On the Promises and Limits of Multi-omics Integration for Deconvolution: The HADACA3 Benchmark

arXiv:2606.05980 · q-bio.QM · Submitted 2026-06-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "On the Promises and Limits of Multi-omics Integration for Deconvolution".

Marcus: Here's the gist of the paper. The paper introduces HADACA3, a community-driven benchmark designed to systematically evaluate multi-omics integration strategies for deconvolution—the process of estimating cellular composition from bulk molecular measurements in complex tissues like tumors.

Ines: This challenge is significant in biology and medicine, especially with the growing availability of multiple molecular data types.

Marcus: The authors present HADACA3 as a comprehensive dataset and benchmarking framework, established through a four-day collaborative competition followed by a large-scale computational test involving over 250,000 analysis pipelines across nine datasets featuring matched DNA methylation (DNAm) and RNA profiles under diverse biological and experimental conditions.

Ines: The benchmark's primary goal is to rigorously test the assumption that combining multi-omic data improves deconvolution performance.

Paper summary: Ines: So this paper introduces HADACA3, which is a community-driven benchmark for deconvolution—that's the process of estimating what cells are in a tissue from bulk measurements like RNA and DNA methylation.

Marcus: It's significant because we often assume that mixing different molecular data types, like RNA and DNAm, will just make the estimation better.

Ines: That's exactly the assumption they set out to test with this benchmark. The paper claims that combining multi-omics data should improve deconvolution performance in complex tissues like tumors.

Yuki: But there's a big question here about whether that combination actually works across different types of biological tissues and noise structures, right?

Ines: Right, and HADACA3 sets up a test where they check preprocessing, feature selection, modeling, and various integration strategies to see what drives the performance.

Marcus: The main finding they report is that DNA methylation alone achieves the highest median performance across all nine datasets tested.

Yuki: So even when we have these multiple data types available, using just DNAm seems to be the most stable and reliable single approach for deconvolution, according to this study.

Ines: But it doesn't mean multi-omics integration is useless; they found that multi-omics strategies can achieve superior top performance in certain specific datasets and pipeline configurations.

Marcus: They point out a real trade-off here between getting the most robust average performance and aiming for the absolute best result.

Ines: Specifically, late integration based on error-weighted averaging shows a strong baseline for combining modalities, while non-linear early integration methods like optimal transport show promise on real biological data because they can capture those complex cross-modal relationships.

Yuki: Early integration methods are sensitive to scale differences and noise heterogeneity across the different modalities, which normalization only partially fixes.

Marcus: That's a fair point; early integration struggles with those kinds of inconsistencies that late integration seems to avoid by separating the inference from the aggregation step.

Ines: The benchmark focuses on pancreatic cancer datasets with cell-type compositions known from in silico, in vitro, and in vivo settings, and they used features from both RNA and DNAm across all nine datasets.

Yuki: They specifically focused on reference profiles for five major cell types found in pancreatic tumors: immune cells, fibroblasts, endothelial cells, and two cancer-cell subtypes called classical and basal-like.

Marcus: The evaluation used twelve complementary metrics to score the performance, including cohort-wise correlations and error metrics like RMSE and Aitchison distance.

Ines: The paper notes that cell-type correlation and Aitchison distance are the most discriminative elements of that aggregate score they calculated.

Yuki: And when they did a linear decomposition of those scores, it seemed like the deconvolution algorithm actually drives performance in late integration, while the integration method dominates early integration because latent embedding approaches perform inconsistently.

Marcus: So, if you're looking at this for large cohorts or complex data where batch effects are an issue, the paper suggests late integration might be more robust than trying to force an early combination across different modalities.

Ines: The paper does flag some limitations; they only focused on pancreatic cancer datasets, and they also noted that RNA-only deconvolution under these conditions had lower quality reference profiles.

Yuki: And there's also the statistical challenge mentioned—with so many pipeline combinations, small score differences might appear consistent without actually being statistically meaningful based on the median aggregate scores.

Marcus: The computational cost for running this full scope of tests was quite high, estimated at over five hundred seventy-five thousand core-hours.

Ines: For someone who only listens to this show, the main point is that while DNA methylation alone is a very reliable single method for deconvolution right now, multi-omics integration isn't automatically better on average across all tests.

Yuki: It changes things by showing us that we need to be careful about which integration strategy we choose based on the specific biology and data quality we are dealing with.

Marcus: This work from HADACA3 gives us a clearer picture of where the current state of multi-omics deconvolution lies, even if it points toward using different methods depending on the situation.

Conclusion: Ines: So, we’re wrapping up our look at HADACA3, which is this new benchmark for testing how well we can mix different molecular data types—like RNA and DNA methylation—to figure out what cells are in a tissue.

Marcus: Yeah, the paper calls it "On the Promises and Limits of Multi-omics Integration for Deconvolution: The HADACA3 Benchmark." It sets up this massive test where they run over two hundred fifty thousand different analysis pipelines across nine different datasets.

Yuki: What this means for us is that we’re finally getting a standardized way to see if combining these data types actually helps or not when we are trying to deconvolve complex samples like tumors.

Ines: The authors found that DNA methylation alone was the most stable single method across all those nine different biological settings.

Marcus: But they also showed that you can get better performance by combining them, depending on how you set up the analysis pipeline and which specific dataset you are looking at.

Yuki: So it seems like the idea isn't that multi-omics is always better, but rather it’s about finding the right strategy for specific problems.

Ines: Right, because they found this trade-off between getting a solid average result and chasing that absolute peak performance in certain setups.

Marcus: Specifically, they found that late integration methods, like error-weighted averaging, give you a very strong baseline when you’re dealing with different data quality.

Yuki: And the early integration stuff—like optimal transport—that seems promising for capturing complex relationships on real biological data, even though it gets sensitive to noise differences between the modalities.

Ines: That sensitivity is the caveat they pointed out; normalization only helps so much when you’re doing early integration.

Marcus: The paper does note some limits, too. They focused only on pancreatic cancer datasets and noted that RNA-only deconvolution struggled because the reference profiles were lower quality than what they needed.

Yuki: And there’s also a statistical hurdle with all those pipeline combinations; sometimes small score differences just look consistent but don't mean anything statistically significant when you look at the median scores.

Ines: So, the authors are suggesting that while single-modality approaches are still strong baselines, we need to be very careful about which integration method we pick based on the specific biology of what we’re studying.

Hugo Barbot, Elise Amblard, Nicolas Homberg, Lucie Lamothe, Morgane Térézol, Mira Ayadi, Aurélia Baurès, Yasmina Kermezli, Carl Herrmann, Sebastien Dejean, Lionel Spinelli

IRMAR CNRS Institute Agro Rennes Angers University Grenoble Alpes University of Grenoble Alpes CNRS LIG Aix Marseille University INSERM MMG BioQuant IPMB Heidelberg University IMT Aix Marseille University CNRS Univ Toulouse Univ Grenoble Alpes INSERM IAB IGDR CNRS Univ Rennes

q-bio.QM

Submitted: 2026-06-04

Updated: 2026-11-12

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The paper introduces HADACA3, a community-driven benchmark designed to systematically evaluate multi-omics integration strategies for deconvolution—the process of estimating cellular composition…

Key concepts

Deconvolution
This is the process of estimating what different cell types are present in a bulk sample of tissue (like a tumor) based on molecular measurements. It's challenging because the raw data doesn't directly reveal cell identities.
Multi-omics Integration
This refers to combining different types of molecular data, such as DNA methylation and RNA profiles, to get a more complete picture than using just one type of data. The paper tests if this combination actually makes the estimation process better.
Late Integration
This strategy involves performing inference on each data type separately first, and then combining the results later using an aggregation method. This approach is noted for being more robust against noise and differences between the various molecular measurements.
Early Integration
This strategy combines different data types early in the analysis pipeline, often by creating a joint representation before making final inferences. While promising for capturing complex relationships, it can be sensitive to noise and scale differences across the input modalities.

Terminology

Summary

The paper introduces HADACA3, a community-driven benchmark designed to systematically evaluate multi-omics integration strategies for deconvolution—the process of estimating cellular composition from bulk molecular measurements in complex tissues like tumors. p. 0 This challenge is significant in biology and medicine, especially with the growing availability of multiple molecular data types. p. 0 The authors present HADACA3 as a comprehensive dataset and benchmarking framework, established through a four-day collaborative competition followed by a large-scale computational test involving over 250,000 analysis pipelines across nine datasets featuring matched DNA methylation (DNAm) and RNA profiles under diverse biological and experimental conditions. p. 0

The benchmark's primary goal is to rigorously test the assumption that combining multi-omic data improves deconvolution performance. The framework jointly evaluates the impact of preprocessing, feature selection, modeling, and various integration strategies. p. 0 A key finding is that DNAm alone achieves the highest median performance across datasets, establishing it as the most stable and reliable single-modality approach. p. 0 However, multi-omics integration strategies can achieve superior top performance in specific datasets and pipeline configurations. p. 0

The study highlights a critical trade-off between robustness and peak performance: while multi-omics integration does not systematically improve average performance over DNAm alone, it can yield better results in tailored settings. p. 0 Specifically, late integration based on error-weighted averaging provides a strong baseline, whereas non-linear early integration methods, such as optimal transport (OT), show promising results on real biological datasets due to their ability to capture complex cross-modal relationships. p. 0 Early integration is noted to be sensitive to scale differences and noise heterogeneity across modalities, which normalization only partially corrects. p. 8 Conversely, late integration avoids these issues by decoupling modality-specific inference from aggregation, leading to improved robustness in heterogeneous settings. p. 9

The benchmark focuses on pancreatic cancer datasets with known cell-type compositions spanning in silico, in vitro, and in vivo settings. The feature space was restricted to the intersection of RNA and DNAm features available across all benchmark datasets and reference profiles. p. 2 Evaluation utilized twelve complementary metrics, including cohort-wise correlations, cell-type correlations, error metrics (RMSE, MAE, Aitchison distance, Jensen-Shannon divergence), and geometric metrics (AID, SDID), with an aggregate score being a weighted geometric mean of these normalized components. p. 3 Analysis revealed that cell-type correlation and Aitchison distance are the most discriminative elements of the aggregate score. p. 8 Furthermore, linear decomposition of aggregate scores indicated that the deconvolution algorithm drives late integration performance, while the integration method dominates early integration, largely because latent embedding approaches perform inconsistently. p. 5

The study concludes by suggesting specific directions for future research: using concatscale or OT for early integration and limeanRMSE for late integration as promising avenues. p. 9 Single-modality approaches remain strong baselines. p. 9 Limitations noted include focusing only on pancreatic cancer, the underperformance of RNA-only deconvolution due to lower quality reference profiles, and the statistical challenge posed by the large number of pipeline combinations where small score differences may appear consistent without being statistically meaningful based on median aggregate scores. p. 8 The computational cost for the full scope was estimated at over 575,000 core-hours. p. 6

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Optimal Transport based representation can align RNA-seq and DNAm distributions through entropic optimal transport, resulting in a joint latent representation that better captures complex cross-modal relationships in real biological settings.

  2. The system can perform cell type deconvolution by using the optimal transport based representation to find the most appropriate shared latent variable Zs, which is then used to estimate cell-type proportions pi such that zmix i approximately equals Zref pi.

  3. The deconvolution algorithm can be chosen from a spectrum including non-negative least squares (NNLS), robust regression (RLR), and ensemble approaches (RLRnnls) to handle noise, outlier features, and reference misspecification effectively.

Related papers