Virtual Screening with Generative Topographic Maps: How Many Maps Are Required?

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Universal generative topographic maps (GTMs) represent chemical space for polypharmacology.
• Individual GTMs, while equivalent on average, show variable performance for specific property landscapes.
• Distinct GTMs capture complementary and synergistic views of biologically relevant chemical space.

Purpose of the Study:

• To evaluate the synergy of multiple universal GTMs for predictive classification.
• To assess the performance of GTMs in virtual screening (VS) using external datasets.
• To determine if cross-validated (CV) accuracy predicts VS success.

Main Methods:

• Employed eight universal GTMs to support predictive classification landscapes.
• Utilized over 600 active/inactive ligand series from the ChEMBL database (v.23).
• Validated performance using external active/decoy sets from the Directory of Useful Decoys (DUD).
• Estimated activity probability for molecules projected onto GTM landscapes, simulating VS.
• Assessed prediction trustworthiness based on applicability domain considerations.

Main Results:

• Cross-validated balanced accuracy on ChEMBL data did not predict VS success.
• Individual GTM predictions for external datasets were often untrustworthy.
• Simultaneous application of all universal GTMs significantly improved VS prediction results.
• Consensus VS using multiple maps consistently outperformed or matched the best individual map.

Conclusions:

• Prioritizing a single universal GTM based on CV results is not advisable for VS.
• Combining multiple GTMs in a consensus approach enhances predictive performance and reliability.
• Synergistic use of diverse GTMs is crucial for robust virtual screening and understanding chemical space.