Twelve years ago, our CEO Andrew Chang joined Genentech as a postdoc and fell in love with library screening, which lets humans explore massive protein sequence-function space. Structure prediction has changed what is possible in this field, but structure alone does not always reach the functions real medicine requires. That conviction, that library screening, next-generation sequencing and language modeling close the gap, is why DeepSeq exists. This month brought it full circle.

We were glad to contribute to a Genentech-led paper in Nature Communications on designing peptides for cancer imaging. Starting from an ML-guided combinatorial library, the team developed disulfide-constrained peptides with subnanomolar affinity for LRRC15, a marker on cancer-associated fibroblasts. PET imaging with one of them showed specific tumor accumulation in a pancreatic cancer model.

One result is worth dwelling on. Crystal structures showed the peptides evolved a distinct binding conformation to engage a flat epitope on LRRC15. Flat surfaces are among the hardest targets to design against, and the molecules that solved it came out of functional selection, not structural prediction. That is the Function-First case in a single experiment.

This also completes a progression we have been building toward in public. Our PLOS Computational Biology work addressed foldability and yield. Our Pharmaceuticals work addressed rapid affinity maturation. This paper carries ML-designed molecules into a living animal.

AI is only as good as the data, and data is only as good as the people who design and run the experiments. Full credit to the Genentech team.

Read the paper: https://doi.org/10.1038/s41467-026-73845-z

‍