expiMap
Summary¶
Lotfollahi, Rybakov, Hrovatin, et al. (2023) present expiMap, a biologically informed deep learning architecture for single-cell reference mapping — integrating newly generated "query" datasets into large reference atlases — that constrains the model's latent space to correspond to interpretable "gene programs" (curated gene sets) rather than an arbitrary learned embedding. Built via transfer learning on top of scArches, expiMap can both refine existing gene programs and learn de novo ones from query data, and the authors demonstrate its use resolving cellular and pathway-level responses to interferon-β stimulation and to COVID-19 disease/treatment across patients and cell types.
Method¶
- Reference training: a variational-autoencoder-style model is trained on reference atlas data with gene expression and study labels, encoding cells into latent variables that each correspond to a known gene program (GP); group lasso and gene-level sparsity regularization prune and refine which genes contribute to each program, feeding into a linear decoder that keeps the mapping between latent dimensions and genes interpretable.
- Query mapping via architectural surgery: following the scArches transfer-learning strategy, most reference-model weights are frozen and new query data is mapped by fine-tuning conditional weights, while new latent-space nodes are added to capture variation not explained by existing programs — either constrained to a predefined but previously inactive gene set ("new constrained") or fully data-driven with L1 regularization ("new unconstrained").
- Independence enforcement: the Hilbert-Schmidt Independence Criterion (HSIC) is used during query optimization to encourage statistical independence between newly learned unconstrained programs and the existing reference programs, reducing redundant or entangled representations.
Performance¶
- Reference mapping benchmark: compared against scArches+scVI, Seurat v4, and Symphony for mapping an interferon-β-stimulated PBMC dataset onto a healthy immune reference, expiMap achieved better batch-correction and bio-conservation scores overall; notably, scArches+scVI failed to correctly mix untreated query monocytes with healthy reference monocytes (average batch correction score 0.47) while expiMap succeeded (0.68), attributed to expiMap's explicit encoding of interferon-β-related gene programs, which helps it distinguish genuinely perturbed cells from unperturbed ones that should integrate normally.
- Ablation: removing the top five gene programs identified as most differential between interferon-β-treated and control cells and retraining caused incorrect mixing of treated and control cells in the reference — direct evidence that the model's interpretable programs are functionally, not just descriptively, responsible for its integration accuracy.
Application: COVID-19 Treatment Response¶
- Mapped PBMC data from two COVID-19 patients (across disease and remission timepoints, both treated with the IL-6-receptor-targeting drug tocilizumab) onto a healthy immune reference plus the interferon-β dataset, identifying 20 annotated cell states including rare populations (megakaryocytes, erythroid progenitors, CD10⁺ B cells).
- CD8⁺ T cells and monocytes from severe/remission timepoints did not integrate cleanly into the healthy reference; differential gene-program testing between these query cells and reference controls implicated RIG-I/MDA5 pattern-recognition-receptor and GPCR-pathway programs as differentially active across disease stages in these cell types.
- Combining expiMap's integrated embedding with CellChat cell-cell-communication inference revealed patient- and timepoint-specific annexin-pathway signaling circuits (implicating CD14⁺/CD16⁺ monocytes, NK cells, and CD8⁺/CD4⁺ T cells) that differed substantially between the two patients despite both ultimately recovering — illustrating the model's use for generating patient-specific, potentially druggable mechanistic hypotheses beyond simple cell-type annotation.
See Also¶
- Biologically Informed Neural Networks — the shared design philosophy across expiMap, P-NET, and BINN.
- scVI — the underlying variational-inference architecture used by the scArches reference-mapping baseline expiMap is compared against.