Molormer

Molormer predicts drug–drug interactions by modeling two-dimensional (2D) molecular structures as spatially encoded molecular graphs processed with a lightweight attention-based Siamese architecture.

Key Features:

• Spatial Information Encoding: Encodes two-dimensional (2D) drug structures as molecular graphs with explicit spatial positional information to capture intramolecular relationships relevant to interactions.
• Lightweight Attention Mechanism: Employs a lightweight attention mechanism with self-attention distilling to retain multi-headed attention benefits while reducing computational and storage requirements.
• Siamese Network Architecture: Uses a Siamese network architecture to maximize use of limited training data and to minimize discrepancies between subnetworks processing drug features.
• Multi-label DDI Datasets: Operates on multi-label drug–drug interaction (DDI) datasets to support multi-label DDI prediction.

Scientific Applications:

• Drug–drug interaction prediction: Predicts DDIs to identify potential synergistic and adverse interactions among drugs.
• Combination therapy and personalized medicine: Supports analysis of multi-drug combination therapies and informs selection of drug combinations in personalized medicine contexts.
• Case-study predictions: Has been applied to predict interactions involving Aliskiren, Selexipag, and Vorapaxar in reported case studies.

Methodology:

Input processing of 2D drug structures; transformation into molecular graphs enriched with spatial information; application of a lightweight attention mechanism (with self-attention distilling) to those graphs; learning via a Siamese network architecture.