Abstract:The increasing availability of large-scale single-cell datasets has enabled the detailed description of cell states across multiple biological conditions and perturbations. In parallel, recent advances in unsupervised machine learning, particularly in transfer learning, have enabled fast and scalable mapping of these new single-cell datasets onto reference atlases. The resulting large-scale machine learning models however often have millions of parameters, rendering interpretation of the newly mapped datasets challenging. Here, we propose expiMap, a deep learning model that enables interpretable reference mapping using biologically understandable entities, such as curated sets of genes and gene programs. The key concept is the substitution of the uninterpretable nodes in an autoencoder’s bottleneck by labeled nodes mapping to interpretable lists of genes, such as gene ontologies, biological pathways, or curated gene sets, for which activities are learned as constraints during reconstruction. This is enabled by the incorporation of predefined gene programs into the reference model, and at the same time allowing the model to learn de novo new programs and refine existing programs during reference mapping. We show that the model retains similar integration performance as existing methods while providing a biologically interpretable framework for understanding cellular behavior. We demonstrate the capabilities of expiMap by applying it to 15 datasets encompassing five different tissues and species. The interpretable nature of the mapping revealed unreported associations between interferon signaling via the RIG-I/MDA5 and GPCRs pathways, with differential behavior in CD8+ T cells and CD14+ monocytes in severe COVID-19, as well as the role of annexins in the cellular communications between lymphoid and myeloid compartments for explaining patient response to the applied drugs. Finally, expiMap enabled the direct comparison of a diverse set of pancreatic beta cells from multiple studies where we observed a strong, previously unreported correlation between the unfolded protein response and asparagine N-linked glycosylation. Altogether, expiMap enables the interpretable mapping of single cell transcriptome data sets across cohorts, disease states and other perturbations. ### Competing Interest Statement Fabian J. Theis consults for Immunai Inc., Singularity Bio B.V., CytoReason Ltd, and Omniscope Ltd, and has ownership interest in Dermagnostix GmbH and Cellarity.

Enhanced Integrated Gradients: improving interpretability of deep learning models using splicing codes as a case study

DeepSS: Exploring Splice Site Motif Through Convolutional Neural Network Directly from DNA Sequence

Biologically Informed Deep Learning to Infer Gene Program Activity in Single Cells

CADD-Splice—improving genome-wide variant effect prediction using deep learning-derived splice scores

DeepIMAGER: Deeply Analyzing Gene Regulatory Networks from scRNA-seq Data

Deep learning of the splicing (epi)genetic code reveals a novel candidate mechanism linking histone modifications to ESC fate decision

DeepIII: Predicting Isoform-Isoform Interactions by Deep Neural Networks and Data Fusion

A mechanistically interpretable neural network for regulatory genomics

DeepSplice: a deep learning approach for accurate prediction of alternative splicing events in the human genome

Identifying DNA Sequence Motifs Using Deep Learning

Reference-informed prediction of alternative splicing and splicing-altering mutations from sequences

Learning interpretable cellular embedding for inferring biological mechanisms underlying single-cell transcriptomics

Splice site recognition - deciphering Exon-Intron transitions for genetic insights using Enhanced integrated Block-Level gated LSTM model

Data Integration of Hybrid Microarray and Single Cell Expression Data to Enhance Gene Network Inference

DeepEnhancer: Predicting Enhancers by Convolutional Neural Networks.

A Deep Learning-Based Sequence Analyzer Incorporating the Transcription Factor Binding Affinity to Dissect the Effects of Non-Coding Genetic Variants

Comprehensive single-cell RNA-seq analysis using deep interpretable generative modeling guided by biological hierarchy knowledge

High-resolution transcription factor binding sites prediction improved performance and interpretability by deep learning method

Interpreting cis -regulatory mechanisms from genomic deep neural networks using surrogate models

Deep learning approaches for non-coding genetic variant effect prediction: current progress and future prospects