🧬 biology

FUSION: a multimodal graph-based structural AI framework for interpretable somatic cancer driver mutation prediction

The paper introduces FUSION, a multimodal graph-based AI framework that integrates protein language model embeddings, AlphaFold2-derived structural topology, and conformational ensembles to accurately predict and interpret somatic cancer driver mutations, particularly for rare missense variants lacking prior clinical evidence.

Original authors: Lucas Moraes dos Santos, Tatiane Senna Bialves, Luana Luiza Bastos, José Gutembergue de Mendonça Silva, Sheila Cruz Araujo, Janaina de Oliveira Melo, Adriano de Paula Sabino, Eduardo Tarazona-Santos

Published 2026-08-31

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Lucas Moraes dos Santos, Tatiane Senna Bialves, Luana Luiza Bastos, José Gutembergue de Mendonça Silva, Sheila Cruz Araujo, Janaina de Oliveira Melo, Adriano de Paula Sabino, Eduardo Tarazona-Santos, Leonardo Henrique França de Lima, Gerd Bruno da Rocha, Eduardo Moraes Reis, Wagner Meira Junior, Raquel Cardoso de Melo-Minardi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cancer begins when the body's cells start to grow out of control, often because of tiny errors in their genetic code. These errors, called mutations, can act like a stuck accelerator in a car, pushing the cell to divide when it should rest. Scientists call the dangerous mutations that drive this process "drivers," while the harmless ones that happen to ride along are called "passengers." For decades, finding the drivers has been like searching for a few specific needles in a massive haystack of genetic data. The challenge is that many drivers are rare; they appear in only a handful of patients or in specific parts of a protein, making them hard to spot with standard tools that rely on seeing the same mistake over and over again. Furthermore, understanding why a mutation causes cancer often requires looking at the three-dimensional shape of the proteins it affects, not just the sequence of letters in the DNA.

A team of researchers has developed a new artificial intelligence system called FUSION to solve this problem. Instead of looking at mutations in isolation, this system builds a detailed map of how a protein is built and how it moves. It combines information about the protein's chemical sequence, its folded shape, and its structural flexibility into a single, unified model. By training on thousands of examples, the system learns to recognize the subtle structural signatures that distinguish a cancer-causing driver from a harmless passenger. The results show that this approach can identify dangerous mutations that other methods miss, including rare variants that do not appear in the usual "hotspots" where cancer is most often found. This capability offers a new way to prioritize which genetic errors deserve further study, potentially helping doctors understand the unique biology of a patient's tumor.

The researchers built FUSION to handle the complexity of human proteins, which are long chains of amino acids that fold into intricate three-dimensional shapes. To do this, they fed the system three different types of information. First, it uses a language model trained on millions of protein sequences to understand the chemical context of each amino acid. Second, it incorporates the physical shape of the protein, generated by a powerful structure-prediction tool that acts like a molecular architect. Third, it includes details about the local environment, such as whether a specific part of the protein is a rigid helix or a flexible loop. The system then connects all these pieces into a network, where every amino acid is a node and the physical distances between them are the links. This allows the AI to see how a change in one spot might ripple through the entire structure.

A key innovation in this work is how the system handles the fact that proteins are not static statues; they wiggle and shift. To capture this movement, the researchers used a generative technique to create thousands of slightly different versions of each protein structure, simulating the natural flexibility of the molecule. They used these variations to teach the AI what a protein looks like when it is in motion, rather than just in one frozen pose. However, when the system makes a final prediction for a new patient, it only needs the single, most likely structure. The training on many moving parts simply helps the model learn the underlying rules of how proteins behave, making it more robust when it encounters a new, unseen mutation.

The team tested FUSION on two major challenges. First, they evaluated it against a large collection of mutations that had been experimentally verified in the lab to see if they caused cancer. In these tests, the system correctly identified driver mutations with high accuracy, outperforming existing methods. It was particularly good at spotting rare variants that appeared only once or twice in large datasets, which are often overlooked by tools that rely on frequency. For example, the system correctly identified a rare mutation in a gene called POT1 as a driver, a finding that aligned with biological evidence about its role in protecting DNA, even though the mutation was too rare for other methods to flag.

The researchers also applied FUSION specifically to pancreatic ductal adenocarcinoma, a deadly form of pancreatic cancer. In this context, the system achieved an even higher level of accuracy, correctly distinguishing drivers from passengers in the vast majority of cases. It successfully identified known drivers in genes like KRAS and TP53, but it also highlighted rare mutations in other genes that had not been previously linked to the disease. One notable discovery was a mutation in a gene called SASH1. This mutation was located in a flexible, disordered region of the protein, far from the usual hotspots where cancer mutations are found. The system flagged it as a potential driver, and further analysis suggested it could disrupt the protein's ability to stop tumor growth, a finding that opens a new avenue for investigation in pancreatic cancer.

Beyond just making predictions, the system offers a way to see why it made those choices. The researchers examined the internal "attention" of the model, which shows which parts of the protein the system focused on when making a decision. For a common mutation in the KRAS protein, the system paid close attention to specific regions known to be critical for the protein's function, including areas that interact with other molecules to control cell growth. It also highlighted a transient pocket near the mutation site, a structural feature that has recently become a target for new cancer drugs. This ability to point to biologically meaningful regions suggests that the system is not just guessing based on patterns in the data, but is actually learning the physical rules that govern how proteins work.

The study demonstrates that combining different types of molecular information into a single graph-based model can significantly improve the detection of cancer drivers. By moving beyond simple lists of mutations and instead modeling the protein as a dynamic, interconnected structure, FUSION provides a more complete picture of how genetic errors lead to disease. While the system is not a cure in itself, it serves as a powerful tool for sorting through the overwhelming amount of genetic data generated by modern sequencing. It helps researchers and clinicians focus their attention on the mutations that are most likely to be driving the cancer, paving the way for more targeted therapies and a deeper understanding of the disease. The work suggests that the future of precision oncology lies in tools that can interpret the complex, three-dimensional language of life, turning raw genetic data into actionable biological insight.

Technical Summary: FUSION – A Multimodal Graph-Based Structural AI Framework

Problem Statement
Accurate identification of somatic cancer driver mutations remains a critical bottleneck in precision oncology, particularly for rare missense variants lacking recurrence evidence or clinical annotation. Existing machine learning approaches often rely on fragmented molecular information, typically utilizing only sequence-derived features or engineered structural descriptors. These methods frequently fail to capture the multiscale nature of mutational effects, which emerge from the interplay between protein sequence, 3D structure, conformational dynamics, and molecular interactions. Furthermore, the scarcity of experimentally validated labels and the long-tailed distribution of somatic mutations (where rare candidates coexist with established hotspots) complicate the distinction between driver and passenger events. Current models often struggle to generalize beyond established mutational hotspots, limiting their utility in prioritizing non-canonical, biologically relevant variants.

Methodology
The authors introduce FUSION, a multimodal artificial intelligence framework designed to integrate heterogeneous molecular data into a unified geometric protein representation. The methodology is structured as follows:

Multimodal Graph Representation: Each protein variant is modeled as an attributed graph $G = (V, E)$ $G = (V, E)$ , where nodes represent amino acid residues and directed edges encode residue–residue relations derived from predicted 3D structures.
- Node Features: Node embeddings are constructed by concatenating: (i) per-residue embeddings from the 33rd layer of the ESM-2 protein language model (capturing biochemical and evolutionary context); (ii) a binary mutation-site indicator; and (iii) learned embeddings for categorical secondary structure annotations, specifically including $\kappa$ -helices (irregular conformations enriched in driver regions).
- Edge Features: Edges encode local geometry via the Euclidean distance between $C_\alpha$ atoms ( $d_{ij} \le 8.0$ Å).
Architecture: The framework employs a Graph Attention Network (GATv2) with three layers and multi-head attention ( $H=4$ ). The GATv2 mechanism functions as an early-fusion module, dynamically computing attention coefficients ( $\alpha_{ij}$ ) that weigh message passing between neighboring residues based on both node and edge features. This allows the model to learn which structural and sequence contexts are most relevant for driver prediction.
Data Augmentation via Conformational Ensembles: To address the static nature of single-structure predictions and capture protein dynamics, the authors employ sequence-conditioned generative modeling. During training, AlphaFlow-MD (and explicit Molecular Dynamics for large proteins) generates conformational ensembles for each variant. These ensembles serve as data augmentation, expanding the structural diversity available for learning without requiring time-resolved simulations at inference.
Training Protocol: The model is trained using a variant-level 5-fold cross-validation protocol. To prevent information leakage, all structures derived from the same variant are assigned to a single fold. The training set includes both the static AlphaFold2 (AF2) structures and the augmented conformational ensembles, while validation relies solely on the static AF2 model to ensure uniform representation.

Key Contributions

Unified Multimodal Framework: FUSION integrates sequence embeddings (ESM-2), secondary structure (including $\kappa$ -helices), and residue-contact topology (from AF2) into a single geometric deep learning architecture, moving beyond reliance on pre-computed descriptors.
Interpretability via Attention: The attention coefficients provide residue-level interpretability, highlighting interactions consistent with known functional regions (e.g., Switch I/II regions in KRAS) and allowing the model to prioritize variants based on structural context rather than recurrence frequency alone.
Conformational Augmentation: The integration of generative conformational ensembles during training allows the model to learn from a broader distribution of structural states, approximating protein dynamics to improve generalization.
Performance on Rare Variants: The framework demonstrates an ability to prioritize plausible non-canonical candidates and rare missense variants that are often overlooked by frequency-based approaches.

Results
FUSION was evaluated across pan-cancer benchmarks and a specific Pancreatic Ductal Adenocarcinoma (PDAC) cohort:

Pan-Cancer Benchmarks:
- On the Kim et al. (2016) dataset (in vivo tumorigenesis assays), FUSION achieved a mean auROC of 0.875 and a macro F1 score of 0.854, outperforming CHASMplus, Network&AA, and CanDrA. Notably, it correctly classified the rare driver POT1 p.G76V, which other methods missed.
- On the Ng et al. (2018) dataset (in vitro assays), FUSION achieved an auROC of 0.806 and a macro F1 of 0.783, surpassing benchmark methods. It successfully recovered 74.20% of true driver variants at a low false-positive rate (FPR = 0.169), including variants outside canonical hotspots (e.g., BRAF p.H269Y, ERBB2 p.P551L).
PDAC Cohort:
- In the PDAC-specific evaluation, FUSION achieved an auROC of 0.941 and a macro F1 of 0.871, outperforming the cancer-type-specific CHASMplus PAAD annotator (auROC 0.917).
- Case Studies:
  - KRAS: Attention analysis for the KRAS p.G12D variant revealed high attention weights on residues in the Switch I and Switch II regions (e.g., Gln61, Pro34), consistent with known allosteric regulatory interfaces and druggable pockets (S-IIP).
  - TP53: The model demonstrated high sensitivity for driver variants located outside canonical hotspot regions (e.g., p.Y205C/D, p.G245A/R/S).
  - SASH1: FUSION prioritized SASH1 p.D564Y, a variant located in a $\kappa$ -helix within an intrinsically disordered region, suggesting a potential novel driver role in PDAC.

Significance and Claims
The paper claims that FUSION supports the hypothesis that cancer driver mutations are multiscale phenomena requiring the integration of sequence, structure, and conformational dynamics. By leveraging multimodal graph learning, the framework improves both predictive performance and biological interpretability. The authors assert that FUSION offers a translational strategy for the prioritization of candidate somatic variants, particularly rare and non-canonical alterations, thereby aiding functional studies and clinical decision-making in precision oncology. The work highlights the limitations of hotspot-centric approaches and emphasizes the value of integrative sequence- and structure-based analyses for identifying actionable alterations in complex malignancies like PDAC. The authors note that future directions include extending the framework to germline variants, integrating multi-omics data, and expanding training cohorts to include underrepresented populations to enhance global applicability.

More like this