FP2VEC

FP2VEC generates trainable embedding-based molecular features to improve quantitative structure–activity relationship (QSAR) classification and regression for chemical compound property prediction.

Key Features:

• Graph Convolution Operations: Applies graph convolution operations on molecular graphs and provides representations reported to outperform extended connectivity fingerprints (ECFP) in tested tasks.
• NLP-inspired Trainable Embeddings: Represents molecules as trainable embedding vectors analogous to words in natural language, enabling dynamic feature learning during model training.
• Convolutional Neural Network Architecture: Integrates a simple convolutional neural network (CNN) architecture adapted from sentence classification to process embedding-based molecular representations.
• Benchmark Performance: Achieves competitive results on several benchmark datasets across classification and regression QSAR tasks when combined with CNN models.
• Multitask Learning Effectiveness: Demonstrates effectiveness in multitask learning scenarios for simultaneous prediction of multiple molecular properties.

Scientific Applications:

• QSAR Modeling: Improves quantitative structure–activity relationship (QSAR) models for predicting chemical compound properties.
• Drug Discovery: Supports drug discovery workflows by enabling prediction of compound properties relevant to lead identification and optimization.
• Toxicology Assessment: Facilitates toxicology assessments through prediction of properties related to compound safety.
• Materials Science Research: Applies to materials science research for predicting molecular properties relevant to material design.

Methodology:

Represents molecules as sets of trainable embedding vectors, applies graph convolution operations on molecular graphs, and processes the embeddings with a simple convolutional neural network (CNN), including in multitask learning settings.