MutationProjector

Targeted DNA sequencing panels are now routine in cancer care, yet most of the information they yield goes unused. Only about 8% of patients are matched to an FDA-approved targeted therapy on the basis of sequencing, usually via a single well-studied gene, while the average sequenced tumor carries roughly eleven alterations — events that are individually too rare to associate with outcomes and that act combinatorially rather than in isolation. MutationProjector, developed by the Ideker lab at UC San Diego with collaborators at Lunit, pretrains one model on tens of thousands of tumor genotypes and reuses its learned embedding across many clinical questions.

The model compresses a tumor's panel-level profile — somatic mutation, copy-number amplification and deletion status across the 468 genes of MSK-IMPACT, plus tumor mutation burden, aneuploidy and dominant mutational signature — into a compact representation of tumor subtype. Attention is constrained by eight curated molecular interaction networks, so signal propagates between genes that physically bind, transcriptionally regulate, phosphorylate, ubiquitinate or genetically interact. That inductive bias lets rare, scattered alterations aggregate into shared pathway-level evidence.

Posted to bioRxiv in September 2025, it follows a one-model/N-tasks design rather than fitting a separate classifier per clinical question, echoing the transfer scheme of single-cell foundation models such as scFoundation.

#Key Features

• Network-constrained graph attention: Each encoder unit runs eight attention heads, each restricted to one interaction network — physical binding, transcriptional regulation, phosphorylation, ubiquitination, genetic interaction, DNA damage response, STRING v12 and PCNet v1.3.
• Multi-task pretraining: A masked-gene self-supervised objective is trained jointly with supervised prediction of cancer type and of tumor-infiltrating lymphocyte status, a mixed strategy also used by AlphaMissense and DNAGPT.
• Covariates as global tokens: Tumor mutation burden, aneuploidy and mutational signatures connect to every gene node in every attention head, broadcasting sample-level context across the graph.
• Attention-based interpretability: Attention-weighted ridge probing recovers the features driving each prediction, surfacing biomarkers such as KMT2A alterations and KRAS–STK11 co-alterations in immunotherapy response.
• Small-cohort transfer: A random forest fit on the frozen embedding solves each downstream task; the immunotherapy classifier needed just 94 labeled patients.
• Open release: Code, pretrained weights and transfer-learned task heads ship on GitHub and Zenodo under the MIT license.

#Technical Details

The architecture loosely follows a Transformer encoder but replaces fully connected attention with GATv2 graph attention layers (via PyTorch Geometric), stacked in two encoder units with residual connections and self-loops disabled. Each gene is encoded as the sum of a learnable identity token and a 10-dimensional mutation embedding capturing its alteration state; continuous covariates are discretized into five bins. Pretraining used 30,328 solid tumors from AACR Project GENIE and TCGA across ten types, from lung adenocarcinoma (n=8,445) down to ovarian cancer (n=87); lymphocyte-infiltration labels came from a ResNet-based classifier applied to TCGA hematoxylin-and-eosin slides. Training ran 100 epochs with AdamW (batch 64, learning rate 0.001, dropout 0.1) on NVIDIA V100 GPUs, minimizing a class-weighted binary cross-entropy loss over the three tasks.

On a held-out 20% split (n=6,066), masked-alteration recovery reached odds ratios of 18.8 for somatic mutations, 41.7 for amplifications and 58.3 for deletions, significantly better than a matched transformer encoder without network knowledge; cancer-type odds ratios ranged from 16.2 (esophageal) to 698.3 (melanoma). Across 2,978 patients in downstream cohorts, a chemotherapy classifier trained on 237 cisplatin-treated tumors and tested on 42 bladder tumors reached a hazard ratio of 0.43 against 0.60 for the strongest baseline; immunotherapy classifiers separated progression-free and overall survival in independent bladder, lung and melanoma cohorts; and a lung adenocarcinoma metastasis classifier achieved an odds ratio of 11.1.

#Applications

MutationProjector targets translational oncology groups holding clinical panel sequencing but lacking the cohort size to train a bespoke predictor per question. With the encoder frozen and a lightweight classifier on top, a lab can fit a new outcome model from a few hundred annotated patients; released heads already cover chemotherapy response, immunotherapy response, lung adenocarcinoma metastasis, and tissue-of-origin scoring for breast, colorectal and lung tumors. The embedding also supports unsupervised work: finding molecularly similar patients, recovering basal versus luminal transcriptional subtypes, and separating HPV-positive head-and-neck and cervical cancers without ever seeing viral status.

#Impact

MutationProjector shows that the pretrain-then-transfer recipe from language and single-cell modeling carries over to sparse, panel-scale cancer genotypes when prior biological knowledge is built into the attention pattern. Its attention-based feature attribution answers a standing objection in clinical AI, recovering established markers such as tumor mutation burden alongside less obvious ones such as chromatin-remodeling co-alterations with KRAS. Limitations are real: pretraining spans only ten solid tumor types, downstream test cohorts number in the tens to low hundreds, all evidence to date is retrospective, and the preprint has not completed peer review. The authors point to pancreatic, prostate and sarcoma cohorts, ICGC data, and added modalities as the next expansion.