Deep learning for assay nuisance compound detection using a gated co-attention graph embedding model (CAGE-Fusion)

Area of Science:

• Computational chemistry
• Drug discovery informatics
• Machine learning in cheminformatics

Background:

• Nuisance compounds in drug discovery cause false positives by interfering with assays through mechanisms like aggregation or reactivity.
• Existing methods struggle to filter these compounds, leading to wasted resources.
• Current deep learning models often process molecular graphs or SMILES sequences independently.

Purpose of the Study:

• To introduce CAGE-Fusion, a multimodal deep learning framework for improved nuisance compound detection.
• To leverage co-attention mechanisms for enhanced molecular representation learning.
• To provide a chemically coherent embedding by integrating graph and sequence data.

Main Methods:

• Developed CAGE-Fusion, a multimodal framework using a gated co-attention mechanism.
• Enabled bidirectional information exchange between graph-based and sequence-based encoders.
• Iteratively refined atom and token embeddings for cross-modal dependency capture.

Main Results:

• CAGE-Fusion achieved competitive performance on MoleculeNet benchmarks.
• Demonstrated high accuracy in nuisance compound classification with macro-averaged PR-AUC of 0.73 and ROC-AUC of 0.94.
• The model effectively visualizes attention-weighted regions for classification decisions.

Conclusions:

• CAGE-Fusion offers a powerful approach for identifying assay nuisance compounds by integrating diverse molecular representations.
• The framework enhances predictive performance and provides interpretable insights.
• This method aids in reducing false positives and optimizing resource allocation in drug discovery.