Medication Recommendation via Dual Molecular Modalities and Multi-Step Enhancement

TL;DR

A bimodal molecular recommendation framework named BiMoRec is proposed, which introduces 3D molecular structures to obtain atomic 3D coordinates and edge indices, overcoming the inherent lack of high-dimensional molecular information in 2D molecular structures.

Abstract

Existing works based on molecular knowledge neglect the 3D geometric structure of molecules and fail to learn the high-dimensional information of medications, leading to structural confusion. Additionally, it does not extract key substructures from a single patient visit, resulting in the failure to identify medication molecules suitable for the current patient visit. To address the above limitations, we propose a bimodal molecular recommendation framework named BiMoRec, which introduces 3D molecular structures to obtain atomic 3D coordinates and edge indices, overcoming the inherent lack of high-dimensional molecular information in 2D molecular structures. To retain the fast training and prediction efficiency of the recommendation system, we use bimodal graph contrastive pretraining to maximize the mutual information between the two molecular modalities, achieving the fusion of 2D and 3D molecular graphs. Additionally, we designed a molecular multi-step enhancement mechanism to re-calibrate the molecular weights. Specifically, we employ a pre-training method that captures both 2D and 3D molecular structure representations, along with substructure representations, and leverages contrastive learning to extract mutual information. We then use the pre-trained encoder to generate molecular representations, enhancing them through a three-step process: intra-visit, molecular per-visit, and latest-visit. Finally, we apply temporal information aggregation to generate the final medication combinations. Our implementation on the MIMIC-III and MIMIC-IV datasets demonstrates that our method achieves state-of-the-art performance.

Figures (7)

• Figure 1: Spatial structure representation of Atropine and Ipratropium. (a) Atropine is a medication for treating depression, while Ipratropium is used for treating obstructive bronchitis; (b) 2D molecular structure representation; (c) 3D molecular structure representation.
• Figure 2: Comparison of molecular embedding similarity and molecular fingerprint similarity, with lighter colors indicating greater differences in similarity. The left side uses only the 2D modality, while the right side uses 2D and 3D modalities.
• Figure 3: The pre-training phase illustrated in the bottom left side of the figure demonstrates the bimodal fusion process of molecular structure and decomposed substructures, and its results are applied to the downstream tasks shown in the top left side. The molecular multi-step enhancement phase in the upper corner explains how molecular knowledge interacts with patient representations at the visit level. It sequentially performs molecular enhancement at three hierarchical levels: intra-visit, molecular per-visit, and latest-visit. The temporal information aggregation phase on the lower right side of the figure models the sequence relationships of the interaction results and maps them to recommendation probabilities, obtaining the final recommendation results.
• Figure 4: Comparison with recent outstanding works across all metrics in MIMIC-III.
• Figure 5: Experiments on parameter sensitivity on MIMIC-III dataset.