Graph Multi-Similarity Learning for Molecular Property Prediction

TL;DR

The paper addresses molecular property prediction and the limitations of contrastive learning that relies on binary positive/negative pairs. It proposes GraphMSL, which learns a generalized multi-similarity metric from self-similarity across multiple modalities (SMILES, images, NMR, fingerprints) and fuses them multimodally, without requiring predefined pairs. The authors establish a convergent similarity property and demonstrate graph-level, node-level, and bi-level variants with strong performance on MoleculeNet benchmarks, together with post-hoc explainability analyses. Overall, GraphMSL provides a flexible, interpretable framework for multimodal molecular representation learning with demonstrated predictive gains and actionable design insights for drug discovery.

Abstract

Enhancing accurate molecular property prediction relies on effective and proficient representation learning. It is crucial to incorporate diverse molecular relationships characterized by multi-similarity (self-similarity and relative similarities) between molecules. However, current molecular representation learning methods fall short in exploring multi-similarity and often underestimate the complexity of relationships between molecules. Additionally, previous multi-similarity approaches require the specification of positive and negative pairs to attribute distinct predefined weights to different relative similarities, which can introduce potential bias. In this work, we introduce Graph Multi-Similarity Learning for Molecular Property Prediction (GraphMSL) framework, along with a novel approach to formulate a generalized multi-similarity metric without the need to define positive and negative pairs. In each of the chemical modality spaces (e.g.,molecular depiction image, fingerprint, NMR, and SMILES) under consideration, we first define a self-similarity metric (i.e., similarity between an anchor molecule and another molecule), and then transform it into a generalized multi-similarity metric for the anchor through a pair weighting function. GraphMSL validates the efficacy of the multi-similarity metric across MoleculeNet datasets. Furthermore, these metrics of all modalities are integrated into a multimodal multi-similarity metric, which showcases the potential to improve the performance. Moreover, the focus of the model can be redirected or customized by altering the fusion function. Last but not least, GraphMSL proves effective in drug discovery evaluations through post-hoc analyses of the learnt representations.

Figures (5)

• Figure 1: Graph Similarity Learning for Molecular Property Prediction (GraphMSL).$\mathcal{\mathrm} t_{i,j}^{g}$, $\mathcal{\mathrm} t_{l,m}^{n}$ represent graph-level and node-level target similarity, respectively. $\mathcal{\mathrm} d_{i,j}^{g}$, $\mathcal{\mathrm} d_{l,m}^{n}$ represent the similarity for graph-level and node-level embeddings, respectively. Here, $i$ and $j$ are the indices of molecular graphs in the graph pool, and $l$ and $m$ are the indices of atomic nodes in the node pool. Unlike the general contrastive learning framework shown in Appendix Figure \ref{['fig:traditional-cl']}, GraphMSL doesn't need to define positive or negative pairs and is capable of learning continuous ordering from target similarity.
• Figure 2: T-SNE visualization depicting the ESOL molecule embeddings alongside molecules within the highlighted region. Each point in the heatmap corresponds to the embeddings of respective molecules in ESOL, with color indicating solubility levels. Red denotes higher solubility, while blue indicates lower solubility.
• Figure 3: T-SNE Visualization of BACE embedding and clustering based on minimum positive subgraph (MPS). MPS represents minimum positive subgraph of a positive molecule; MCS represents maximum common subgraph of several positive molecules, sharing the same MPS. Pink nodes represent MPS, blue nodes depict molecules, and edge colors indicate binding potential differences. Red edges denote successful designs (original higher than MPS), while blue indicates less efficient designs (original lower than MPS).
• Figure 1.1: Illustration of Different Types of Similarities.
• Figure 1.2: General framework of Intra-Modality Graph Contrastive Learning. It relies on definition of positive and negative pairs.

Theorems & Definitions (3)

• Theorem 3.1: Theorem of Convergent Similarity learning
• Theorem 2.1: Theorem of Convergent Similarity learning
• proof