Descriptions:
Alex Rives — head of science at Biohub and the researcher who led training of the first transformer language model for protein biology at Meta AI in 2018 — joins the Latent Space AI for Science podcast to discuss scaling laws in computational biology, the ESMC model, and his long-held conviction that the “bitter lesson” (scale and compute beating hand-crafted methods) will reshape protein science just as it reshaped natural language AI.
Rives explains the theoretical basis for why next-token prediction transfers to proteins: evolution imposes co-evolutionary constraints on amino acid sequences — residues in structural contact cannot be chosen independently — creating exactly the kind of statistical regularity that transformer models capture. Over successive model generations, each roughly an order of magnitude larger, his team observed emergent biological capabilities that tracked the scaling curve. The conversation goes deep on ESMC, EvolutionaryScale’s world-modeling approach, where the model acts as a searchable latent space for protein molecules satisfying arbitrary design criteria.
The most technically substantive section covers ESMC’s recent results designing antibodies and SCFVs (single-chain antibody fragments) — a class of therapeutic significantly harder to design than mini-binders, and largely out of reach for AlphaFold-style methods because antibody evolution favors diversity rather than conservation. Rives reports reaching therapeutic-range binding affinities in a small number of experimental trials. The discussion also explains why the AlphaFold paradigm struggles with antibodies (absence of multiple sequence alignments, inverted evolutionary pressure) and how ESMC’s generative search approach sidesteps that limitation.
📺 Source: Latent Space · Published May 27, 2026
🏷️ Format: Podcast
