PaSCient

Single-cell foundation models have largely been built to represent individual cells: given an expression profile, return an embedding that captures cell type and state. But disease is rarely a property of one cell. It is a property of a tissue ecosystem — many interacting cell types shifting together — and the unit that clinicians and drug developers actually care about is the patient, not the cell. PaSCient closes that gap by learning a single vector representation of a whole patient sample directly from the scRNA-seq profiles of all the cells it contains.

The model was developed at Genentech Research & Early Development, with collaborators at Yale, Stanford, and Roche Informatics. It was first posted as a preprint in November 2024 and published in Cell Systems in 2026. Rather than averaging cells into a pseudo-bulk profile, which discards cellular composition and is easily dominated by batch effects, PaSCient uses a learned attention mechanism to decide which cells matter for a given patient's disease state, and keeps those weights available afterwards as interpretable importance scores.

PaSCient is trained once, as a disease model, across a large aggregated atlas and then applied to new samples without refitting. Its outputs support dimensionality reduction and patient-cohort visualization, prioritization of the cell types and genes driving a diagnosis, patient subgroup discovery, and prediction of disease severity or treatment response.

#Key Features

• Patient-level embeddings from single-cell input: Consumes the full cell-by-gene matrix for a sample and emits one vector per patient, preserving multi-cellular context that pseudo-bulk aggregation destroys.
• Learned attention pooling: A softmax-attention layer scores every cell in the sample and weights it accordingly, which ablations show outperforms mean pooling, gated attention, and a transformer aggregator.
• Fine-grained interpretability: Integrated gradients produce attributions for every cell-gene pair, which can be averaged to rank individual cells, cell types, genes, or genes within a specific cell type.
• Imbalance-aware training: A resampling strategy that oversamples rare disease and tissue classes measurably improves performance on an atlas where COVID-19 accounts for roughly 9% of samples and multiple sclerosis about 2%.
• Implicit clinical signal: On held-out COVID-19 cohorts never seen in training, the embedding space organizes by disease severity rather than by study, and the predicted diagnosis probability tracks severity.

#Technical Details

PaSCient's architecture is inspired by DeepSets: a learnable cell encoder maps each cell's expression vector into a shared embedding space, a softmax-attention aggregator collapses that set of cell embeddings into a single patient embedding, and a multilayer perceptron classifier predicts the disease label. The whole stack is trained end-to-end by minimizing cross-entropy against observed disease-state labels. Training used 24.3 million droplet-based scRNA-seq profiles from more than 5,000 patient samples covering 135 disease-state labels across 413 studies and 189 tissues, all sourced from the publicly accessible CELLxGENE corpus and split 60/20/20 with all samples from a study held in the same partition. Evaluated by weighted F1 on disease classification, PaSCient outperforms pseudo-bulk and cell-type-proportion baselines, embeddings from the single-cell foundation models CellPLM and SCimilarity paired with either kNN or MLP classifiers, and the domain-specific ScRAT method. Swapping mean pooling for non-linear attention alone improved weighted F1 by 16.6%.

#Applications

PaSCient is aimed at translational researchers who have single-cell data from patient cohorts and want to reason at the level of the individual rather than the cell. Concrete uses include classifying disease state from a tissue biopsy or blood draw, ranking the cell types and cell-type-specific genes that drive a diagnosis to generate mechanistic hypotheses, stratifying a cohort into subgroups for trial design, and reading out severity or treatment response from an embedding that was never explicitly trained on those labels. In the paper's COVID-19 case study, the model independently recovered monocytes and platelets as the most predictive cell types and interferon-response genes including IFITM3 and IFI27 as the top monocyte features.

#Impact

PaSCient reframes what a single-cell foundation model is for, moving the unit of representation from the cell to the patient and showing that the shift is worth real accuracy on disease classification. It also delivers a useful negative result for the field: a plain pseudo-bulk baseline beat several general-purpose single-cell foundation models on this task, a reminder that cell-level pretraining does not automatically transfer to patient-level questions. The code is released under Apache-2.0 on GitHub, archived on Zenodo, and ships an inference notebook demonstrating how to apply a fixed checkpoint to new data. The main barrier to adoption is that the trained weights are distributed on request by email rather than as a self-serve download, so reproducing the model's outputs requires contacting the authors.