Intology's Locus turns autonomous AI R&D on molecular discovery: matched compute, 1.5–2.8× fewer antibody candidates

The autonomous discovery system Locus is now applied to molecular discovery rather than only language-model post-training and ML research engineering; it must propose and run experiments itself with no starting model, limited data and noisy evaluation. The prior condition is the existing state-of-the-art methods from BioHub and Boltz and frontier LLM agents such as Codex and Claude Code, which struggle to generalize from the small initial dataset and to overcome the high evaluation variance it imposes. On a proprietary commercial molecular property prediction task, Locus trained a model from scratch and its final model, scored once on a fully held-out test set, beat the baselines and met the task's predefined performance threshold; the task reports only relative metrics to preserve confidentiality.

Locus works by running several experiments in parallel and sampling new experiments from evaluation feedback, using purpose-built infrastructure for concurrent experiments and high resource utilization. It splits problem difficulty into three axes, evaluation cost, evaluation variance and the size of the experimental search space, and organizes experiments around them. On Boltz-2 adaptation across six targets (BChE, CYP19A1, EGFR, MAO-B, MAP4K1 and PLK4), each run received 12 H100 GPU-hours, with ChEMBL data split into temporally and structurally held-out validation and test sets.

For antibody design, Locus writes and revises workflow code around given generative and scoring models, covering sequence initialization, mutation strategies, structural objectives and candidate ranking. The BoltzGen paper's workflow runs an expensive generation campaign followed by strict filtering, while the filtering criteria Locus learns retain more useful candidates, so it reaches stronger binders from fewer generated candidates. In the ESMC/ESMFold2 setting, Locus's mean ipSAE for the top two candidates per target was 0.797, against 0.758 for Codex and 0.704 for the ESMFold2 baseline; Locus's candidate panel scored higher than ESMFold2 in 90.3% of matched target-generation comparisons.

The demonstrated scope is three computational tasks: a proprietary commercial molecular property prediction task, Boltz-2 potency adaptation on six ChEMBL targets, and in silico antibody hit-finding built on BoltzGen/Boltz-2 and ESMC/ESMFold2. Antibody binding strength is scored by OpenDDE's ipSAE as a proxy, and the source states that fully validating the binding affinity of the designed antibodies will require wet-lab testing.

Antibody binding scores are structure-prediction-confidence proxies rather than measured affinities; the molecular property prediction task reports performance only in relative form because of commercial confidentiality; and in the antibody design tasks, Claude Opus models, GPT-6 Astra and later Sol-family models triggered biological safety classifiers and could not be used for optimization, so both Locus and the baseline used GPT-5.6 Sol High.