🔬ESMFold2: The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub

Large-scale protein language models, specifically the Evolutionary Scale Modeling (ESM) family, are revolutionizing biological research by treating protein sequences as a language to predict evolutionary constraints. By training on vast metagenomic datasets rather than relying on traditional multiple sequence alignments, these models achieve superior representational fidelity and enable the design of complex molecules like antibodies and protein binders. Mechanistic interpretability techniques, such as sparse autoencoders, reveal that these models autonomously learn hierarchical biological features, ranging from basic biochemical properties to functional motifs. Biohub’s current research focuses on scaling this paradigm to the cellular level, integrating frontier artificial intelligence with high-throughput experimental data generation. This approach aims to create predictive digital representations of biology, ultimately transforming scientific discovery by allowing researchers to simulate complex physiological interventions and accelerate the development of novel therapeutic modalities.