scOPE

A somatic mutation is legible in single-cell RNA-seq only when its gene is expressed, the mutant allele is transcribed, and reads happen to span the variant position. Miss any of those and a cell simply has no variant read — which may mean wild type, or may mean unobserved. The mutations that define a tumor's clonal architecture are therefore among the features scRNA-seq is least equipped to report. Bulk tumor cohorts are the mirror image: matched genotype and expression for hundreds of patients, averaged over every cell.

scOPE (single-cell Oncological Prediction Explorer) treats the gap between them as a domain transfer problem rather than a classification problem. For each cancer type it factorizes a bulk expression matrix into a compact latent space, trains one regularized classifier per driver against matched genotype, and then freezes the gene set, preprocessing statistics, latent loadings, and classifier weights before any single-cell data is seen. Gene-matched single cells are projected through that unchanged map to yield a continuous score per cell per driver program. Because supervision comes from bulk genotype labels rather than pretraining on unlabelled atlases, it sits apart from single-cell models such as Geneformer and scGPT: what transfers is a specific genotype–phenotype association, not a general representation of cell state.

The central result, from Oregon Health & Science University with Amherst College, is deliberately a negative one: transfer is selective, not general. Across 158 audited driver–cancer models spanning seven malignancies, 102 met predefined claim-safety criteria and only 11 reached out-of-fold AUROC ≥ 0.90. Deciding which programs survive the move, and abstaining when none do, is the framework's main job.

#Key Features

• Frozen transfer map: No single-cell mutation labels, patient labels, cell-type annotations, or embedding coordinates are used to refit components or update classifier parameters — the bulk-fitted map is applied to the target cohort unchanged.
• Programs, not allele calls: A score ranks a cell by similarity to a bulk-derived, driver-associated transcriptional program. Bulk mutation labels can also encode lineage, tumor purity, or co-occurring alterations, so the score is not a posterior probability of genotype.
• Cell-state residual scores: Raw scores are elevated in healthy cells whenever a bulk program tracks lineage. Subtracting the matched healthy-reference median for a cell's annotated state gives the residual used for every claim-bearing comparison.
• Ground-truth-free confidence: A weighted geometric mean of bulk transferability, spatial coherence in the expression-derived neighborhood graph, score concentration, and agreement with an inferred copy-number axis ranks programs where genotype will never exist.

#Technical Details

Each cancer-specific bulk matrix is standardized and factorized by truncated SVD at k = 30 components; one L2-regularized logistic classifier (C = 1.0, balanced class weights, liblinear) is fit per driver in that latent space. Training cohorts are BeatAML for AML and TCGA for the six solid tumors, with source matrices of up to 651 AML, 903 BRCA, 408 CRC, 155 GBM, 567 LUAD, 172 PAAD, and 466 SKCM tumors. Bulk discrimination uses five shuffled folds with the entire pipeline refit inside each training fold. Median out-of-fold AUROC ranges from 0.788 in CRC down to 0.597 in SKCM; the strongest individual models are AML NPM1 (0.971), GBM IDH1 (0.963), PAAD KRAS (0.944), and BRCA TP53 (0.931).

Transfer targets are seven public scRNA-seq cohorts totalling 190,817 retained cells, disjoint from the bulk training data. A label-free moment-matching step rescales the gene-wise location and scale of the target cohort to the bulk reference; the loadings and classifiers themselves never change. Against expressed-mutation labels in the van Galen AML cohort, residual scores separated NPM1-labelled cells at AUROC 0.753, with concordant enrichment for TP53 and DNMT3A; the confidence score ranked the three independently supported AML programs above the rest at AUROC 0.85 across 12 truth-evaluable drivers.

#Applications

The practical target is the archive of scRNA-seq cohorts that were never genotyped and never will be. There, scOPE supplies a per-cell reading of driver-associated transcriptional activity that can nominate cells for targeted variant recovery, flag samples worth matched DNA profiling, track a program across treatment timepoints — the NPM1 program contracted during therapy across multiple patients — and complement expression-derived copy-number inference, where it resolved discordant malignant populations invisible to aneuploidy alone. The scope-bio package implements both phases over AnnData objects.

#Impact

scOPE's contribution is as much methodological as predictive: it separates model performance from biological interpretation with an explicit evidence ladder, and it reports that even well-supported programs occupy restricted transcriptional subspaces rather than uniformly marking mutation-positive tumors. Two limits shape its use today. The architecture is classical machine learning — a linear factorization plus per-driver logistic regression — which buys interpretable gene loadings but caps the structure it can capture. And no fitted checkpoint is distributed: the MIT-licensed package ships the framework, so a user must refit the bulk phase on a reference cohort such as TCGA or BeatAML before projecting their own cells. The work remains a preprint awaiting peer review.