Using information from historical high-throughput screens to predict active compounds.

Area of Science:

• Drug discovery and development
• Computational chemistry
• Machine learning in cheminformatics

Background:

• High-throughput screening (HTS) is crucial for hit identification in drug discovery.
• Modern drug discovery increasingly uses complex phenotypic screens, necessitating smarter compound selection strategies.
• Machine learning (ML) models are commonly trained on chemical fingerprints for hit expansion.

Purpose of the Study:

• To investigate an alternative ML approach using high-throughput screening (HTS) fingerprints for hit expansion.
• To evaluate the performance of ML models trained on HTS fingerprints against traditional chemical fingerprints.
• To explore the potential of combining HTS and chemical fingerprints for improved hit retrieval.

Main Methods:

• Constructed HTS fingerprints from in-house and public assay data (PubChem).
• Trained and compared three ML methods (Random Forest, Logistic Regression, Naïve Bayes) using HTS and Morgan2 chemical fingerprints.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUC) and enrichment factors.

Main Results:

• Random Forest (RF) models trained on HTS fingerprints achieved high AUC (>0.8) for 78% of internal assays.
• RF with HTS fingerprints outperformed Logistic Regression with chemical fingerprints in most assays.
• HTS fingerprints facilitated the retrieval of more diverse chemical structures (chemotypes).
• Heterogeneous classifier fusion combining both fingerprint types showed comparable or superior performance.

Conclusions:

• ML combined with HTS fingerprints is a promising strategy for hit expansion in drug discovery.
• This integrated approach effectively utilizes information from both assay outcomes and chemical structures.
• The method demonstrated strong performance in validation studies, including prospective compound selection.