DIGITAL BIOLOGY
Features
BMS partners with Chai Discovery to find therapeutic antibody candidates
The agreement gives BMS use of Chai's molecular folding and design models across its portfolio. The collaboration is intended to support antibody candidate discovery; it does not disclose specific therapeutic targets or the number of programs involved. Chai announced the partnership without financial terms.
Sanofi and AWS describe SWEL, a system connecting scientific data with AI workflows
SWEL sits above a scientific data layer containing more than 20 petabytes from hundreds of connected lab instruments. Sanofi aspires to reduce the number of molecules needing wet lab validation by 50% and deploy AI workflows ten times faster; these are goals, not measured results. The system has scaled to support more than 50 scientific workflows.
Benchling introduces chemistry reaction planning and Inductive Bio's ADMET models
Chemistry Reactions models single-step synthesis and links molecules to experiments, biological results and computational predictions in Benchling. It recalculates stoichiometry as conditions change. Inductive Bio's models let chemists predict absorption, distribution, metabolism, excretion and toxicity before physical testing and compare predictions with results. Benchling's chemistry capabilities remain in limited availability.
SandboxAQ launches AQPotency to predict potency without a solved protein structure
The model runs on ordinary computing hardware and reports confidence alongside a flag indicating whether a target falls within its reliable range. It is available through Claude via Model Context Protocol and SandboxAQ's website. Pricing starts at $1 per 1,000 comparisons. SandboxAQ says eight customer programs have used the model with experimentally validated impact.
Blinded antibody benchmark finds most models worse than random on one ranking task
AIntibody experimentally tested 511 AI-designed or predicted antibodies from 29 organizations across three tasks targeting the SARS-CoV-2 receptor-binding domain. Several groups produced developable antibodies with affinities below 100 picomolar, but success did not transfer between tasks. When ranking high-affinity clones within sequence clusters, all but one model performed worse than random selection. Out-of-library design results varied widely.