Unifying Disjoint Phenotypic Contexts: A Multimodal Soft Contrastive Approach to Identify DILI Activity Cliffs

<title>Abstract</title> <p>Predicting clinical toxicity, such as drug-induced liver injury (DILI), remains a central challenge in drug discovery. This difficulty is compounded by the presence of "activity cliffs"---structurally similar compounds with divergent safety profiles that are notoriously difficult to distinguish. Traditional structure-based representations systematically fail on these edge cases, prompting the integration of functional biological readouts (e.g., morphology and transcriptomics) to provide phenotypic context. However, current multimodal approaches are bottlenecked by two severe limitations: they require complete multi-assay data for every compound, preventing scalable pretraining on disjoint datasets, and they rely on hard contrastive objectives that impose the biologically implausible assumption that all distinct compounds are equally dissimilar negatives. We introduce Bio$\mathcal{X}$Mol, a multimodal representation learning framework that solves both issues. Bio$\mathcal{X}$Mol anchors disjoint phenotypic datasets (JUMP-CP and LINCS L1000) to molecular structures and replaces hard negatives with a soft contrastive objective, utilizing momentum-based teacher networks to weight negative repulsion by underlying biological similarity. When evaluated on a held-out set of clinical DILI activity cliffs, Bio$\mathcal{X}$Mol achieves 80.7\% pairwise ranking accuracy. Crucially, a hard contrastive baseline trained on the exact same data and architecture collapses to 51.0\% (random chance), while traditional ECFP fingerprints achieve 60.7\%. By simply modifying the contrastive objective to respect biological gradients, Bio$\mathcal{X}$Mol yields a 58\% relative performance increase over the hard contrastive baseline and a 33\% relative increase over structural fingerprints, demonstrating that soft phenotypic alignment is essential for accurately identifying toxicity cliffs.</p>