MolGraph-xLSTM as a graph-based dual-level xLSTM framework for enhanced molecular representation and interpretability

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Predicting molecular properties is crucial for accelerating drug discovery.
• Graph Neural Networks (GNNs) are widely used for molecular representation learning but struggle with long-range dependencies.
• Developing advanced computational methods is essential to enhance drug discovery efficiency.

Purpose of the Study:

• To introduce MolGraph-xLSTM, a novel graph-based xLSTM model designed to improve molecular property prediction.
• To enhance feature extraction and effectively model long-range interactions in molecules.
• To provide a more effective computational tool for drug discovery.

Main Methods:

• Processing molecular graphs at both atom-level and motif-level scales.
• Utilizing a GNN-based xLSTM framework with jumping knowledge for local feature extraction and multilayer information aggregation.
• Refining embeddings using a multi-head mixture of experts (MHMoE) for enhanced expressiveness.
• Validating the model on 21 datasets from MoleculeNet and Therapeutics Data Commons (TDC) benchmarks.

Main Results:

• MolGraph-xLSTM achieved significant improvements on MoleculeNet: 3.18% average AUROC increase for classification and 3.83% RMSE reduction for regression.
• On the TDC benchmark, the model showed a 2.56% AUROC improvement and a 3.71% RMSE reduction on average.
• These results demonstrate superior performance compared to baseline methods in both classification and regression tasks.

Conclusions:

• MolGraph-xLSTM effectively captures long-range molecular interactions, outperforming existing GNNs.
• The dual-scale graph processing and MHMoE refinement contribute to enhanced molecular representation learning.
• The model shows strong generalizability and effectiveness for computational drug discovery tasks.