Multi-task learning with a natural metric for quantitative structure activity relationship learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Quantitative structure-activity relationship (QSAR) modeling aims to predict compound activity from molecular structure.
• Multi-task learning (MTL) can enhance QSAR by exploiting shared information across related drug targets and assays.
• Existing QSAR methods often treat each drug target independently, potentially missing valuable cross-target correlations.

Purpose of the Study:

• To investigate the effectiveness of MTL for QSAR prediction across multiple drug targets.
• To introduce and evaluate the utility of evolutionary distance between drug targets as a task relatedness metric in MTL.
• To compare instance-based and feature-based MTL approaches against a single-task learning baseline.

Main Methods:

• Utilized curated ChEMBL datasets containing compound activity data for 1091 drug targets.
• Implemented a single-task learning baseline using random forests for individual target prediction.
• Applied feature-based and instance-based MTL strategies, incorporating evolutionary target distance to measure task relatedness.

Main Results:

• Instance-based MTL significantly outperformed the baseline and feature-based MTL on 741 out of 1091 drug targets.
• Feature-based MTL showed superior performance on 179 targets, while the baseline was best for 171 targets.
• The integration of evolutionary target distance improved MTL QSAR performance.

Conclusions:

• MTL, particularly instance-based approaches, offers a powerful strategy for enhancing QSAR predictions.
• Leveraging target similarity through evolutionary distance is a key factor in successful MTL QSAR.
• MTL enables effective QSAR learning even with limited data for specific targets by transferring knowledge from related targets.