How to Train an AI For Science

Author: Johnny Hu (moderator), Yusuf Roohani, Daniel Balsam, Peyton Greenside Collection: UNLOCK-2026-Sessions

Summary

This panel discusses the practical challenges of building AI models for biological science, featuring three different perspectives: academic foundation model development (Yusuf Roohani, Arc Institute), mechanistic interpretability tooling (Daniel Balsam, Goodfire), and applied AI-first antibody and protein design (Peyton Greenside, Big Hat Biosciences). The moderator is Johnny Hu, Principal at Menlo Ventures, a 50-year-old VC firm that is an investor in Zara, Axium, Measra, and Goodfire.

Yusuf Roohani leads AI/ML at Arc Institute and recently published the STACK architecture paper, which combines observational data (cell lines, patient samples, primary cells) with dynamic/mechanistic/causal information. He argues the AlphaFold-style approach — scaling one modality of data — will not work for most biology, and that the right scaling axis for biology remains an open and exciting question. He also describes SC Base Count, a paper using AI agents to reprocess the NIH's Sequence Read Archive (SRA), the largest repository of transcriptomics data, into a uniformly processed format released publicly. Arc Institute (nonprofit) open sources its models; they released EVO 2, a genomics foundation model, open source.

Daniel Balsam of Goodfire builds mechanistic interpretability tooling — an "AI debugger" that maps the internal representations of biology foundation models to human-understandable biological concepts. Goodfire collaborated with Arc Institute and Mayo Clinic on using internal representations of EVO 2 to diagnose pathogenicity from ClinVar variants — even though EVO 2 was trained only on the human reference genome (not variant data), it had latent capability to stratify pathogenicity. His company's thesis is that models should be designed like written software: debuggable, editable, and understandable, not grown like organic matter.

Peyton Greenside co-founded Big Hat Biosciences, an AI-driven protein and antibody design company, and focuses on generating fit-for-purpose experimental data to fine-tune models for specific drug design problems. She argues that data quality matters more than data quantity: reducing a dataset to 25% of its original size by applying stringent quality filters has boosted model performance relative to training on the full noisy dataset. Big Hat focuses on closing the lab-to-model loop as quickly as possible, treating data generation speed as the primary constraint. They use publicly available structural and sequence models but build proprietary fine-tuned models on top of them, treating their data pipeline as the core differentiator.