Language models now design proteins for drug targets in lab tests

Anthropic's Claude models achieved binding success rates exceeding industry norms in early-stage protein design experiments, with potential to slash drug development costs for protein-based therapies.

Anthropic conducted two experiments using Claude models to design proteins for drug targets. In the first, 1,320 protein designs were tested in labs, yielding 354 that bound to targets (26.8% hit rate). Claude’s models achieved 49% binding success when ranked by their own criteria, significantly outpacing industry standards of 10–15% (per proteinbase.com). For specific targets like TREM2, 72 out of 90 designs bound; for RBX1, Claude’s best design bound at 3.9 nM KD—10 times tighter than the contest winner’s 45 nM KD. Compute costs for multi-target campaigns were $50,000 (vs. $10,000 per single target), with validation by contract labs Adaptyv Bio and Twist Bioscience.

The mechanism involves Claude analyzing structural data from NMR and LC-MS in minutes—23 minutes for NMR, 19 for LC-MS—then generating protein designs that bind targets. This bypasses traditional trial-and-error methods, accelerating early-stage discovery while maintaining lab validation.

For healthcare, this could cut costs for protein-based therapies by 50%+ by reducing the need for expensive, slow experimental iterations. Faster target-specific designs mean more accessible treatments for conditions like autoimmune disorders where protein targets are critical. However, results vary widely by target type: some designs (e.g., MBP) showed zero binding, while TNFα designs traced to only four protein scaffolds.

Independent review of results is pending. What matters next: whether Claude’s designs hold up under full clinical testing and whether the 50%+ cost reduction translates to real-world access for patients in resource-limited settings. The source notes results vary by target, so this isn’t a universal solution but a promising step for specific therapeutic areas.

Source: The Decoder