Biomedical Artificial Intelligence · global
Identifying Drug Interaction Risks Earlier: Inductive Ranks First in Blind CYP Inhibition Challenge
A blind test involving 750 compounds and four metabolic enzymes provides concrete validation for AI-assisted drug design. But the top two entries fell within the same statistical performance tier, and the test molecules closely resembled the training data, requiring careful interpretation of the results’ applicability.
Even if a drug candidate acts on its therapeutic target, it may encounter obstacles by interfering with the metabolism of other drugs. Anticipating these risks during molecular design could give researchers an opportunity to adjust structures earlier. On November 6, Inductive announced that it had placed first in OpenADMET’s blind CYP direct inhibition challenge, adding a testable result to the evidence on whether AI can assist with this work.
The challenge required participants to predict the direct inhibitory potency of 750 compounds against four enzymes: CYP3A4, CYP2D6, CYP2C9, and CYP1A2. The measured results remained hidden until the competition ended. These cytochrome P450 enzymes participate in the metabolism of many small-molecule drugs. When one is inhibited, concomitant drugs that depend on it for clearance may accumulate in the body, increasing exposure and potentially raising the risk of toxicity.
Results released by the organizers on November 5 showed that the direct inhibition track attracted 271 participants, with 103 eligible entries included in the final leaderboard. Inductive’s model, submitted under the name Whippoorwill, achieved an overall ST-RAE error metric of 0.341 ± 0.010 and a Spearman rank correlation coefficient of 0.815. However, the second-place entry had an error metric of 0.343 ± 0.010, and both were placed in the same statistical performance tier. Ranking first therefore does not establish superiority over the second-place entry. Final rankings were calculated using the complete test set, while the live leaderboard during the competition used only half of it. Entries also had to meet eligibility requirements concerning model reports, accounts, and other criteria.
This validation rests on laboratory measurements rather than clinical outcomes. According to the organizers, Octant and the Fraser laboratory at the University of California, San Francisco, jointly generated the data using biochemical enzyme assays. Reactions for three enzymes were read using fluorescence signals, while CYP2D6 substrate depletion was measured by mass spectrometry. The winning task involved predicting direct inhibitory potency and was distinct from another track identifying whether inhibition increases over time. The results therefore cannot be interpreted as covering all mechanisms of CYP interactions.
Inductive’s technical report states that the team combined two-dimensional and three-dimensional graph neural networks with a TabM network using pretrained molecular embeddings. The three-dimensional models used low-energy molecular conformations and did not use CYP protein structures. In addition to measurements supplied for the competition, the training data included public and proprietary data on CYP, the lipophilicity measure LogD, and the acid dissociation constant pKa. The result therefore reflects the combined performance of the models, data resources, and training strategies, and cannot be attributed to any single algorithm alone.
The selection of the test set also defines what the results mean. The organizers selected 25 potent inhibitory compounds for each of three enzymes, then purchased ten structurally similar molecules for each, creating a set of 750 test compounds that were all measured against the four enzymes. Inductive also explained in its report that these molecules were analogs of 75 training compounds, prompting the team to design validation splits specifically to simulate this similarity distribution. This resembles the setting in which medicinal chemists continually modify existing molecules, but it does not yet establish whether the models are equally reliable when encountering unfamiliar chemical scaffolds.
For drug discovery and development, this blind test supports using AI to prioritize experiments and compare the metabolic risks of closely related molecules. It does not demonstrate clinical efficacy for drug candidates, nor can it directly determine whether concomitant medications are safe for patients. Moving from biochemical inhibitory potency to interactions in humans requires additional evidence, including actual drug exposure. The organizers have unblinded and released the complete inhibition dataset, allowing other teams to reanalyze the models’ strengths and errors and providing a shared starting point for further validation.