Anastasia Litinetskaya

Abstract Constructing joint representations from multimodal single-cell datasets is crucial for understanding cellular heterogeneity and function. Traditional methods, such as factor analysis and kNN-based approaches, face computational limitations with scalability across large datasets and multiple modalities. In this work, we demonstrate the product-of-experts VAE-based model, which offers a flexible, scalable solution for integrating multimodal data, allowing for the seamless mapping of both unimodal and multimodal queries onto a reference atlas. We evaluate how different strategies for combining modalities in the VAE framework impact query-to-reference mapping across diverse datasets, including CITE-seq and spatial metabolomics. Our benchmarks assess batch effect correction, biological signal preservation, and imputation of missing modalities. We showcase our approach in a mosaic setting, integrating CITE-seq and multiome data to accurately map unimodal and multimodal queries into the joint latent space. We extend this to spatial data by integrating gene expression and metabolomics from paired Visium and MALDI-MSI slides, achieving a high correlation in metabolite predictions from spatial gene expression. Our results demonstrate that this VAE-based framework is scalable, robust, and easily applicable across multiple modalities, providing a powerful tool for data imputation, querying, and biological discovery.

Abstract To deliver clinically relevant insights from large patient cohorts profiled with single-cell technologies, a key challenge is to relate sample-level and single-cell measurements. We present MultiMIL, a deep learning framework that applies attention-based multiple-instance learning for phenotype prediction and cell state identification. We applied MultiMIL to peripheral blood mononuclear cells from COVID-19 patients, the Human Lung Cell Atlas, and a spatial proteomics breast cancer dataset, demonstrating how our model can be utilized to find phenotype-associated cell states, learn phenotype-informed sample representations, and expand disease signatures.