MoAble

MoAble predicts mechanisms of action (MoAs) of novel compounds by coembedding compound structures and transcriptomic signatures to infer transcriptomic responses and enable pathway enrichment–based MoA prediction.

Key Features:

• Deep learning coembedding model: Uses a deep learning-based coembedding approach to map compound structures and transcriptomic signatures into a shared embedding space.
• Structure-to-transcriptome prediction: Generates low-dimensional representations of compound-induced transcriptomic signatures directly from chemical structure data without requiring measured compound signatures.
• Connectivity-based MoA inference: Infers MoAs by analyzing connectivity between embedding vectors of compounds and embeddings derived from genetic perturbation datasets.
• Pathway enrichment analysis: Applies pathway enrichment analysis to predicted transcriptomic responses or embedding-derived signals to assign MoAs at the pathway level.
• Integration of structural and transcriptomic data: Integrates compound structure information with predicted transcriptomic responses for functional interpretation.
• Comparative performance: Demonstrates accuracy comparable to methods that utilize measured compound transcriptomic signatures.

Scientific Applications:

• Mechanism of action prediction: Predicts MoAs of novel compounds when experimental transcriptomic signatures are unavailable.
• Early-stage drug development: Supports MoA hypothesis generation and candidate prioritization in early-stage drug discovery from structure-only data.
• Connectivity mapping with genetic perturbations: Links compounds to genetic perturbation datasets via embedding connectivity for functional annotation.
• Pathway-level interpretation: Enables pathway-level interpretation of predicted compound effects through enrichment analysis.

Methodology:

Train a deep learning-based coembedding model to jointly embed compound structures and transcriptomic signatures into a unified low-dimensional space; generate predicted transcriptomic responses from compound structures; assess connectivity between compound embeddings and embeddings from genetic perturbation datasets; and apply pathway enrichment analysis on predicted signatures or embedding-derived signals.