In drug discovery, the data wins the benchmark

Four in a row

Inductive has won the OpenADMET CYP Inhibition Challenge, marking its fourth consecutive first-place finish in major blind ADMET prediction competitions. The company ranked ahead of more than 270 participants, including NVIDIA, which placed third, and Recursion, which placed fourth. The challenge required predicting pIC50 against four CYP enzymes — CYP3A4, CYP2D6, CYP2C9 and CYP1A2 — across 750 compounds with withheld measurements. CYP enzymes metabolise the majority of small molecule drugs, so the prediction task sits directly on the path to a usable medicine. m:…4082

Inductive's Beacon models are powered by a pre-competitive data consortium and have been tested on 19 critical drug discovery endpoints. The company's previous wins were the ASAP-Polaris-OpenADMET Challenge in 2025, the OpenADMET-ExpansionRx Challenge in February 2026 and the OpenADMET PXR Challenge in July 2026. Ben Birnbaum, at Inductive, states that data is the bottleneck for AI in chemistry. Four straight wins with the same explanation is a coherent story: the moat is the dataset, not the architecture. m:…4082

Context at a million base pairs

Deep Genomics developed DeepRNAi to predict siRNA activity and off-target effects across the transcriptome, and the model uses Biological Foundation Models to incorporate global genomic context. The company's flagship foundation models include BigRNA, Enigma and REPRESS, and they take into account up to a million base pairs of sequence context. Incorporating these models nearly doubles the performance of DeepRNAi, according to the company, and its performance rivals experimental screening methods. m:…8824

The case studies show why context matters. DeepRNAi correctly predicted that Alnylam's siRNAs targeting the distal 3' UTR of the PLG transcript would be inactive. A common variant in the IRF5 gene, with a 58% global allele frequency, creates a proximal polyadenylation signal, and DeepRNAi with foundation models identifies that siRNAs tiling the IRF5 variant region are unlikely to work for a large portion of the population. A model that only sees the local sequence would miss both. Deep Genomics says it will soon release the DeepRNAi pre-print. m:…8824

The pharma deals

Chai Discovery announced a collaboration with Roche to integrate its AI-driven molecular design platform into Roche's drug discovery portfolio, using its AI models and platform for drug discovery programs. The partnership aims to accelerate the development of protein-based therapeutics across multiple therapeutic areas. Chai-3 is Chai Discovery's latest AI model designed for protein-based therapeutics, and the company says it has announced multiple collaborations with leading pharmaceutical companies throughout 2026. Chai Discovery was founded in 2024 by Joshua Meier, Jack Dent, Matthew McPartlon and Jacques Boitreaud. m:…5039

The preprint in this week's set takes a different angle on the same problem. FoldAbS repurposes frozen protein folding models as foundation encoders for antibody screening, capturing interaction dynamics beyond basic structure prediction. A unified screening architecture can evaluate specificity, affinity and developability within a single model, and FoldAbS-Px achieved performance gains of 13.1% in specificity, 9.2% in affinity and 6.2% in developability compared with baseline methods. The idea is to reuse what protein folding models already learned rather than training a new model from scratch for each property. m:…3910

What this means

The week's drug discovery news converges on one claim: the winners are the teams with the best data and the best reuse of existing models. Inductive wins on a consortium dataset. Deep Genomics wins on a million base pairs of context. Chai wins pharma deals with a model built for protein therapeutics. FoldAbS wins by repurposing a folding model instead of training a new one. None of these is a new architecture announcement. All four are about what the model is fed and what it is built on. m:…4082 m:…8824 m:…5039 m:…3910

For founders and investors in AI drug discovery, the implication is to diligence the data before the model. Inductive's Josh Haimson states that its tools help chemists make higher-quality decisions, which is a modest claim backed by four blinded wins. The field's credibility now rests on blinded challenges and withheld measurements, not on self-reported benchmarks. That is a healthy standard, and it is one the rest of AI could learn from. m:…4082