Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships

Area of Science:

• Computational biology and cheminformatics
• Drug discovery and development
• Cellular imaging and phenotypic profiling

Background:

• Traditional compound selection relies on chemical similarity, often missing diverse molecules with similar biological effects.
• This structure-centric approach overlooks compounds with therapeutic potential due to functional convergence.
• Limitations in exploring chemical space hinder the identification of novel bioactive agents.

Purpose of the Study:

• To develop a training-free, transfer learning method for large-scale compound preselection using deep phenotypic profiling.
• To enable robust comparison of cellular phenotypic signatures across vast datasets like JUMP-CP.
• To overcome the limitations of structure-based methods in drug discovery.

Main Methods:

• Leveraged deep phenotypic profiling of human cells using the JUMP-CP dataset (112,480 compounds).
• Employed a transfer learning-based approach for training-free, large-scale compound preselection.
• Validated the method across 65 high-throughput assays, including in vitro and in cellulo systems.

Main Results:

• The method efficiently enriches biologically active compounds, surpassing structure-centric approaches.
• Discovered thousands of compound activity cliffs, revealing atom-level bioactivity determinants.
• Demonstrated that structurally diverse compounds can induce convergent or opposite phenotypes across 30 pathways.

Conclusions:

• Phenotypic profiling offers a transformative, scalable framework for navigating chemical space biologically.
• The approach enables discovery of bioactive compounds, novel mechanisms, and target-pathway relationships.
• Phenoseeker tool facilitates access to compounds with similar phenotypic profiles for broader research community.