All-Assay-Max2 pQSAR: Activity Predictions as Accurate as Four-Concentration IC₅₀s for 8558 Novartis Assays

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Profile-quantitative structure-activity relationship (pQSAR) is a two-step machine learning approach.
• Previous pQSAR models were developed for 728 kinase assays.

Purpose of the Study:

• To develop a large-scale pQSAR model using over 11,000 diverse biochemical and cellular assays.
• To evaluate the performance and applicability of the scaled pQSAR model.

Main Methods:

• Training single-assay random forest regression models on biochemical and cellular pIC50 assays using Morgan 2 fingerprints.
• Building partial least squares (PLS) models using the profile of pIC50 predictions from random forest models.
• Reducing the profile by including only random forest models that correlate with the assay being modeled.

Main Results:

• The scaled pQSAR model achieved a median correlation of r_ext^2 = 0.53, significantly outperforming individual random forest models (median r_ext^2 = 0.05).
• 72% of the pQSAR models met the success threshold of r_ext^2 > 0.30, totaling 8558 successful models.
• Successful models covered all 51 annotated target subclasses and 4196 phenotypic assays, demonstrating broad applicability.

Conclusions:

• The large-scale pQSAR model demonstrates unprecedented scope, accuracy, and applicability domain for drug discovery.
• pQSAR is effective across diverse assay types, including phenotypic assays, and applicable to virtually any disease area.
• The continuously updated pQSAR models provide valuable predictions for virtual screening, selectivity design, and toxicity assessment.