Multi-view biomedical foundation models for molecule-target and property prediction

This paper introduces MMELON, a multi-view biomedical foundation model that integrates graph, image, and text representations of molecules to robustly predict properties and identify potential Alzheimer's disease treatments across over 120 tasks.

The Gist

Imagine you are trying to find the perfect key to open a very specific, complex lock. In the world of medicine, the "lock" is a disease-causing protein in your body, and the "key" is a new drug molecule. For decades, scientists have tried to predict which keys fit which locks using computers, but it's been like trying to guess the shape of a key by only looking at a single photo of it. Sometimes the photo helps, sometimes it doesn't.

This paper introduces a new, super-smart computer system called MMELON (Multi-view Molecular Embedding with Late Fusion) that solves this problem by looking at molecules from three different angles at once, just like a human would.

Here is the simple breakdown of how it works, using some everyday analogies:

1. The Problem: One View Isn't Enough

Imagine you are trying to describe a famous building (like the Eiffel Tower) to a friend who has never seen it.

• The "Text" View: You describe it using words: "It's a tall iron tower with a lattice structure." (This is like SMILES, a text code scientists use for molecules).
• The "Graph" View: You draw a map of how the metal beams connect to each other. (This is like a Graph, showing atoms and bonds).
• The "Image" View: You show them a photograph of the tower. (This is like a 2D Image of the molecule).

Previous computer models usually picked just one of these descriptions. If they only used the text, they might miss how the structure actually looks. If they only used the photo, they might miss the chemical rules of how the atoms connect.

2. The Solution: The "All-Seeing" Team

The researchers built a team of three expert AI "scouts," each trained on a massive library of 200 million molecules.

• Scout 1 (Text): Reads the chemical "recipe."
• Scout 2 (Graph): Analyzes the chemical "blueprint."
• Scout 3 (Image): Studies the chemical "photograph."

Instead of forcing them to agree on one answer immediately, the MMELON system lets them work independently first. Then, it brings them together in a "meeting room" (the Aggregator).

3. The "Late Fusion" Meeting

This is the clever part. When the system needs to predict if a molecule will work as a drug, it asks the three scouts for their opinions.

• The system doesn't just average their answers. It learns who to trust more for the specific job at hand.
• Think of it like a jury. For a task involving complex 3D shapes, the "Image" scout might get a higher vote. For a task involving chemical reactions, the "Graph" scout might get the loudest voice.
• The system assigns a "weight" to each scout. If the Image scout is right 90% of the time for a specific disease, the system listens to them more. If the Text scout is better for another task, it listens to them instead.

This "Late Fusion" means the system is flexible. It doesn't force a single way of thinking; it combines the strengths of all three.

4. The Real-World Test: Fighting Alzheimer's

To prove this system works, the researchers used it to hunt for new treatments for Alzheimer's disease.

• The Target: They looked at 33 specific proteins (GPCRs) linked to Alzheimer's.
• The Search: They scanned through thousands of existing drugs and natural compounds found in our gut (metabolites) to see which ones might stick to these proteins like a key in a lock.
• The Discovery: The AI found some very promising candidates.
  • It identified a gut chemical called acetyl-glutamine that might interact with a protein called FPR1.
  • It found that Glutathione (a powerful antioxidant we get from food) could also be a strong binder.
• The Proof: The researchers didn't just trust the computer. They used 3D modeling to visualize exactly how these molecules would sit inside the protein. The computer's "guess" matched perfectly with the physical reality, showing that the AI had correctly identified the "key" features.

Why This Matters

In the past, if a computer model failed to find a drug, scientists might have had to start over with a different method. With MMELON, the system is robust. Even if one "scout" (like the text reader) is confused, the other two (the image and graph readers) can still guide the system to the right answer.

In short: This paper describes a new AI that doesn't just "read" or "see" molecules; it understands them deeply by combining text, maps, and pictures. This helps scientists find new drugs faster, cheaper, and with more confidence, potentially leading to better treatments for diseases like Alzheimer's.

Technical Summary

1. Problem Statement

Drug discovery is a complex, costly process with low success rates, heavily reliant on predicting the chemical and biological properties of candidate molecules. While machine learning has aided this process, a significant bottleneck remains in molecular representation.

• Limitation of Single Views: Existing foundation models typically rely on a single representation (e.g., text/SMILES, 2D graphs, or images). Each view has inherent strengths and weaknesses; for instance, sequence-based models capture human-generated knowledge well but may miss intrinsic geometric symmetries, while graph models capture topology but struggle to scale to massive parameter counts without specific architectural adjustments.
• Data Scarcity: Learning generalized latent representations is difficult due to limited labeled data for specific downstream tasks and the vastness of chemical space.
• Gap: There is a lack of scalable, multi-modal foundation models that effectively integrate complementary views (text, graph, image) on massive datasets (>100M molecules) to improve robustness across diverse drug discovery tasks.

2. Methodology: MMELON

The authors propose MMELON (Multi-view Molecular Embedding with Late Fusion), a flexible architecture that aggregates multiple molecular views within a foundation model setting.

A. Single-View Encoders

Three distinct pre-trained encoders are developed, each based on a different modality:

1. Text View: Based on MolFormer. It processes SMILES strings using a Transformer architecture with rotary positional embeddings and linear attention.
2. Graph View: Based on TokenGT. It treats atoms and bonds as tokens in a sequence. A novel pre-training task, Betti number prediction, is introduced to capture topological features (connected components and cycles) alongside standard node masking and edge prediction.
3. Image View: Based on ImageMol. It uses a CNN (ResNet-18 backbone) to process 2D molecular images, capturing spatial arrangements, symmetry, and functional groups.

B. Pre-training Strategy

• Dataset: The Graph and Text models are pre-trained on a curated dataset of 200 million molecules derived from PubChem and ZINC22. The Image model is pre-trained on 10 million compounds from PubChem.
• Tasks:
  • Text: Masked Language Modeling (MLM).
  • Image: Multi-granularity clustering, rationality discrimination, jigsaw puzzle prediction, and image reconstruction.
  • Graph: Node feature masking, edge prediction, and the novel Betti number prediction (topological invariants).

C. Late Fusion Architecture

Instead of early fusion (combining inputs), MMELON employs a Late Fusion strategy:

1. Independent Encoders: The three single-view models act as pre-trained encoders.
2. Aggregator Module: An attention-based sub-network combines the embeddings ($z_m$) from the three views into a joint multi-view embedding ($z^{mv}$).
   • The contribution of each view is weighted by a learned coefficient $\alpha_m$: $z^{mv} \propto \sum \alpha_m z_m$.
   • The weights are derived via a soft-max operation on the embeddings, allowing the model to dynamically adjust the importance of each view based on the downstream task.
3. Secondary Pre-training: The aggregator is pre-trained on a 10M molecule subset using an embedding reconstruction task to initialize weights before fine-tuning.

3. Key Contributions

• MMELON Architecture: A novel multi-view foundation model that successfully integrates Graph, Image, and Text modalities using a learnable late-fusion mechanism.
• Scale: Pre-training on 200M molecules, significantly larger than many previous multi-modal attempts, demonstrating scalability.
• Novel Pre-training Task: Introduction of Betti number prediction for graph transformers to capture global topological context, complementing local edge/node features.
• Interpretability: The attention weights ($\alpha$) provide transparency, revealing which molecular representation (view) is most critical for specific prediction tasks.
• Open Source: The model and code are released on GitHub and Hugging Face.

4. Results

A. Embedding Analysis

• Complementarity: Analysis of Euclidean distances shows that the Image view is the most distinct from the others, while Text and Graph are highly correlated ($c=0.7$).
• Validity: Pre-trained embeddings show moderate correlation with traditional molecular fingerprints (Morgan, MACCS, etc.), validating that the models learn chemically meaningful representations.

B. Benchmark Performance

The model was fine-tuned on over 120 downstream tasks, including:

• MoleculeNet: 21 tasks covering solubility, toxicity, and binding.
• ADME/Tox: Cytochrome P450 (CYP) inhibition and ComputationalADME tasks.
• Performance: The multi-view model demonstrates robustness, matching or closely approaching the performance of the best single-view model (typically the Graph model) across almost all tasks. It avoids the "poor result" scenario seen in some single-view models on specific datasets.
• Drug-Target Interaction: When combined with protein embeddings (ESM), the multi-view model outperformed state-of-the-art GNN implementations on the Davis kinase dataset.

C. Case Study: Alzheimer's Disease (AD) & GPCRs

The authors applied the model to a large-scale virtual screening of 106 G Protein-Coupled Receptors (GPCRs), focusing on 33 GPCRs linked to Alzheimer's Disease.

• Prediction: The model identified strong binders from a set of 515 gut metabolites and 2,504 FDA-approved drugs.
• Validation:
  • FPR1: Identified acetyl-glutamine (a gut metabolite) and Glutathione (GSH) as strong binders. 3D docking confirmed binding modes consistent with known catalytic sites.
  • ADA2A: Identified fructose 1,6-biphosphate and isosorbide dinitrate as top candidates.
• Attention Analysis: The model's attention maps aligned with chemical intuition (e.g., highlighting amide groups in acetyl-glutamine and phosphate/nitro groups in other candidates), confirming the model learns relevant pharmacophores.

5. Significance and Future Directions

• Robustness in Discovery: MMELON provides a reliable tool for drug discovery by mitigating the risks associated with relying on a single molecular representation. It ensures high performance even when a specific view (e.g., text) is suboptimal for a particular task.
• Scalability: The success of late fusion on 200M molecules suggests that multi-modal foundation models can scale effectively, a crucial step for generalizable AI in biomedicine.
• Interpretability: The ability to quantify the contribution of each view ($\alpha$) offers insights into the "why" behind predictions, aiding chemists in understanding model decisions.
• Future Work: The authors note that the current views are 1D/2D. Future iterations will incorporate 3D conformations and binding context to further enhance the representation of molecular geometry and interaction dynamics.

In summary, MMELON represents a significant advancement in biomedical AI by demonstrating that integrating diverse molecular views through a scalable, late-fusion foundation model leads to robust, interpretable, and high-performing predictions for complex drug discovery challenges.