Zweig, A. | bioRxiv Science

bioRxiv2025

Robust Integration of Sparse Single-Cell Alternative Splicing and Gene Expression Data with SpliceVI

Alternative splicing (AS) and gene expression (GE) are tightly related regulatory processes, critical for defining cell types and states, yet are rarely modeled together in single-cell analyses. This hinders a comprehensive understanding of cellular identity. We address this by introducing SpliceVI, adapted from MultiVI (Multi-modal Variational Inference) to specifically handle AS. Applied to a large multisample mouse Smart-seq2 dataset (n = 142, 315 cells/nuclei), SpliceVI jointly learns from both AS and GE using a partial variational autoencoder that effectively handles the sparsity and missingness of splicing data. We show that SpliceVIs joint embeddings are more expressive and informative of biological correlates like age than a GE-only approach (scVI). SpliceVI also uncovers splicingbased differences between neuronal subclusters. This approach reveals the distinct yet synergistic relationship between AS and GE in shaping cellular diversity in mouse.

systems biology↗

bioRxiv2025

Domain-Invariant Feature Learning for Patient-Level Phenotype Prediction from Single-Cell Data

Accurate prediction of patient-level disease status from single-cell RNA sequencing (scRNA-seq) data is critical to enabling precision diagnostics. However, study-specific artifacts induce spurious correlations that limit generalization and interpretability. We studied this problem in the context of Multiple Instance Learning (MIL), a framework where each patient is modeled as a set of single-cell profiles. To improve robustness to domain shifts, we propose an adversarial and metric-based approach that learns domain-invariant representations while preserving task-relevant biological variation. We benchmarked our method on a systemic lupus erythematosus (SLE) dataset with synthetically added spurious features and evaluated its performance on two real-world scRNA-seq atlases: a cross-tissue immune dataset and a COVID-19 severity atlas. Across all settings, we observed consistent improvements in out-of-domain accuracy and more biologically faithful model attributions. Our findings establish a new standard for robust, interpretable patient-level prediction under domain shifts using scRNA-seq.

bioinformatics↗