Loeb, R. | bioRxiv Science

bioRxiv2026

Comprehensive benchmarking of batch integration methods for spatial transcriptomics using a large-scale cancer atlas

Spatial transcriptomics (ST) enables spatially-resolved gene expression measurement, providing insights into tissue architecture and disease biology. However, batch effects from sequencing protocols, sample processing, and other technical factors can confound biological signals. Although batch correction has been extensively studied in single-cell transcriptomics, spatial integration methods lack rigorous benchmarking on large real-world datasets. This study benchmarks 11 representation-learning methods across three categories--linear, graph-based and probabilistic methods using Owkins MOSAIC Window dataset, a large-scale spatial transcriptomics atlas of human cancers. Methods are evaluated across three criteria: batch correction, biological conservation, and spatial conservation. We also propose a new integration metric to assess robustness of representations to domain shifts and generalizability to unseen samples. Probabilistic methods (scVIVA, scVI) outperform linear and graph-based approaches in batch correction and biological conservation. On the other hand Graph-based methods excelled at spatial conservation but underperformed in batch integration. Out-of-distribution evaluation reveals that sophisticated methods show reduced peformance on unseen samples while linear methods maintain robust generalization, highlighting trade-offs between integration quality and generalizability that should guide method selection for real-world applications.

cancer biology↗

bioRxiv2024

Robust Evaluation of Deep Learning-based Representation Methods for Survival and Gene Essentiality Prediction on Bulk RNA-seq Data

Deep learning (DL) has shown potential to provide powerful representations of bulk RNA-seq data in cancer research. However, there is no consensus regarding the impact of design choices of DL approaches on the performance of the learned representation, including the model architecture, the training methodology and the various hyperparameters. To address this problem, we evaluate the performance of various design choices of DL representation learning methods using TCGA and DepMap pan-cancer datasets, and assess their predictive power for survival and gene essentiality predictions. We demonstrate that non DL-based baseline methods achieve comparable or superior performance compared to more complex models on survival predictions tasks. DL representation methods, however, are the most efficient to predict the gene essentiality of cell lines. We show that auto-encoders (AE) are consistently improved by techniques such as masking and multi-head training. Our results suggest that the impact of DL representations and of pre-training are highly task- and architecture-dependent, highlighting the need for adopting rigorous evaluation guidelines. These guidelines for robust evaluation are implemented in a pipeline made available to the research community.

bioinformatics↗