando Martinez, Amir Khosrowshahi, Ryan Zarcone, Grant Rotskoff†
Conformational ensemble generated by Topos-1
Topos-1 outperformed AlphaFold2, Boltz-2, Chai-1, and BioEmu on key experimental benchmarks, cutting error in predicted protein size (radius of gyration) by 43–79% while running ~1,000× faster than molecular dynamics. As project lead, I built the first prototypes from scratch, co-led pretraining, and drove the research through the technical report.