Multimodal Contrast-Enhanced Molecular Representation Learning and Property Prediction

Abstract

Molecular representation learning (MRL) has garnered significant attention due to its pivotal role in downstream applications such as molecular property prediction and drug discovery. In most MRL approaches, molecules are encoded into 2D topological graphs via graph neural network (GNN), which suffers from over-smoothing issues and limited receptive fields. Furthermore, most GNN models fail to utilize the 3D spatial structural information that determines molecular physicochemical properties and biological activity. To this end, here we propose multimodal contrast-enhanced molecular representation learning (MCMRL). This approach utilizes both the 2D topological information and 3D structural information of molecules for contrastive learning to enhance molecular graph representations. Further, it integrates additional molecular fingerprint information and feature fusion techniques to incorporate multimodal knowledge, yielding more reliable and generalizable molecular representations. MCMRL is pre-trained on ~10 million unlabeled molecules from PubChem, followed by various downstream benchmark tasks. Experimental results demonstrate that MCMRL achieves superior performance in 9 out of 13 benchmark tests for molecular property prediction, validating its effectiveness in molecular representation learning. Furthermore, potential molecular drugs binding to biological target protein DRD2 screened by MCMRL representation show promising affinity score, which also demonstrates the efficacy of the proposed method.

1. Introduction

Molecular property prediction is widely accepted as a fundamental step in computational drug and material discovery, with numerous methods relying on accurate molecular property for molecular evaluation, screening, and generation [1,2]. For this purpose, reliable and generalizable molecular representation learning is an essential task and has gained increasing attention in this area. Many conventional molecular representations, such as SMILES [3], ECFP [4] and RDKFP [5], have been developed and widely adopted. However, given the enormous magnitude of the chemical space of potential pharmacologically active molecules (~10⁶⁰) [6], universal molecular representation learning methods are urgently in need and still face fundamental challenges [7].

In recent years, graph neural networks (GNNs) have emerged as the mainstream approach for molecular representation learning (MRL) based on deep neural networks (DNNs) [8]. By effectively mapping molecular structures onto topological graphs (where atoms serve as nodes and chemical bonds as edges), GNNs capture topological information to predict molecular properties [9,10]. However, their inherent limitations have become a persistent bottleneck for performance improvement: GNNs suffer from severe over-smoothing and restricted receptive fields during graph convolution and message passing, resulting in inadequate representation of complex molecular features [11,12,13]. Attention FP propagates node information by extending graph attention mechanisms to capture local atomic environments, even identifying latent intramolecular critical bonds. Yet, it remains confined to modeling topological structures while neglecting non-local spatial correlations [14]. D-MPNN employs a specialized message-passing architecture to aggregate molecular graph structural information yet similarly encodes only 2D topological features without considering 3D geometric information [15]. Variants such as MGCN [16] advance explicit modeling of intramolecular quantum interactions yet remain tailored for quantum property prediction. They fail to generalize to general physicochemical prediction tasks and lack integration of comprehensive 3D structural features—crucial for determining molecular physicochemical properties and biological activity [17,18,19].

Self-supervised learning (SSL) alleviates the scarcity of annotated data in MRL by mining large-scale unlabeled molecular datasets, having found widespread application in natural language processing and computer vision [20,21,22,23]. However, its implementation in MRL remains nascent and lacks chemical depth. Existing SSL approaches suffer from narrow feature modeling and oversimplified pre-training designs, failing to match the semantic complexity inherent in molecular structure–property relationships. Early SSL-based MRL research, such as PretrainGNN [24], utilized solely molecular topological information while completely disregarding 3D geometry. Pre-training was conducted exclusively through masking and predicting nodes, edges, or contextual features within 2D topological graphs. GraphMVP [25] attempted to address this by emphasizing alignment and consistency between 2D and 3D views in SSL yet failed to incorporate molecular functional group information. Another representative SSL approach, FG-BERT [26], employed a functional group masking strategy for pre-training to optimize 2D topological substructure modeling, but similarly excluded 3D spatial information, thereby failing to capture molecular structural similarities and differences determined by geometric structure. In summary, existing SSL-based MRL approaches either entirely disregard 3D geometric information or achieve only superficial alignment between 2D and 3D views without deep integration. Their pre-training tasks underestimate molecular properties’ sensitivity to subtle structural alterations—minor structural changes may induce drastic property shifts. Existing methods struggle to extract chemically meaningful semantic representations, let alone achieve accurate and generalizable MRL [27,28,29,30,31,32,33].

The main limitations of existing self-supervised molecular representation learning methods can be summarized as the following two aspects. First, the integration of molecular 2D topological features and 3D spatial geometric features remains limited to superficial alignment, failing to achieve deep cross-modal enhancement and feature interaction [34,35,36,37]. This impedes the extraction of the intricate correlations between molecular structure and properties. Second, they neglect prior chemical knowledge embedded in conventional molecular fingerprints, such as functional groups and substructures. Feature learning from a single modality or partial modalities struggles to cover the complex physicochemical properties of molecules. These two shortcomings result in insufficient generalization capabilities for predicting molecular properties and an inability to effectively learn chemically meaningful molecular representations.

To address these challenges, here we propose multimodal contrast-enhanced molecular representation learning (MCMRL). By conducting contrastive learning between 2D topological information and 3D structural information, it optimizes molecular graph representations. In addition, the incorporation of additional molecular fingerprint information, coupled with feature fusion techniques, fully integrates multimodal molecular knowledge to construct more robust and generalizable molecular representations. To validate the effectiveness of the proposed MCMRL method, we evaluated it against multiple state-of-the-art (SOTA) baseline models across 13 molecular property prediction benchmark datasets, achieving superior results on nine tasks. Furthermore, we screened potential molecular drugs binding to biological target protein DRD2 using MCMRL representation. The proposed candidate molecules show promising affinity scores.

Our contributions can be summarized as follows:

1. We propose a novel multimodal contrast-enhanced molecular representation learning model that innovatively integrates multimodal information including 2D topology, 3D spatial structure, and functional group fingerprints, overcoming the limitations of existing methods that rely on single-modal or partial-modal feature learning.

2. We construct a cross-attention-based feature fusion module that achieves deep integration of local topological embeddings, global geometric embeddings, and functional group fingerprint information, yielding reliable and generalizable molecular feature representations.

3. Benchmark dataset experiments demonstrate that our approach achieves significant performance improvements in molecular property prediction tasks. Case study on potential drugs of DRD2 further proves the efficacy of the proposed approach.

2. Materials and Methods

2.1. 2D Encoder Module

In our work, a 2D molecule graph is defined as $G = (V,E)$, where $V$ and $E$ are nodes (atoms) and edges (chemical bonds), respectively. Since GNN can make use of the information transfer mechanism so that the model can fully learn the local topological information, we use GNN to exploit the neighborhood aggregation operation in the 2D Encoder Module, which updates the node representation iteratively. The aggregation update rule for a node feature on the $\mathrm{k}$th layer of a GNN is given in Equations (1) and (2):

$${\mathrm{a}}_{\mathrm{v}}^{\left(\mathrm{k}\right)}={\mathrm{A}\mathrm{g}\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}}^{\left(\mathrm{k}\right)}\left(\left\{{\mathrm{h}}_{\mathrm{u}}^{\mathrm{k}-1}:\mathrm{u}\in \mathcal{N}\left(\mathrm{v}\right)\right\}\right),$$

(1)

$${\mathrm{h}}_{\mathrm{v}}^{\left(\mathrm{k}\right)}={\mathrm{C}\mathrm{o}\mathrm{m}\mathrm{b}\mathrm{i}\mathrm{n}\mathrm{e}}^{\left(\mathrm{k}\right)}\left(\left\{{\mathrm{h}}_{\mathrm{u}}^{\mathrm{k}-1},{\mathrm{a}}_{\mathrm{v}}^{\left(\mathrm{K}\right)}\right\}\right),$$

(2)

where ${\mathrm{h}}_{\mathrm{v}}^{\left(\mathrm{k}\right)}$ is the feature of node $\mathrm{v}$ at the $\mathrm{k}$th layer and ${\mathrm{h}}_{\mathrm{v}}^{\left(0\right)}$ is initialized by node feature ${\mathrm{h}}_{\mathrm{v}}$. $\mathcal{N}\left(\mathrm{v}\right)$ denotes the set of all the neighbors of node v. To further extract a graph-level feature ${\mathrm{h}}_{\mathrm{G}}$, readout operation integrates all the node features among the graph $\mathrm{G}$ as given in Equation (3):

$${\mathrm{h}}_{\mathrm{G}}={\mathrm{R}\mathrm{E}\mathrm{A}\mathrm{D}\mathrm{O}\mathrm{U}\mathrm{T}}^{\left(\mathrm{k}\right)}\left(\left\{{\mathrm{h}}_{\mathrm{u}}^{\mathrm{k}}:\mathrm{v}\in \mathrm{G}\right\}\right)$$

(3)

In our work, we build GNN encoders based on GIN. While GIN utilizes an MLP and weighted summation of node features in the aggregation, both are simple yet generic graph convolutional operations. Additionally, we implement widely used mean pooling as the readout.

2.2. 3D Encoder Module

In recent years, learning 3D geometric representations has achieved great progress in molecular modeling. 3D molecular graphs incorporate the spatial positions of atoms, which need not be static since atoms continuously move along potential energy surfaces in real-world scenarios. The 3D structure at local minima on this surface is termed a molecular conformation. To fully leverage molecular geometric information and better characterize global features, this paper defines a 3D molecular graph as $\mathrm{G}=\left(\mathrm{X},\mathrm{R}\right)$, where $\mathrm{X}$ represents nodes (atoms). Compared to 2D molecular graphs, this approach introduces an additional coordinate dimension $\mathrm{R}$ for nodes and employs a 3D graph neural network to encode the representation embedding of molecular conformations. Specifically, it can be expressed as

$${\mathrm{h}}_{3\mathrm{D}}=\mathrm{G}\mathrm{N}\mathrm{N}\left({\mathrm{T}}_{3\mathrm{D}}\right({\mathrm{g}}_{3\mathrm{D}}\left)\right)=\mathrm{G}\mathrm{N}\mathrm{N}\left({\mathrm{T}}_{3\mathrm{D}}\right(\mathrm{X},\mathrm{R}\left)\right),$$

(4)

where $\mathrm{R}$ is the 3D-coordinate matrix and ${\mathrm{T}}_{3\mathrm{D}}$ is the 3D transformation. Note that further information such as plane and torsion angles can be solved from the positions. SchNet is composed of the following key steps:

$${\mathrm{z}}_{\mathrm{i}}^{\left(0\right)}=\mathrm{e}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{d}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}({\mathrm{x}}_{\mathrm{i}}),$$

(5)

$${\mathrm{z}}_{\mathrm{i}}^{(\mathrm{t}+1)}=\mathrm{M}\mathrm{L}\mathrm{P}(\sum _{\mathrm{j}=1}^{\mathrm{n}}\mathrm{f}({\mathrm{x}}_{\mathrm{j}}^{(\mathrm{t}-1)},{\mathrm{r}}_{\mathrm{i}},{\mathrm{r}}_{\mathrm{j}})),$$

(6)

$${\mathrm{h}}_{\mathrm{i}}=\mathrm{M}\mathrm{L}\mathrm{P}\left({\mathrm{z}}_{\mathrm{i}}^{\left(\mathrm{K}\right)}\right),$$

(7)

where $\mathrm{K}$ is the number of hidden layers, and

$$\mathrm{f}({\mathrm{x}}_{\mathrm{j}},{\mathrm{r}}_{\mathrm{i}},{\mathrm{r}}_{\mathrm{j}})={\mathrm{x}}_{\mathrm{j}}·{\mathrm{e}}_{\mathrm{k}}({\mathrm{r}}_{\mathrm{i}}-{\mathrm{r}}_{\mathrm{j}})={\mathrm{r}}_{\mathrm{j}}·\mathrm{e}\mathrm{x}\mathrm{p}(-\mathsf{\gamma }{‖‖{\mathrm{r}}_{\mathrm{i}}-{\mathrm{r}}_{\mathrm{j}}‖2-\mathsf{\mu }‖}_2^2),$$

(8)

is the continuous-filter convolution layer, enabling the modeling of continuous positions of atoms. The 3D structure of each molecule is generated with RDKit toolbox [5]. The conformation is initialized with Experimental Torsion Knowledge Distance Geometry (ETKDG) algorithm and optimized with Merck Molecular Force Field (MMFF94) before model pre-training and downstream tasks. Considering the processing speed and model efficiency, the molecular 3D conformation is currently simplified by a representative 3D structure.

2.3. Molecular Fingerprints Encoder Module

We use here three complementary fingerprints MACCS, PubChem and Pharmacophore ErG fingerprints, as detailed below.

MACCS Fingerprint: A fingerprint based on a substructure key using SMARTS mode. MACCS contains the most atomic properties, bond properties, and atomic neighborhoods in different topological separations, which is implicative for drug discovery. We chose the short variant of the 166 bits for this study.

PubChem Fingerprint: An 881-bit fingerprint based on a substructure key with broad chemical structure coverage.

Pharmacophore ErG fingerprint: This uses the extended reduce graph (ErG) method and pharmacophore-type node descriptions are applied to encode molecular properties.

Among these, MACCS can effectively capture local functional groups (such as carboxyl groups and pyridine groups). PubChem can supplement large-scale features such as macrocycles and stereocenters. ErG can encode 3D pharmacophore distributions, addressing the limitations of the previous two 2D fingerprints. Therefore, the combination of the three achieves a three-level complementary approach of ‘local-global-interaction,’ significantly enhancing the accuracy, robustness, and generalization capabilities of drug discovery task.

Combining these three into a hybrid fingerprint to Equation (9),

$$\mathrm{F}\mathrm{P}=\mathrm{C}\mathrm{O}\mathrm{N}\mathrm{C}\mathrm{A}\mathrm{T}\left({\mathrm{F}\mathrm{P}}_{\mathrm{M}\mathrm{A}\mathrm{C}\mathrm{C}\mathrm{S}},{\mathrm{F}\mathrm{P}}_{\mathrm{P}\mathrm{u}\mathrm{b}\mathrm{C}\mathrm{h}\mathrm{e}\mathrm{m}},{\mathrm{F}\mathrm{P}}_{\mathrm{P}\mathrm{h}\mathrm{a}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{c}\mathrm{o}\mathrm{p}\mathrm{h}\mathrm{o}\mathrm{r}\mathrm{e}\mathrm{E}\mathrm{r}\mathrm{G}}\right),$$

(9)

The fingerprints vector was input into the artificial neural network (ANN) to obtain the following representation Equation (10):

$$\mathrm{V}=\mathrm{W}·\mathrm{F}\mathrm{P}+\mathrm{b},$$

(10)

2.4. Fusion Module

The computational methods are illustrated in Equations (11) and (12). To efficiently integrate 2D local topological features, 3D global geometric features, and molecular fingerprint knowledge, MCMRL employs a feature fusion module based on a cross-attention mechanism. The 2D molecular graph and 3D molecular graph are encoded respectively by a 2D graph neural network module and a 3D graph neural network module. The 3D graph neural network aims to capture global molecular information, whilst this module more finely represents the geometric information within the 3D molecular graph. However, employing solely a 3D graph neural network architecture may result in the loss of local molecular information (such as atoms, bonds, and connectivity relationships), whereas the 2D graph neural network can capture the local node information of the molecule. Furthermore, we introduce three complementary types of molecular fingerprint information as supplementary knowledge to enhance the model’s ability to accurately characterize molecular features. These two hierarchical information types—molecular graphs and molecular fingerprints—are fed into a fusion module. Through an interactive attention mechanism, they are integrated to achieve the fusion of graph structure and functional group information. Ultimately, through a series of operations, a robust representation integrating local topological features, global collective features, and supplementary molecular fingerprint information is obtained. This yields the final molecular representation: the 2D molecular graph embedding ${\mathrm{h}}_{\mathrm{T}}$ originates from the 2D graph neural network, while the 3D molecular graph embedding ${\mathrm{h}}_{\mathrm{G}}$ derives from the 3D graph neural network module. These two are concatenated to form ${\mathrm{h}}_{\mathrm{C}}$, which is input alongside the molecular fingerprint ${\mathrm{h}}_{\mathrm{F}\mathrm{P}}$ into the feature fusion module. This module computes the interactive attention between graph structure and functional group information, using ${\mathrm{h}}_{\mathrm{C}}$ as the query vector and ${\mathrm{h}}_{\mathrm{F}\mathrm{P}}$ as the key–value pair. The specific computational workflow of the fusion module is illustrated in the figure.

$$\mathrm{Q},\mathrm{K},\mathrm{V}={\mathrm{L}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}_{\mathrm{Q}}\left({\mathrm{h}}_{\mathrm{C}}\right),{\mathrm{L}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}_{\mathrm{K}}\left({\mathrm{h}}_{\mathrm{F}\mathrm{P}}\right),{\mathrm{L}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}}_{\mathrm{V}}\left({\mathrm{h}}_{\mathrm{F}\mathrm{P}}\right),$$

(11)

$$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{Q},\mathrm{K},\mathrm{V}\right)=\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\displaystyle \frac{\mathrm{Q}{\mathrm{K}}^{\mathrm{T}}}{\sqrt{\mathrm{d}}}}\right)\mathrm{V},$$

(12)

By applying layer normalization (LN) following the multi-head attention mechanism and feedforward network, the molecular representation H is ultimately obtained. This fusion network aggregates and embeds local atomic information from the 2D molecular graph, global molecular information from the 3D molecular graph, and molecular fingerprint information. Through an interactive attention mechanism, it assigns higher weights to important atoms throughout the entire molecule.

2.5. Contrastive Learning Module

To achieve effective alignment between 2D topological features and 3D geometric features, a projection head is required to map both modalities onto the same vector space. This projection head takes as input the 512-dimensional topological features output by the 2D GIN model and the 512-dimensional geometric features output by the 3D SchNet model. This projection head achieves dimensionality reduction and spatial alignment through two layers of linear transformations: The first linear layer contains 512 neurons and employs the ReLU activation function to perform nonlinear feature transformation. The second linear layer contains 256 neurons and omits an activation function to preserve feature continuity, ultimately outputting a 256-dimensional modal feature vector. The projected features are fed into the InfoNCE loss calculation module, which maximizes similarity between features from the same molecule while minimizing similarity between features from different molecules, thereby achieving contrastive learning optimization for multimodal representation. Specifically, after obtaining the 2D molecular graph and 3D molecular graph representations for each molecule, we maximize the mutual information between the 2D molecular graph and 3D molecular graph representations from the same molecule using InfoNCE as the contrastive loss, while distinguishing between the two representations from different molecules. Specifically, we first use a projection head to graph the two representations into the same vector space, then select positive and negative pairs for contrastive learning. During training, the 2D molecular graph and 3D molecular graph representations from molecules are used as input, where the 2D molecular graph and 3D molecular graph representations from the same molecule are used as positive sample pairs, and the two representations from different molecules are used as negative sample pairs, thereby generating positive pairs and negative sample pairs. More specifically, the $\mathrm{i}$th 2D molecular graph representation and the $\mathrm{i}$th 3D molecular graph representation form a unique positive sample pair, while the remaining 2D molecular graph representations and the $\mathrm{i}$th 3D molecular graph representation form negative sample pairs. Therefore, the InfoNCE loss for the $\mathrm{i}$th 2D molecular graph is

$${\mathrm{L}}_{\mathrm{i}}^{\mathrm{t}}=-\log \left({\displaystyle \frac{\exp \left(\mathrm{s}\mathrm{i}\mathrm{m}\left({\mathrm{z}}_{\mathrm{i}}^{\mathrm{t}},{\mathrm{z}}_{\mathrm{i}}^{\mathrm{g}}\right)/\mathsf{\tau }\right)}{{\sum }_{\mathrm{j}=1}^{\mathrm{N}}\exp \left(\mathrm{s}\mathrm{i}\mathrm{m}\left({\mathrm{z}}_{\mathrm{i}}^{\mathrm{t}},{\mathrm{z}}_{\mathrm{j}}^{\mathrm{g}}\right)/\mathsf{\tau }\right)}}\right),$$

(13)

where $\mathrm{t}$ and $\mathrm{g}$ represent 2D molecular graphs and 3D molecular graphs, respectively, $\mathrm{s}\mathrm{i}\mathrm{m}(·)$ is a pairwise similarity function using cosine similarity, and $\mathsf{\tau }$ is the temperature parameter. Similarly, the loss function for the $\mathrm{i}$th 3D molecular graph is

$${\mathrm{L}}_{\mathrm{i}}^{\mathrm{g}}=-\mathrm{l}\mathrm{o}\mathrm{g}\left({\displaystyle \frac{\exp \left(\mathrm{s}\mathrm{i}\mathrm{m}\left({\mathrm{z}}_{\mathrm{i}}^{\mathrm{g}},{\mathrm{z}}_{\mathrm{i}}^{\mathrm{t}}\right)/\mathsf{\tau }\right)}{{\sum }_{\mathrm{j}=1}^{\mathrm{N}}\exp \left(\mathrm{s}\mathrm{i}\mathrm{m}\left({\mathrm{z}}_{\mathrm{i}}^{\mathrm{g}},{\mathrm{z}}_{\mathrm{j}}^{\mathrm{t}}\right)/\mathsf{\tau }\right)}}\right),$$

(14)

In summary, the final InfoNCE loss function is as follows:

$${\mathrm{L}}_{\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{o}}={\displaystyle \frac{1}{2\mathrm{N}}}\sum _{\mathrm{k}\in \mathrm{t},\mathrm{g}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{L}}_{\mathrm{i}}^{\mathrm{k}}.$$

(15)

3. Results

3.1. Overview of MCMRL

Multimodal contrast-enhanced molecular representation learning (MCMRL) is a self-supervised contrastive learning framework for MRL that optimizes molecular graph representations through contrastive learning of 2D topological information with 3D structural information. Furthermore, it incorporates additional molecular fingerprint information and feature fusion techniques to construct more robust and generalizable molecular representations. Within this framework, a 2D encoder extracts local topological information via a Graph Isomorphism Network (GIN); the 3D encoder utilizes a 3D graph neural network [38] to extract global spatial structural features; the fingerprint encoder employs an artificial neural network to extract fingerprint information; and feature fusion techniques adopt a cross-attention mechanism. The basic workflow comprises two stages: pre-training and fine-tuning, as shown in Figure 1.

The MCMRL model employs contrastive learning [39,40], which optimizes molecular graph representations by minimizing distances between duplicate representations of the same molecule while maximizing distances between distinct molecular representations [41]. It further integrates three complementary molecular fingerprint representations—namely, the MACCS fingerprint [42], the Pharmacophore ErG fingerprint [43], and the PubChem fingerprint [44]. These are fused with the molecular graph representations enhanced by contrastive learning. The three representation modalities form a ‘local-global-interaction’ hierarchical complementary framework [45], achieved through a cross-attention [46] module. This integrates multimodal molecular knowledge to construct a more robust and universal molecular representation, termed the MCMRL representation (Figure 1a). For a batch of N molecular SMILES, we simultaneously construct a 2D molecular graph (where nodes represent atoms and edges represent chemical bonds) for the 2D encoder, and a 3D molecular structure (containing atomic coordinates and atomic types) for the 3D encoder. The two embeddings of the same molecule form a positive sample pair, while embeddings from different molecules form negative sample pairs (Figure 1b). The NT-Xent loss function is employed to maximize consistency among positive pairs and minimize correlation among negative pairs. The MCMRL model undergoes contrastive learning enhanced using approximately 10 million unlabeled molecules from the PubChem database [47]. The entire fine-tuning process is completed through supervised learning on the target molecular property database. Further details of the MCMRL framework are provided in Section 2.

3.2. Molecular Property Prediction Performance

To validate the effectiveness of MCMRL, we conducted benchmarking tests on multiple challenging classification and regression tasks from MoleculeNet [48].

Table 1. Test performance of different models on seven classification benchmarks.

Dataset	BBBP	Tox21	ClinTox	HIV	BACE	SIDER	MUV
GCN [9]	71.8 ± 0.9	70.9 ± 2.6	62.5 ± 2.8	74.0 ± 3.0	71.6 ± 2.0	53.6 ± 3.2	71.6 ± 4.0
GIN [10]	65.8 ± 4.5	74.0 ± 0.8	58.0 ± 4.4	75.3 ± 1.9	70.1 ± 5.4	57.3 ± 1.6	71.8 ± 2.5
PretrainGNN [24]	68.7 ± 1.3	78.1 ± 0.6	87.6 ± 1.5	71.1 ± 0.5	84.5 ± 0.7	62.7 ± 0.8	80.1 ± 2.1
Attentive FP [14]	64.3 ± 1.8	76.1 ± 0.5	84.7 ± 0.3	75.7 ± 1.4	78.4 ± 0.0	60.6 ± 3.2	76.6 ± 1.5
D-MPNN [15]	71.2 ± 3.8	68.9 ± 1.3	90.5 ± 5.3	75.0 ± 2.1	85.3 ± 5.3	63.2 ± 2.3	76.2 ± 2.8
MGCN [16]	85.0 ± 6.4	70.7 ± 1.6	63.4 ± 4.2	73.8 ± 1.6	73.4 ± 3.0	55.2 ± 1.8	70.2 ± 3.4
GraphMVP [25]	72.4 ± 1.6	75.9 ± 0.5	79.1 ± 2.8	77.0 ± 1.2	81.2 ± 0.9	63.9 ± 1.2	77.7 ± 1.9
FG-BERT [26]	70.2 ± 0.9	78.4 ± 0.8	83.2 ± 1.6	77.4 ± 1.0	84.5 ± 1.5	64.0 ± 0.7	75.3 ± 2.4
MCMRL	74.1 ± 0.6	79.7 ± 1.2	91.3 ± 1.8	80.3 ± 2.1	84.6 ± 0.4	67.6 ± 0.7	81.8 ± 1.0

The mean and standard deviation of the ROC-AUC (%) for each benchmark were reported. Best-performing methods for each benchmark are marked in bold.

presents the comparison of ROC-AUC (%)—the area under the receiver operating characteristic curve—between our MCMRL model and supervised learning as well as self-supervised/pre-trained baseline models across classification tasks. The mean and standard deviation of three independent runs are reported. Compared to other self-supervised learning or pre-training strategies, the MCMRL framework achieves superior performance on five out of seven benchmarks. These improvements demonstrate that MCMRL provides a powerful self-supervised learning strategy for MRL.

Table 2. Test performance of different models on six regression benchmarks.

Dataset	FreeSolv	ESOL	Lipo	QM7	QM8	QM9
GCN [9]	2.87 ± 0.14	1.43 ± 0.05	1.43 ± 0.05	122.9 ± 2.2	0.037 ± 0.001	5.796 ± 1.969
GIN [10]	2.76 ± 0.18	1.45 ± 0.02	0.85 ± 0.07	124.8 ± 0.7	0.037 ± 0.001	4.741 ± 0.912
PretrainGNN [24]	2.76 ± 0.02	1.10 ± 0.06	0.74 ± 0.10	113.2 ± 6.0	0.020 ± 0.002	4.081 ± 0.001
Attentive FP [14]	2.07 ± 0.18	0.98 ± 0.02	0.72 ± 0.10	72.0 ± 2.7	0.018 ± 0.001	2.156 ± 0.001
D-MPNN [15]	2.18 ± 0.91	0.98 ± 0.26	0.65 ± 0.05	105.8 ± 13.2	0.018 ± 0.002	3.241 ± 0.119
MGCN [16]	3.35 ± 0.01	1.27 ± 0.15	1.11 ± 0.04	77.6 ± 4.7	0.022 ± 0.002	0.050 ± 0.002
GraphMVP [25]	—	1.029	0.681	—	—	—
FG-BERT [26]	—	0.94 ± 0.03	0.66 ± 0.01	—	—	—
MCMRL	1.79 ± 0.25	0.91 ± 0.01	0.65 ± 0.05	89.0 ± 3.2	0.017 ± 0.001	2.095 ± 0.216

The mean and standard deviation of the test RMSE (FreeSolv, ESOL, Lipo) or MAE (QM7, QM8, and QM9) are reported. The best-performing methods for each benchmark are marked in bold; —means no result reported for the corresponding models.

presents the performance of MCMRL and baseline models on regression benchmarks. FreeSolv, ESOL, and Lipo employ root mean square error (RMSE) as the evaluation metric, while QM7, QM8, and QM9 follow MoleculeNet’s recommendation to use mean absolute error (MAE). Compared to classification tasks, regression tasks are more challenging as they deal with manually defined discrete labels. Table 2 reveals the following observations: (1) Across six benchmarks, MCMRL outperforms on four benchmarks and achieves near-equivalent performance on the remaining QM7 and QM9 benchmarks. Compared to the Attentive FP model, MCMRL demonstrates superior performance on all five regression datasets. For instance, on the FreeSolv and ESOL databases, it achieves 28% and 7% improvements over the Attentive FP approach, respectively. (2) Compared to supervised learning models, MCMRL demonstrates competitive performance in most scenarios. For instance, on the Lipo dataset, MCMRL achieves results comparable to the top-performing supervised D-MPNN. However, on the QM9 dataset, MCMRL cannot compete with the supervised MGCN. This is because these two models are specifically designed for quantum interactions. Notably, while MGCN excels on datasets related to quantum mechanical properties (i.e., QM7, QM8, and QM9), it does not demonstrate superiority over other supervised learning baselines in the remaining benchmarks. Furthermore, MCMRL pre-training maintains effectiveness on the challenging QM9 benchmark. On the QM9 dataset, MCMRL outperforms other supervised learning baseline models, validating its efficacy.

Table 1 and Table 2 both demonstrate that in certain scenarios, MCMRL exhibits superior predictive accuracy compared to other pre-trained/self-supervised learning benchmarks. Notably, MCMRL benefits from pre-training on large-scale unlabeled databases, and the application of unlabeled data grants MCMRL significant advantages over other benchmark methods in terms of chemical space generalization capability and molecular property diversity.

3.3. Evaluation of Contrastive Learning Pre-Training and Multimodal Fusion

To validate the enhancement of graph representation performance in MCMRL models through cross-dimensional contrastive learning augmentation, we compared MCMRL molecular graph embeddings generated with and without contrastive learning augmentation. The dataset comprises two representations for 1000 molecules, including 1000 positive pairs and 999,900 negative pairs (where 2D–3D feature pairs from the same molecule are positive examples, and pairs from different molecules are negative examples). We computed the cosine similarity between each feature pair, defined as ${\displaystyle \frac{(\mathrm{u},\mathrm{v})}{‖\mathrm{u}‖‖\mathrm{v}‖}}$ [49]. Statistical analysis indicates (as shown in Figure 2) that within the graph representations of the MCMRL model prior to contrastive learning reinforcement, similarity values for positive molecular pairs predominantly cluster between 0.4 and 0.8, with an average of approximately 0.6. Conversely, negative molecular pairs exhibited similarity values concentrated between 0.3 and 0.7, averaging approximately 0.5. This indicates that prior to cross-dimensional contrastive learning enhancement, the graph representations of MCMRL struggled to distinguish diverse molecules. As both representations employ atomic type encoding, they share a limited amount of identical information. Consequently, the presence of relatively high similarity among positive samples before contrastive learning enhancement is a normal phenomenon. For positive sample pairs in the MCMRL model’s graph representations after contrastive learning reinforcement, similarity values predominantly cluster within the 0.8 to 1.0 range, with an average of approximately 0.93. Negative sample pairs, conversely, exhibit similarity values predominantly distributed between −0.2 and 0.2, with an average approaching zero. This conclusively demonstrates that our contrastive learning enhancement enables the MCMRL model’s graph representations to fully learn similarities among positive samples while distinguishing differences among negative samples. This outcome robustly validates the significant enhancement of cross-dimensional contrastive learning to the performance of MCMRL model graph representations.

To systematically evaluate the performance of different molecular representation methods, we compared multiple combinations including 2D features, 3D features, fingerprint (FP) features, their fusions, and variants with certain components removed (no 2D, no 3D, no FP, no fusion). The two figures present evaluation results using ROC-AUC for classification tasks and RMSE for regression tasks across various benchmark datasets. As shown in Figure 3, full MCMRL achieves the best average performance across the vast majority of molecular property prediction tasks, encompassing both classification and regression. This excellence stems from its comprehensive fusion of 2D, 3D, and FP multimodal features, enabling the capture of complementary molecular information from multiple dimensions. Although its advantage is less pronounced on a few classification datasets like SIDER, and single 2D features perform poorly in esol regression tasks, these instances precisely highlight the importance of multimodal information when tackling complex tasks. Therefore, this experiment demonstrates that removing any feature category or fusion module degrades model performance, confirming the irreplaceable role of our topological, geometric, and molecular fingerprint information alongside the fusion module in providing critical insights.

3.4. Investigation on MCMRL’s Graph Representation

To validate that our contrastive learning enhancement method enables MCMRL to fully learn topological and spatial geometric information, demonstrating that the multimodal contrastive learning-enhanced graph representations possess robust representational capabilities, we employed t-distributed stochastic neighbor embedding (t-SNE) [50] to analyze the graph representations. The t-SNE algorithm maps closely related molecular representations to adjacent points in a 2D space. Figure 4 displays the t-SNE 2D embedding plot for 100,000 molecules from the validation set, color-coded by molecular weight. We have also added randomly selected molecules to illustrate how MCMRL’s graphical representations distinguish between similar and dissimilar molecules. For example, the top three molecules above share a similar structure, each featuring a chlorine atom bonded to a benzene ring. The bottom five molecules all contain a benzene ring and a carbonyl group. This demonstrates that even without downstream task fine-tuning, our pre-training method enables the model to learn intrinsic relationships between molecules—molecules with similar properties often share comparable features.

To further evaluate MCMRL’s graph representation capability, we compare its graph representations with traditional molecular FP scores (e.g., ECFP and RDKFP). Specifically, given a query molecule, we extract its graph representation via MCMRL and compute its cosine distance from all reference molecules in the pre-trained database. All reference molecules were then sorted by representation distance and uniformly divided into 20 buckets based on their sorted percentile. Lower percentile thresholds indicate molecules more closely related to the query, as their MCMRL’s graph representations are closer. Within each bin, 5000 molecules were randomly selected, and their dice F-score similarity with the query was computed. Figure 5 shows the mean and standard deviation of FP similarity within each bin. ECFP tends to yield lower similarities than RDKFP, as the former encompasses a broader range of features related to molecular activity. However, both ECFP and RDKFP similarities decrease as MCMRL’s graph representation distances increase. The mean RDKFP similarity for the top 5% is ~0.91, dropping to ~0.53 for the bottom 5%. Similarly, the average ECFP similarity decreased from ~0.50 in the top 5% to ~0.27 in the bottom 5%. Despite fluctuations with increasing percentage thresholds, the overall trend of MCMRL’s graph representation aligns with chemical FP similarity, indicating that distances between MCMRL graph representations effectively reflect molecular similarity. Furthermore, based on MCMRL’s graph representations similarity, molecules closer to the query exhibit high similarity, while distant molecules show even lower similarity than RDKFP. The overall decline trend is more pronounced than in both molecular fingerprints, indicating that our MCMRL’s graph representations possess stronger molecular representation capabilities than these two fingerprints and more effectively reflect molecular similarity.

3.5. Case Study: Potential Drug for DRD2

To further demonstrate the efficacy of the molecular representation by MCMRL, a case study of screening potential drugs for DRD2 was conducted. The 6CM4 crystal structure downloaded from the PDB database is a complex of the DRD2 with risperidone (8NU). The receptor portion of this structure was used for the docking study, and the co-crystallized ligand risperidone was extracted for subsequent docking validation. We employed the DRD2 ligand 8NU [51] as the query molecule, calculating the latent space distances between 8NU and 100,000 randomly selected molecules based on their MCMRL representations. These distances were then sorted and divided into 10 bins. We employed molecular docking technology (AutoDock Vina Version 1.1.2) [52] to calculate the binding affinity of 100 randomly selected molecules from each bin with the DRD2 receptor. These results were compared against the experimentally resolved DRD2-8NU complex structure 6CM4 [53]. As shown in Figure 6, binding affinity increases with the potential spatial distance from the 8NU molecule. A clear overall upward trend is evident. Notably, molecules in the first cluster exhibit significantly lower binding affinities than subsequent clusters, with an average of −11.78—demonstrating a decisive advantage. These molecules share the highest similarity with 8NU. This study successfully identified high-affinity DRD2-binding molecules based on MCMRL feature similarity, fully validating the powerful representation capabilities of MCMRL.

Next, Figure 7 presents a 2D schematic diagram of the 10 molecules most closely related to 8NU within the MCMRL representation domain. We used PyMOL (Version 3.1.6.1) [54] to generate molecular docking structures based on the binding affinities of these 10 molecules with the DRD2 protein, while annotating their binding affinities to DRD2. The binding affinities of these 10 molecules to DRD2 ranged from −12.7 to −10.5, all highly consistent with 8NU’s affinity value of −12.1. Observation of these selected molecules reveals significant structural and functional group similarities, with the lead molecule differing from 8NU by only one hydrogen atom. Through contrastive learning on large-scale unlabeled datasets, MCMRL automatically embeds molecules into a representative feature space and distinguishes compounds in a chemically meaningful manner, further demonstrating its ability to learn chemically meaningful representations. More details of the molecular docking procedure are as shown in Appendix D.

4. Discussion

In this study, we explore the molecular representation learning method enhanced by multimodal contrastive learning. The contrastive learning mechanism in the proposed MCMRL enables the model to capture fine-grained correlations between molecular 2D topological and 3D geometric features, with the cosine similarity of positive sample pairs elevated to 0.93 after pre-training. This improvement essentially stems from the model’s learning of the fundamental chemical rule that topological structure determines spatial geometry, and spatial geometry reflects physicochemical properties, which addresses the critical limitation of existing methods in inadequately capturing structure–property correlations. For multimodal feature fusion, the 2D topological features ensure the accuracy of local bonding information (e.g., atomic connections and bond types) of molecules, the 3D geometric features characterize the spatial steric hindrance and intermolecular interaction potential that dominate molecular biological activity and physicochemical behavior, and the molecular fingerprint features introduce prior chemical knowledge of functional groups, substructures and pharmacophore distributions. The multimodal characterization of molecular properties through the fusion of these three types of features constitutes the core reason for MCMRL achieving superior performance on 9 out of 13 benchmark datasets for molecular property prediction.

Despite its promising performance, MCMRL still has certain limitations that need to be improved. First, the generation of 3D molecular conformation relies on classical molecular mechanics force fields, which leads to limited prediction accuracy for macromolecules with complex spatial structures and flexible conformations. Secondly, the model’s prediction performance for quantum mechanical properties (e.g., dipole moment of QM9) is still inferior to MGCN, a model specially designed for quantum interaction modeling, due to the lack of a dedicated feature extraction module for quantum chemical features.

Given the above limitations, some promising directions of MRL can be further investigated as future works. For example, the 3D conformational space, or at least multiple conformers of the molecule, could be incorporated to improve the accuracy of 3D feature characterization. In addition, quantum chemical descriptors, together with more generic fusion and enhanced strategy for the multimodal information, would contribute to improve the efficacy and accuracy of MRL.