Antibodies designed with fewer iterations: Locus automates molecular research

Locus, developed by Intology, automates molecule discovery and the adaptation of biological models, with promising results still to be confirmed in the lab.

Identifying a promising molecule often requires multiplying experiments, with no certainty that the results obtained on a small number of compounds will reproduce on others. Specialized AI models can reduce this exploration, but adapting them to limited experimental data remains difficult. A new series of works led by Intology shows that an autonomous research system can improve the performance of these models by choosing for itself which experiments to run and which methods to test.

This system, called Locus, operates in three areas of molecular discovery: chemical property prediction, estimation of the potency of small molecules, and antibody design. Rather than applying a predefined procedure, it formulates several hypotheses, carries out computational experiments in parallel, and adapts its research based on the results obtained. The work published by Intology reports performance superior to that of several specialized methods and agents based on Claude Code or Codex, in the configurations evaluated. Nearly one million simulated molecules to offset the lack of data

The first experiment concerns a commercial task of molecular property prediction whose data remain confidential. In this type of problem, certain properties can require several hours of processor-based simulation for a single molecule. The challenge is to develop a model capable of providing sufficiently accurate estimates while reducing the time needed for each prediction.

The researchers are initially provided with a small set of annotated molecules and a validation set. No specialized model is supplied to them at the outset. Faced with the same constraints, Claude Code, using Opus 5.5 Max, and Codex, using GPT-6 Astra High, obtain models whose generalization remains limited according to the published tests.

The strategy followed by Locus consists in progressively producing new training data from simulations. The system thus assembles a set approaching one million annotated molecules, whereas Claude Code generates around 100,000 examples and Codex does not exploit this possibility in the reported experiment.

This increase in data is accompanied by an exploration of different model sizes, up to several billion parameters. The results presented indicate a prediction error lower than that of competing solutions on a test set kept separate from training. The final model also reaches the performance threshold set for this task.

The absolute figures are nevertheless not published, due to the proprietary nature of the data. The comparisons rely on relative measures, which limits independent assessment of the scale of the reported progress. Adapting Boltz-2 to molecules it knows poorly

The second series of experiments focuses on predicting the potency of small molecules, an important piece of information in the early phases of pharmaceutical research. The researchers are more specifically interested in IC₅₀, a measure of the concentration required to inhibit a given biological activity by half.

A general-purpose model such as Boltz-2 can estimate certain properties of interactions between proteins and molecules. Its results, however, deteriorate when the compounds studied or the biological targets differ greatly from the examples seen during training. With few experimental measurements available, a direct adjustment also risks producing a model overly fitted to the initial data.

The experiment covers six biological targets from ChEMBL, a database devoted to bioactive molecules. For each one, the researchers have at least 120 ligands associated with experimental IC₅₀ measurements. The data are split between training, validation, and test, with the compounds reserved for final evaluation remaining inaccessible to the systems during their research.

With a budget of 12 H100 GPU hours per run, Locus achieves a mean Spearman correlation of 0.622 between predictions and experimental measurements. Codex obtains 0.569, Claude Code 0.450, and the Boltz-2.1 API 0.225. This metric mainly measures the ability to correctly rank molecules by potency, rather than the absolute accuracy of each predicted value.

The results also show a smaller gap between validation performance and test-set performance for Locus. The system explores more adaptation strategies, from fine-tuning certain parts of Boltz-2 to combining its predictions with complementary chemical features. This experimental diversity appears to help limit the overfitting observed with restricted data. Designing antibodies with fewer candidates

The third part of the study shifts the problem from prediction to the design of new molecules. In antibody research, a common approach consists in generating a large number of candidates, estimating their binding ability to a biological target, and then applying several filters to select the most promising ones. This procedure can consume substantial computing resources without guaranteeing that the best candidates will be retained.

The researchers asked Locus to build its own design procedures using existing models, notably BoltzGen and Boltz-2, then ESMC and ESMFold2. It could modify the generation steps, selection criteria, and ranking methods without having to systematically retrain the underlying models.

In the experiments combining BoltzGen and Boltz-2, ten biological targets are set aside for final evaluation. The procedures developed by Locus produce candidates that are better ranked according to the chosen criterion, while exploring smaller sets of molecules than the reference methods. The researchers attribute this result in particular to selection rules less likely to eliminate interesting candidates prematurely.

A second series of tests, based on ESMC and ESMFold2, yields an average score of 0.797 for the two best antibodies for each target, versus 0.758 for Codex and 0.704 for the reference ESMFold2 method. In this protocol, Locus’s solution matches the best score in the reference set while examining only 16.7% of the number of candidates, that is,