Wells split the year between building open-source AI methods for protein design and applying the tools to design novel proteins. He released ProFam, a protein family language model for functional protein design, alongside one of the largest open datasets of protein families, then pivoted to structure-based protein design and extending these methods to multi-state design: building binder-triggered switches into proteins. The first switches out of that pipeline, allosteric enzymes designed as heart-attack biosensors, reached the lab this year: with the first round of testing validating the experimental assay. Even though there were no successful designs in the first round, the negative data is improving the computational design pipeline for future rounds.

ProFam-1 shipped open-source with code, weights, and a dataset downloaded over 400 times. Our work on binder design resulted in lab-confirmed hits against 2 therapeutically relevant protein targets (Nipah Virus and 15PGDH).

The first switches out of that pipeline, allosteric enzymes designed as heart-attack biosensors, reached the lab this year: with the first round of testing validating the experimental assay. Even though there were no successful designs in the first round, the negative data is improving the computational design pipeline for future rounds.

Protein language models, ProFam included, are good at lead optimisation: taking a protein that already works and making it better. But the highest-value problems in protein design are not about improving what exists; they are about unlocking capabilities that have no precedent in nature. For those, structure-based models are the more powerful paradigm.

Jude holds a PhD in ML for Drug Discovery from UCL. He has worked on protein informatics and structural generation via advanced ML. Jude has published 8+ publications or pre-prints in the last 3 years.