Activity landscape image analysis using convolutional neural networks.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Activity landscapes (ALs) visually represent compound similarity and activity data, aiding in the understanding of structure-activity relationships (SARs).
• Three-dimensional (3D) ALs offer complex topographical representations of SARs, analogous to geographical maps.
• Current graphical SAR analysis has not extensively utilized image analysis techniques for 3D ALs.

Purpose of the Study:

• To explore the potential of image analysis for characterizing 3D activity landscapes (ALs).
• To apply machine learning algorithms, including CNNs, SVMs, and RFs, to classify 3D AL images and their variants.
• To assess the effectiveness of different image features, particularly topological ones, in predicting SAR characteristics.

Main Methods:

• Construction of 3D ALs for various compound activity classes.
• Generation of 3D AL image variants with differing topology and information content.
• Application of Convolutional Neural Networks (CNNs) directly on 3D AL images.
• Extraction of 1D features from transformed 3D AL images for Support Vector Machine (SVM) and Random Forest (RF) models.
• Training SVM and RF models on features derived from edge filtering of 3D AL images.

Main Results:

• Machine learning models accurately distinguished between 3D AL image variants based on topology and information content.
• CNNs achieved the highest classification accuracy by directly learning from 3D AL images.
• Predictive performance for all models (CNN, SVM, RF) was optimal for image variants highlighting topological elevation.
• SVM models trained on edge-filtered images demonstrated high classification accuracy, underscoring the importance of topological features.

Conclusions:

• Image analysis of 3D ALs shows significant potential for graphical SAR exploration.
• Topological features within 3D AL images are critical for systematic SAR characterization and prediction.
• This proof-of-concept study validates machine learning-based image analysis as a viable approach for inferring SAR characteristics from 3D ALs.