Fusing Graphs and Sequences for Molecular Representation

198 concepts152 papers1 proposal

To a computer, a molecule can be many things. It can be a string of text characters (a SMILES string) capturing its chemical grammar. It can be a two-dimensional graph where nodes represent atoms and edges represent bonds. Or it can be a three-dimensional cloud of coordinates tracking spatial geometry.

Historically, deep learning models for drug discovery and materials science forced a choice. Researchers either used Recurrent Neural Networks or Transformers to process text-like molecular sequences, or they used Graph Neural Networks (GNNs) to process structural topologies.

However, relying on a single modality introduces fundamental blind spots. Purely graph-based models excel at capturing local atom connectivity but often struggle with long-range dependencies, global symmetries, or specific chemical nuances like chirality and ionic states. Conversely, sequence-based models capture chemical notation and global context beautifully but lack explicit awareness of physical, spatial connectivity.

To overcome these limitations, a powerful paradigm has emerged: multi-modal molecular representation. By fusing graphs, sequences, and expert-defined descriptors, modern biochemical AI frameworks build a more holistic, robust, and generalizable understanding of molecular space.


Why Hybrid Representations Work: Complementary Information

The core intuition behind multi-modal molecular learning is that different representations capture orthogonal, complementary facets of chemical reality.

Consider a task like predicting drug-target affinity or chemical reaction yields. A minor structural modification can completely alter a molecule's biological activity or reaction outcome. As Kehan Guo demonstrates in the development of the Uncertainty-Aware Multimodal model (UAM), combining sequence-based and graph-based features allows a network to capture subtle structural nuances that single-modality models overlook, significantly improving yield predictions where minor variations dictate success or failure.

┌──────────────────────────────────────────────────────────┐
│                   MOLECULAR MODALITIES                   │
└──────────────────────────────────────────────────────────┘
             │                      │
             ▼                      ▼
   ┌───────────────────┐  ┌───────────────────┐
   │  SMILES Sequence  │  │  2D/3D Topology   │
   └───────────────────┘  └───────────────────┘
             │                      │
             ▼                      ▼
   ┌───────────────────┐  ┌───────────────────┐
   │    Transformer    │  │    Graph Net      │
   │ (Chemical Grammar)│  │ (Spatial/Bonds)   │
   └───────────────────┘  └───────────────────┘
             │                      │
             └──────────┬───────────┘
                        ▼
             ┌─────────────────────┐
             │ Joint Latent Space  │
             └─────────────────────┘

By combining these perspectives, models achieve a "dual-view" or "multi-view" understanding:

  • The Sequential View: Captures local chemical grammar, functional groups, and textual patterns.
  • The Topological/Geometric View: Captures explicit atom connectivity, bond distances, and three-dimensional spatial relationships.

For instance, Xiangfeng Yan showed in the GSATDTA framework that sequences effectively preserve specific properties like chirality and ionic groups, while graph structures are superior at capturing spatial topology. Fusing them compensates for the individual limitations of each format.


Fusion Architectures: How Modalities Communicate

Integrating text and graphs requires architectures that can translate between disparate data structures. Three primary strategies dominate the literature:

1. Dual-Stream and Multi-Encoder Networks

The most straightforward approach processes each modality through a dedicated encoder before fusing their latent representations. In the DeepDrug framework, Qijin Yin et al. utilize a dual-stream architecture that feeds sequence data into convolutional layers and structural data into a GNN, capturing both linear chemical information and 3D geometric properties simultaneously to predict drug interactions. Similarly, Srivathsan Badrinarayanan et al. designed Multi-Peptide with a dual-encoder setup—a transformer for sequence processing and a GNN for spatial analysis—to map peptide properties.

2. Cross-Modality Contrastive Alignment

Instead of merely concatenating features at the end, some frameworks use contrastive learning to align the latent spaces of different modalities during pre-training. Bing-Xue Du et al. developed CMMS-GCL, which uses graph contrastive learning to maximize the mutual information between a molecule's SMILES string representation and its structural molecular graph. This forces the model to find consistent, underlying chemical concepts across both text and topology.

Taking this further, Ao Shen et al. introduced MoleSG, which employs a non-overlapping masking strategy across both SMILES and graph encoders, using a unified Transformer backbone to reconstruct the masked portions of both modalities, ensuring the model captures complementary information.

3. Cross-Modal Projectors for Large Language Models

With the rise of Large Language Models (LLMs), researchers are now projecting structural graph features directly into the text token space of language models. Zhiyuan Liu et al. developed MolCA, which uses a cross-modal projector to translate 2D graph features into "soft prompts" that an LLM can understand. This allows the language model to ground its pre-existing knowledge of chemical notation in explicit structural topology, leading to superior performance in molecule captioning and IUPAC name prediction. As He Cao et al. demonstrate with InstructMol, anchoring topological data with sequential notation provides a vital structural context that helps language models navigate complex chemical spaces reliably.


Robustness and Generalization

The ultimate test for any molecular machine learning model is out-of-distribution (OOD) generalization—the ability to make accurate predictions for novel molecular scaffolds or entirely new chemical families unseen during training.

Pure GNNs often struggle when encountering novel topologies because their message-passing operations are tightly bound to the specific connectivity patterns of the training data. Augmenting these models with auxiliary features—such as 1D fingerprints, physical metadata, or sequence embeddings—acts as a regularizer.

Chloe Engler Hart et al. demonstrated this effect in Mol2CCS by integrating GNNs with parallel modules processing 1D molecular fingerprints and physical metadata. This multi-modal approach significantly improved the robustness of property predictions when navigating diverse, unfamiliar chemical frameworks.

Similarly, Bharath Ramsundar implemented a hybrid featurization strategy in AMPL, combining traditional circular fingerprints with physicochemical descriptors and learned graph-based latent vectors. This hybrid approach ensures that even if the GNN component encounters an unfamiliar atomic arrangement, the model can fall back on explicit, expert-defined chemical descriptors and global sequence patterns to guide its predictions.

By moving away from single-modality silos and embracing hybrid representations, modern biochemical AI models operate much like human chemists: they read the formula, visualize the 2D structure, and reason about the 3D shape all at once.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_molecular_representation_c7021b71,
  title = {hybrid molecular representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-molecular-representation/rdr_c7021b71e3f84b668dae3fb66551f5d3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_hybrid_molecular_representation_c7021b71_v1,
  title = {hybrid molecular representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-molecular-representation/rdr_c7021b71e3f84b668dae3fb66551f5d3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake