Single Cell Data Analysis Algorithms

National Institute Of Diabetes And Digestive And Kidney Diseases

Investigators

Abstract

The increasing volume and complexity of single-cell data provide unprecedented opportunities for dissecting cellular differentiation, but also demand new computational strategies that can extract both predictive accuracy and mechanistic understanding. Black-box deep learning models excel at reconstructing complex patterns yet offer limited biological interpretability, while purely mechanistic models are often intractable to fit from sparse and noisy measurements. To bridge these approaches, we developed a hybrid framework that integrates deep generative modeling with explicit regulatory modeling. We applied a variational autoencoder to the Drosophila melanogaster blastoderm gene expression atlas to learn a compact latent representation, and trained a neural ODE in this latent space to capture smooth temporal dynamics across developmental time points. By decoding trajectories back to the full gene space, we obtained consistent derivative estimates, which enabled the fitting of a mechanistic Hill-function model of gene regulation with interpretable coupling parameters. This two-stage strategy allows us to reconstruct cellular trajectories, impute missing data, and perform targeted “what-if” interventions by directly editing regulatory couplings in the Hill model. The approach therefore preserves the expressive power of deep learning while yielding a mechanistic framework that supports systematic predictions under perturbations, providing a path toward interpretable modeling of developmental gene regulatory networks. A common step in scRNA-seq analysis is the selection of marker genes—a small subset of genes whose expression profiles can distinguish between different sub-populations of cells. Identifying a minimally sufficient set of these genes is a non-trivial task. Current methods, from statistical tests to deep learning models, show poor concordance, with less than 50% overlap in the top 20 marker genes they identify. This highlights the need for more robust and explanatory analyses. Deep learning models are particularly well-suited for this predictive task, as their architecture is inherently capable of capturing the complex, non-linear relationships between genes that may distinguish a cell type. While these models offer impressive predictive performance, their "black-box" nature provides scant biological understanding on its own. To extract mechanistic relevance from these trained models, we have developed SensX, a model-agnostic explainable AI (XAI) framework that uses global sensitivity analysis (GSA) to explain what a model has learned. The framework operates by analyzing simultaneous perturbations in all genes to determine important high-order interactions. We also defined SensX landscapes, which provide critical information about the scale and magnitude of gene expressions relevant to a specific cell type. We applied SensX to explain a model trained on the Human Lung Cell Atlas (HLCA), which contains data from over a million cells. The analysis revealed that only 0.5% of genes are important to accurately classify 21 different cell types. Retraining models with these smaller gene subsets showed that fewer than 10 genes are sufficient to distinguish some cell types. This analysis also revealed that more complex models learn slightly different sets of marker genes, potentially pointing to niche sets of genes critical for distinguishing cell types with higher accuracy. A manuscript is in preparation. As a model-agnostic XAI framework, SensX is designed to be both scalable and efficient. Benchmarks on synthetic data sets show that SensX outperforms current state-of-the-art XAI methods in accuracy (up to 50% higher) and computation time (up to 158 times faster), with higher consistency in all cases. Moreover, SensX is the only model-agnostic method that scaled to explain vision transformer (ViT) models with more than 150,000 input features. In that context, SensX not only validated the models by confirming that the features they learned were intuitively accurate but also revealed potential biases inherent to the model architecture. A manuscript is in preparation.