A molecule is not a single static entity; it is a complex, multi-faceted object. To a computer, it can be a 1D string of text characters (SMILES), a 2D topological graph of atoms and bonds, a 2D visual depiction, or a dynamic 3D spatial conformation governed by quantum mechanics.
Traditional molecular machine learning models typically pick one of these representations and treat it as exhaustive. However, as Zhichun Guo et al. note in Graph-based Molecular Representation Learning, restricting a model to a single perspective forces it to compress multi-dimensional chemical realities into a narrow bottleneck, often leading to information loss. Multi-view molecular representation learning addresses this by explicitly modeling and fusing these diverse facets to build a holistic, robust understanding of chemical structures.
The Spectrum of Molecular Views
To understand why multi-view learning works, we must first look at the distinct, complementary information that different chemical modalities provide:
- 1D Sequences and Text: SMILES strings provide a compact, grammar-based representation of a molecule's composition. Beyond structural strings, natural language descriptions provide high-level semantic context. Carl Edwards et al. demonstrated in Text2Mol: Cross-Modal Molecule Retrieval with Natural Language Queries that aligning text queries with structural representations unlocks cross-modal retrieval capabilities, allowing models to map abstract human concepts (like "solubility" or "toxicity") directly to chemical structures.
- 2D Graphs and Images: 2D graphs map explicit atom-and-bond connectivity, capturing local topology. Interestingly, the authors of MolEmb: Multimodal Large Language Models Can Be Strong Molecular Embedding Models (Xinjian Zhao et al.) show that combining symbolic SMILES with 2D visual depictions provides structural cues that complement sequential information, making embeddings significantly more robust.
- 3D Geometry: Physical properties, chirality, and binding affinities depend heavily on spatial conformations. Weitao Du et al. developed Molecule Joint Auto-Encoding: Trajectory Pretraining with 2D and 3D Diffusion to jointly learn 2D topology and 3D spatial arrangements, capturing the essential physical links between connectivity and shape.
- Higher-Order and Contextual Knowledge: Beyond raw structure, molecules possess macro-level properties. Anatol Ehrlich et al. utilize abstract representations like junction trees and extended reduced graphs in XIMP: Cross Graph Inter-Message Passing for Molecular Property Prediction. Similarly, Xiaoyi Liu et al. incorporate pharmacophore hypergraphs in HyperPhS: a pharmacophore-guided multimodal representation framework for metabolic stability prediction through contrastive hypergraph learning, while Mukun Chen et al. integrate external knowledge graphs to capture elemental and pharmacological context in Knowledge-aware contrastive heterogeneous molecular graph learning.
[1D SMILES String] <---\
[2D Molecular Graph] <----+- [Multi-View Fusion Engine] ---> [Holistic Embedding]
[3D Conformation] <---/
How Fusion Works: Aligning and Intersecting Views
Simply concatenating vectors from different encoders is rarely enough; the model must understand how the views relate to one another. Two primary mechanisms drive this integration:
Contrastive Alignment
One approach is to force different views of the same molecule into a shared semantic space. Yaosen Min et al. argue in Molecular Graph Contrastive Learning with Parameterized Explainable Augmentations that expanding beyond simple dual-view frameworks to richer contrastive objectives captures a wider diversity of chemical features.
For instance, Sumin Ha et al. introduce MV-CLAM: Multi-View Molecular Interpretation with Cross-Modal Projection via Language Model, which uses a specialized Multi-Querying Transformer alongside fine-grained contrastive loss to bridge 1D, 2D, and 3D views. By aligning these representations, the model ensures that geometric properties like chirality are preserved alongside topological connectivity.
Cross-View Attention and Message Passing
Instead of just aligning final embeddings, models can allow views to communicate dynamically during encoding. In DeepRSMA: a cross-fusion-based deep learning method for RNA–small molecule binding affinity prediction, Zhijian Huang et al. employ a cross-fusion attention mechanism to integrate sequence and graph views.
Similarly, Runze Wang et al. propose MolHF: Molecular Heterogeneous Attributes Fusion for Drug-Target Affinity Prediction on Heterogeneity, fusing local atomic associations, global topology, and sequential bond arrangements via attention to capture complex chemical behaviors that individual representations miss. Jiali Tang et al. achieve this in CROP: Integrating Topological and Spatial Structures via Cross-View Prefixes for Molecular LLMs by using a "cross-view prefix" mechanism guided by an LLM to dynamically fuse topological and spatial views.
Mitigating Shortcut Learning
A critical, intuitive reason why multi-view learning works is that it prevents models from exploiting superficial correlations, known as shortcut learning. Single-view models often suffer from this flaw: a graph neural network might memorize specific subgraphs that happen to correlate with a target property in a small training set, without learning actual chemical principles.
However, multi-view learning must be designed carefully. Runhan Shi et al. point out in Prediction of chemical reaction yields with large-scale multi-view pre-training that simply piling on representations can cause information redundancy, where the model still finds superficial shortcuts. The key is identifying complementary views that provide distinct information. When done correctly, forcing a model to reconcile a 1D string, a 2D graph, and a 3D conformer ensures that it extracts robust, invariant chemical laws rather than dataset artifacts.
Generalization and the "Cold-Start" Advantage
The ultimate test of a molecular representation model is its ability to generalize to unseen data. In drug discovery, "cold-start" scenarios—where a model must predict binding affinities for entirely novel drugs or targets—are notoriously difficult.
By leveraging multi-view pre-training, models achieve significantly better generalizability. Xike Ouyang et al. demonstrate this in Improving generalizability of drug-target binding prediction by pre-trained multi-view molecular representations, showing that fusing global sequence-based features with local structural graphs allows models to maintain high predictive accuracy even when data is scarce.
This is echoed by Yanqing Niu et al. in MiRNA-Drug Resistance Association Prediction Through the Attentive Multimodal Graph Convolutional Network and Bozitao Zhong et al. in ReactZyme: A Benchmark for Enzyme-Reaction Prediction, where multi-view representations enable flexible, reaction-based retrieval tasks across unseen protein families. By looking at molecules from multiple perspectives simultaneously, these systems learn representations that reflect true chemical identity rather than single-modality biases.
Key Foundational Work
MiRNA-Drug Resistance Association Prediction Through the Attentive Multimodal Graph Convolutional Network (2022) — contributed 1 concept elements to this direction.
Text2Mol: Cross-Modal Molecule Retrieval with Natural Language Queries (2021) — contributed 1 concept elements to this direction.
Molecular Graph Contrastive Learning with Parameterized Explainable Augmentations (2021) — contributed 1 concept elements to this direction.