Target-driven machine learning-enabled virtual screening (TAME-VS) enables prioritization of AKR1C3 inhibitors.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Machine learning (ML)-guided virtual screening shows promise for accelerating drug discovery.
• Current ML methods struggle with generalization to new chemical scaffolds and lack prospective validation.
• Aldo-keto reductase 1C3 (AKR1C3) is a key enzyme in cancer progression and chemoresistance.

Purpose of the Study:

• To develop and validate a machine learning-guided virtual screening workflow (TAME-VS) for target-driven drug discovery.
• To address challenges in generalization and prospective evaluation of ML models in drug screening.
• To identify novel inhibitors of AKR1C3 and assess their functional relevance in chemoresistant cancer cells.

Main Methods:

• Implemented TAME-VS, integrating automated target expansion, bioactivity data acquisition, and supervised learning.
• Utilized homolog-derived information to overcome data sparsity in classification model construction.
• Conducted retrospective validation on diverse inhibitor series and prospective validation on novel derivatives.

Main Results:

• TAME-VS demonstrated generalization to chemically distinct scaffolds, surpassing similarity-based prioritization.
• Model-guided prioritization identified active inhibitors more effectively than rational design alone under synthesis constraints.
• Identified S34-1035 and S34-1040 as potent, selective AKR1C3 inhibitors that restored doxorubicin sensitivity in resistant breast cancer cells.

Conclusions:

• TAME-VS offers a robust and generalizable approach for accelerating early-stage drug discovery.
• The identified AKR1C3 inhibitors hold potential for treating chemoresistant cancers.
• A publicly accessible TAME-VS server was developed to promote reproducible research in the field.