Intology Locus: an AI research system for molecules

What happened?

On November 3, Intology showed how its autonomous AI research system Locus develops models for drug discovery (Intology). Locus designs, trains and evaluates models on its own without human intervention. Intology compared it with Codex and Claude Code.

The results

• Binding affinity: Locus adapted Boltz-2 to six target proteins (including EGFR, MAO-B and PLK4) in 12 H100-hours. It reached a Spearman correlation of 0.622, versus 0.225 for the Boltz-2.1 API, 0.450 for Claude Code (Opus 5 Max) and 0.569 for Codex (GPT-6 Astra Max).
• Property prediction: on a proprietary task it scaled the solution to about 1 million synthetic labels and billions of parameters.
• Antibody workflows: it needed 1.5 to 2.8 times fewer candidates to reach the same result.
• ESM-based workflow: a top-2 score of 0.797 versus 0.758 for Codex and 0.704 for ESMFold2, with about 6x better candidate efficiency.

The limits

The antibody results rely on an in silico proxy metric (ipSAE); wet lab validation is still pending. One notable detail: Claude Opus and GPT-6 Astra triggered biosecurity classifiers on the antibody tasks, so Intology used GPT-5.6 Sol High there.

Why it matters

AI that does AI research has mostly been shown on machine learning benchmarks. Here it builds better models than general coding agents in a real industry field, molecular design. Experts from Benchling and Boltz praised the results. The real test will be the wet lab, and how much safety classifiers end up limiting this kind of work.