scTransMIL: Transformer MIL Traces Cancer from Sample to Cell to Gene

Overview

scTransMIL is a computational framework published in Nature Communications (September 14, 2026) that connects sample-level cancer classification, cell-level disease relevance scoring, and gene-level attention ranking using a Transformer-based multiple instance learning (MIL) architecture. The method treats a single-cell sequencing sample as a "bag" of cells and learns which cells contribute most to the sample-level label, enabling back-tracing from a positive sample prediction to individual malignant cells and candidate driver genes.

Model Architecture and Training

The model employs a gated encoder and a sample aggregator built on Transformer layers to capture contextual gene co-expression patterns across cells. It performs two-stage prediction: first distinguishing tumor from normal tissue, then predicting the specific cancer type among 30 classes. Training, validation, and test sets are split by donor (999 donors total) to avoid data leakage from the same individual. The dataset comprises approximately 400,000 cells across 1,563 samples (including sub-samples).

Sample-Level Classification Performance

Benchmarking against multiple baselines using Accuracy, AUC, and F1 across varying training data proportions shows scTransMIL's sample-level performance improves with more data and consistently outperforms comparison methods. The model reliably predicts whether a sample is cancerous or normal.

Metastatic Origin Prediction

In an independent breast cancer (BRCA) metastasis cohort of 37 patients and ~40,000 cells, scTransMIL predicted the primary cancer origin for both primary and metastatic samples. The majority of predictions pointed to breast cancer, demonstrating the model's ability to recognize the tissue of origin even after metastasis.

Cell-Level Disease Relevance Scoring

To validate cell-level scores, the authors compared scTransMIL's disease-relevance scores against copy number variation (CNV)-inferred malignant cell annotations. Using CNV as reference, scTransMIL achieved an AUC of 0.71, outperforming three comparison methods (0.67, 0.58, 0.56). This confirms the model can identify malignant cells within a mixed population without explicit cell-type labels.

Early Detection Potential

A thyroid cancer-adjacent sample with near-normal histology received a positive prediction score of 0.64. High scores concentrated in epithelial cells exhibiting pre-malignant features, illustrating scTransMIL's ability to detect sparse malignant signals before overt morphological changes.

Gene-Level Attention and Candidate Discovery

Beyond differential expression, scTransMIL's gene attention mechanism ranks genes by their contextual importance in the learned expression patterns. In acute lymphoblastic leukemia (ALL), the model highlighted BLK , a gene with existing biological support that was missed by standard differential expression analyses. The authors caution that attention ranking suggests candidates but does not prove causality.

Independent Validation in Rheumatoid Arthritis

A separate study in Advanced Science applied scTransMIL alongside SCimilarity and a dual-threshold strategy to 79 synovial samples from rheumatoid arthritis (RA) patients. The pipeline converged on OGT (O-GlcNAc transferase) as a top candidate.

Mechanistic and Experimental Follow-up

Computational predictions were validated experimentally: patient tissues showed OGT dysregulation; cellular assays and fibroblast-like synoviocyte-specific OGT knockout mice supported its pathogenic role. Mechanistically, OGT-mediated O-GlcNAcylation stabilizes SAP130 , leading to epigenetic suppression of BTG2 and driving aberrant synovial cell behavior.

Therapeutic Targeting in Preclinical Models

The team designed an AS1411-OSMI targeted degradation molecule. In collagen-induced arthritis (CIA) mice, treatment reduced disease severity, and combination with etanercept (ETA) yielded additional benefit. Bone structure and histology improvements were also documented. These results remain preclinical (mouse models only).

Conclusions and Caveats

scTransMIL provides a layered investigative path: from sample diagnosis to cellular localization to gene prioritization. The RA case demonstrates how computational leads can be connected to tissue, cellular, and animal experiments. The authors emphasize that high cell scores do not equal definitive malignant identity, and candidate genes are not confirmed targets. The framework's value lies in giving researchers more evidence to decide which experimental leads to pursue next amid growing single-cell data volumes.