Multimodal IC50 prediction

PyTorch models fusing cell painting microscopy with SMILES chemical representations across compound libraries totaling 39K compounds.

AstraZeneca · Drug Discovery ML

Problem

Predicting compound activity (IC50) from chemical structure alone leaves information on the table. Cell painting — high-content imaging that captures cellular morphological response to compound treatment — provides a rich phenotypic signature that complements structure-based representations. Models that fuse the two should outperform either modality alone, but practical multimodal training brings its own challenges: heterogeneous data volumes, alignment between image and chemical representations, and joint embedding regularization.

Approach

PyTorch-based multimodal architecture combining:

• Cell painting microscopy branch — convolutional encoder over the multi-channel imaging stack, learning a phenotypic embedding of compound treatment effect.
• Chemical structure branch — SMILES tokenizer feeding a sequence model to produce a structural embedding.
• Joint head — fusion of the two embeddings into IC50 regression with uncertainty estimation, trained across compound libraries totaling 39K compounds.

Result

Deployed as part of internal compound activity prediction workflows. Multimodal fusion outperforms either single modality and provides interpretable embeddings that surface structure-phenotype correspondences useful for downstream target reasoning.

Stack

PyTorch, MONAI for image preprocessing, RDKit for SMILES handling, scikit-learn for evaluation tooling.