UNLOCK 2026 Biology For AI The Next Breakthroughs Are Methods, Not Models Henry Lee, Cultivarium

Author: Henry Lee, Cultivarium Collection: UNLOCK-2026-Sessions Source: Biology For AI: The Next Breakthroughs Are Methods, Not Models

Summary

Henry Lee, representing Cultivarium (a nonprofit research institution in Boston, Massachusetts), argues that the dominant bottleneck for AI in biology is not model architecture but biological data generation methods. His core thesis: approximately 1 trillion species exist on Earth, biological science has meaningfully characterized only about 200 organisms, approximately 5 model organisms receive the vast majority of lab attention, and even after 100+ years of studying E. coli, 40% of its genes remain of unknown function. For most organisms, the experiments needed to generate training data simply do not yet exist, so bringing AI into the picture "just yet" will not help.

Lee presents three vignettes of Cultivarium's technical approach to closing this gap. First, GenomeSPOT — a machine learning model that predicts organism growth conditions (oxygen, temperature, salinity, pH) from genome sequence alone, applied to 52,000 organisms across 3,300 environments using all known metagenomic sequences. Second, a high-throughput molecular toolkit that screens thousands of combinations of DNA parts per organism, pooling organisms to test communities and generate hundreds of thousands to millions of data points per experiment; currently applied to 5,600 bacteria across 12 bacterial phyla (of ~90 known). Third, a technology stack for lab automation including Prism (AR glasses recording scientist experiments for video-language model documentation), Hermes (an AI planning and execution system for experiments), and the "Work Cell Armadillo" built on Medra hardware — with Leaf Labs as an additional hardware partner.

Lee announces a preview of Primordia, a biological world model tracking DNA transfer between organisms. The bacteria-only portion of Primordia currently covers 143,000 microbes and 14 million DNA exchanges, with edges labeled by transfer mechanism (viral or bacterial conjugation) and annotated by biome (human gut, animal gut, plant, marine, freshwater, soil). An antibiotic resistance application is demonstrated: mapping how ESKAPE (Enterococcus, Staphylococcus, Klebsiella, Acinetobacter, Pseudomonas, Enterobacter) pathogens exchange resistance genes with other bacteria in different biomes.