Astellas trained its own antibody language models on NVIDIA BioNeMo to predict antibody properties before measurement.

Standard protein language models underperformed on antibody-specific properties, so Astellas built astABpLM using BioNeMo's ESM-1nv framework, with separate heavy-chain and light-chain models trained on 2.4 billion antibody sequences from the OAS database. It also built a conformational screening workflow using torsional diffusion models. The page reports improved accuracy in predicting antibody physical properties, a 50-60x speedup in conformational screening (15 seconds versus several hours to a day), and training times of about 65 hours (heavy chain) and 37 hours (light chain) on one DGX A100.

NVIDIA is a computing and AI company whose BioNeMo and Clara platforms provide models and accelerated tools for drug discovery, genomics and medical imaging.

This sits in the Build stage of the AI use-case lifecycle: Named examples from the foundation-model labs and platforms pharma teams use to actually construct, train and generate the model or application.