Beacon-2 turns human efficacious dose into a computable reward: agent cuts predicted dose 17-fold

Beacon-2 replaces potency alone with human efficacious dose as the preclinical optimization target, giving agents a single number that integrates potency and pharmacokinetics to iterate against. AI drug discovery had concentrated on predicting potency, and optimizing an incomplete reward function invites reward hacking; multi-parameter scores or hard property cutoffs risk optimizing the wrong goal or discarding promising candidates. On 325 compounds in the ExpansionRx blind challenge with complete measured ADME data, Beacon-2's fully blind predicted doses matched doses derived from measured in vitro data through the PK/PD module, with a Spearman rank correlation of 0.79 and an average fold error of 2.4x.

The system chains three configurable modules: an ADMET-PK module predicts in vitro properties such as permeability and hepatocyte stability plus preclinical PK properties where in vivo data exist, a potency module gives structure-informed potency estimates, and a PK/PD module combines both through physiologically informed mechanistic models into a dose. The potency module offers two routes: automated free energy perturbation on in-house infrastructure with auto-posing workflows, or a fine-tunable co-folding affinity model currently in alpha. Across 20 partner programs on the Compass platform spanning different chemical spaces and stages, Beacon-2 reached an average Spearman rank correlation of 0.61; when two compounds in a program differed by 3x or more in measured dose, it selected the lower-dose design over 81% of the time.

In a public case study on the SARS-CoV-2 Mac1 inhibitor AVI-6451, the medicinal chemistry agent Indy with Beacon-2 cut predicted daily dose roughly 17-fold within a budget of 5 rounds of 10 compounds each, with the best compound, IB-47, found in the final round. The same agent starting from the same compound with the same budget but optimizing potency alone saw dose stagnate or worsen, and its most potent predicted compound, IB-93, had a predicted dose 17-fold worse than the original. Three checks on the top 10 compounds from each run favored dose optimization: mean QED of 0.64 versus 0.52, three medicinal chemists choosing the Beacon-2 compound in 87% of 10 blinded pairs, and a recomputation with a rat in vivo clearance and volume-of-distribution PK configuration giving 20x lower average dose versus 5x.

The demonstrated scope is a retrospective blind prediction on the ExpansionRx public challenge dataset, prospective tracking across 20 live partner programs, and one computational case study starting from AVI-6451; both dose evaluations used an assumed 10 nM reference potency. It suits medicinal and computational chemists interested in lead optimization and reward design for AI agents.

Projected doses ultimately need wet-lab validation, and the team says it is working toward letting agents equipped with Beacon-2 iterate toward low dose with the lab in the loop. The co-folding affinity model in the potency module is currently in alpha. The public blind test used an assumed 10 nM reference potency because potency data were not part of the challenge, and the 20-program dose projections used the same assumed value.