On the Promises and Limits of Multi-omics Integration for Deconvolution: The HADACA3 Benchmark
Understanding the cellular composition of complex tissues, such as tumors, is a key challenge in biology and medicine. A common approach, known as deconvolution, aims to estimate Understanding the cellular composition of complex tissues, such as tumours, is a key challenge in biology and medicine. Deconvolution methods estimate cell-type proportions from bulk molecular measurements, and the increasing availability of multiple molecular modalities has raised the expectation that their integration should improve estimation accuracy. However, whether and under which conditions multi-omics integration actually improves deconvolution remains unclear. Here we present HADACA3, a community-driven competition followed by a modular benchmark of more than 18,000 pipelines combining pre-processing, feature selection, deconvolution and integration of RNA and DNA methylation (DNAm), across nine pancreatic cancer datasets. With the available references, DNAm alone achieves the highest median performance, outperforming RNA alone and most integration strategies. Dataset-specific optimisation of generic integration pipelines rarely improves upon DNAm alone: the best configurations mostly reduce the weight of RNA, and their gains do not generalise across datasets. The only consistent improvement, confirmed on unseen data, comes from an integration rule that estimates cancer cell types from DNAm alone and averages both modalities for non-cancer cell types. These findings challenge the assumption that integrating more molecular data necessarily improves deconvolution. Instead, effective multi-omics integration relies on prior knowledge of the cell-type-specific strengths of each modality, rather than on generic statistical integration strategies.