Subject-level disease prediction from scRNA-seq, pairing cell-type-grouped scGPT pretraining with a Reactome pathway-constrained decoder.
No providers recorded yet. Browse all providers
Single-cell RNA sequencing foundation models are usually evaluated on cell-level tasks — annotation, batch integration, perturbation response — but the clinical question is often at the level of the patient: does this subject have the disease, and which biological processes drove that call. Bridging cell-level representations to subject-level labels normally means averaging embeddings across a sample, which discards the cellular heterogeneity carrying the signal and leaves the prediction as opaque as the transformer that produced it.
eSPred, developed at Duke University, is a framework that adapts an existing single-cell foundation model to that setting through three coordinated changes. During pretraining, cells of the same type are grouped four at a time and their expression profiles concatenated into a single input, so the masked-expression objective learns cell-type-specific relationships alongside gene-gene ones. During fine-tuning, a pathway-integrated decoder replaces the usual [CLS] token pooling with a hierarchical network whose connectivity is restricted to known gene-pathway edges from Reactome. At prediction time, a hierarchical classifier labels each cell and aggregates cell-level calls into a subject-level outcome by majority vote rather than pooling embeddings first.
The authors build the framework on scGPT, chosen for its pretraining scale of over 10 million cells from 65 CELLxGENE datasets, but present the three components as modular: the same grouping, pathway decoder, and hierarchical classifier can be layered onto other transformer-based single-cell models such as scBERT with adjusted preprocessing.
Preprocessing follows a standard Scanpy pipeline — filtering genes expressed in fewer than 10 cells, normalizing to 10,000 counts per cell, log1p transformation, and selecting 1,200 highly variable genes to match scGPT's original pretraining configuration. Cells are grouped by annotation where available and by Leiden clustering on PCA-reduced data otherwise; leftover cells that cannot form a complete group of four are dropped. The grouped input produces a 512 x 4,800 embedding matrix, split into four 512 x 1,200 per-cell matrices, each passed through the pathway decoder to yield a 512-dimensional cell embedding.
Benchmarks cover three cohorts: COVID-19 (80 subjects, 422,220 cells, 29 cell types after grouping), lupus (274 subjects, 1.26 million cells, 11 immune cell types), and lung disease (318 subjects spanning LUAD, LUSC, NSCLC, and COPD, 1.28 million cells, 33 cell types), each with a 70/30 subject-level split. eSPred reaches F1 scores of 0.877, 0.910, and 0.960 respectively, against a base scGPT configuration at 0.772, 0.637, and 0.784. An ablation separates the contributions: cell-type concatenation alone lifts F1 to 0.824/0.824/0.836, the pathway decoder alone lifts lupus to 0.849, and the hierarchical classifier improves every configuration it is added to. Baselines include ProtoCell4P, DeepGeneX, logistic regression, random forest, SVM, and an MLP, all given the same 1,200 genes as input.
The target use is subject-level clinical prediction from a single-cell profile — classifying disease versus control, and by extension disease subtype, from a patient's dissociated sample. Because the decoder is pathway-structured, each prediction comes with a ranked list of implicated processes: in the COVID-19 cohort the top pathways were signal transduction, immune system, and protein metabolism, consistent with established SARS-CoV-2 biology, and in lupus the model highlighted GPCR signaling. That combination suits translational groups who need a defensible mechanistic account alongside a classifier output, and the modular design lets a lab swap in a newer single-cell backbone without redesigning the downstream stack.
eSPred is a preprint and has not been peer reviewed, and no code or pretrained checkpoint has been released. Its contribution is methodological: it demonstrates that how cells are batched into a foundation model's context and how gene embeddings are routed during fine-tuning both materially affect subject-level accuracy, with the largest gains on the most class-imbalanced cohort. The authors are explicit about the limits of pathway-constrained architectures — predefined Reactome annotations may not capture context-specific regulation, and the importance rankings carry no significance testing, so they indicate predictive contribution rather than statistical enrichment. Evaluation rests on a single 70/30 split per dataset across three cohorts, so generalization to new tissues and assay platforms remains to be established.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.