Empowering chemical structures with biological insights for scalable phenotypic virtual screening

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Scalable identification of bioactive compounds is crucial for drug discovery.
• Current methods face a trade-off between scalability (structural screening) and biological insight (phenotypic profiling).
• Extracting robust biological signals from noisy data for inference without biological data is a key challenge.

Purpose of the Study:

• To present DECODE (DEcomposing Cellular Observations of Drug Effects), a framework enabling structure-based in silico biological profiling.
• To empower chemical representations with intrinsic biological semantics.
• To bridge the gap between structural screening and phenotypic profiling.

Main Methods:

• DECODE leverages limited paired transcriptomic and morphological data as supervisory signals.
• It extracts measurement-invariant biological fingerprints from chemical structures.
• Modality-specific variation is explicitly filtered.

Main Results:

• DECODE improves functional retrieval and early active-compound prioritization across various virtual screening benchmarks.
• The framework demonstrates enhanced performance in held-out retrieval and scaffold-split tasks.
• UMAP-clustering benchmarks show DECODE's effectiveness in compound prioritization.

Conclusions:

• DECODE provides a novel approach to in silico biological profiling by integrating chemical structure with biological semantics.
• The framework facilitates more efficient and effective drug discovery pipelines.
• DECODE's ability to extract biological signals without requiring inference-time biological data represents a significant advancement.