Explainable multi-task learning for multi-modality biological data analysis

Xin Tang, Jiawei Zhang, Yichun He, Xinhe Zhang, Zuwan Lin, Sebastian Partarrieu, Emma Bou Hanna, Zhaolin Ren, Hao Shen, Yuhong Yang, Xiao Wang, Na Li, Jie Ding, Jia LiuNature Communications 2023 · DOI

Abstract

Current biotechnologies can simultaneously measure multiple high-dimensional modalities (e.g., RNA, DNA accessibility, and protein) from the same cells. A combination of different analytical tasks (e.g., multi-modal integration and cross-modal analysis) is required to comprehensively understand such data, inferring how gene regulation drives biological diversity and functions. However, current analytical methods are designed to perform a single task, only providing a partial picture of the multi-modal data. Here, we present UnitedNet, an explainable multi-task deep neural network capable of integrating different tasks to analyze single-cell multi-modality data. Applied to various multi-modality datasets (e.g., Patch-seq, multiome ATAC + gene expression, and spatial transcriptomics), UnitedNet demonstrates similar or better accuracy in multi-modal integration and cross-modal prediction compared with state-of-the-art methods. Moreover, by dissecting the trained UnitedNet with the explainable machine learning algorithm, we can directly quantify the relationship between gene expression and other modalities with cell-type specificity. UnitedNet is a comprehensive end-to-end framework that could be broadly applicable to single-cell multi-modality biology. This framework has the potential to facilitate the discovery of cell-type-specific regulation kinetics across transcriptomics and other modalities.

Introduction

Modern "multi-omic" biotechnologies allow researchers to observe different layers of biological information—such as gene expression (RNA), chromatin accessibility (ATAC), and protein abundance—within the exact same cell. However, the computational landscape remains fragmented. Most existing tools are "single-task" specialists, focusing either on multi-modal integration (clustering cells into types) or cross-modal prediction (predicting one modality from another). This fragmented approach provides only a partial picture of cellular function and often fails to account for the high noise and heterogeneity inherent in biological data. UnitedNet addresses these challenges by providing a unified, end-to-end framework that handles multiple analytical tasks simultaneously while remaining interpretable.

The UnitedNet Architecture

UnitedNet utilizes an encoder-decoder-discriminator architecture designed for robust multi-modal fusion. Its core innovation lies in a multi-task learning scheme that alternates between joint group identification and cross-modal translation. Encoders project disparate modalities into a shared latent space, where cells are clustered into biological types.

UnitedNet architecture: Encoders project modalities into a shared latent space, while decoders and discriminators facilitate cross-modal prediction and adversarial training.

To ensure high-fidelity results, the model incorporates adversarial training, where a discriminator network forces predicted data to look indistinguishable from real biological measurements. Furthermore, an adaptive weighting mechanism automatically assigns lower weights to noisier modalities, preventing low-quality data from degrading the overall analysis.

Explainable Multi-Task Learning and Synergy

A key insight of the UnitedNet framework is the synergy between tasks: joint training for cross-modal prediction actually improves the alignment of the shared latent space, leading to more accurate cell-type clustering than single-task methods. To move beyond "black-box" predictions, the researchers integrated the SHAP (SHapley Additive exPlanations) algorithm. This allows the model to quantify exactly how much a specific feature (such as a gene) contributes to a cell's identity or a predicted physical trait.

Explainable multi-task learning workflow: Task 1 focuses on joint group identification, while Task 2 handles cross-modal prediction. Interpretation modules quantify feature-to-group and feature-to-feature relevance.

Results and Biological Discovery

UnitedNet was validated across a diverse array of benchmarks, including Patch-seq, multiome ATAC + RNA, and spatial transcriptomics. In unsupervised cell-type identification, UnitedNet consistently outperformed or matched state-of-the-art methods like Schema, totalVI, and Weighted Nearest Neighbor (WNN), as measured by the Adjusted Rand Index (ARI). In cross-modal tasks, it achieved higher $R^2$ values and AUC than standard coupled autoencoders.

The model's robustness was demonstrated in simulations where one modality had 90% "dropout" (missing data); UnitedNet’s adaptive weighting maintained high accuracy while other methods failed. Furthermore, the model successfully performed batch correction across 12 batches of human brain tissue, neutralizing technical effects. Beyond performance metrics, UnitedNet facilitated biological discovery by identifying the gene Lrrc38 as a key driver of firing rates in Pvalb neurons, recapitulating known biology and suggesting new regulatory mechanisms.

Conclusion

UnitedNet represents a significant advancement in the analysis of single-cell multi-modality data. By unifying integration and prediction within an explainable architecture, it provides a comprehensive tool for discovering cell-type-specific regulation kinetics. While future work may involve automated hyperparameter tuning and expanding to tasks like trajectory inference, UnitedNet currently stands as a powerful end-to-end solution for modern biological research.

Summary updated

Abstract

Current biotechnologies can simultaneously measure multiple high-dimensional modalities (e.g., RNA, DNA accessibility, and protein) from the same cells. A combination of different analytical tasks (e.g., multi-modal integration and cross-modal analysis) is required to comprehensively understand such data, inferring how gene regulation drives biological diversity and functions. However, current analytical methods are designed to perform a single task, only providing a partial picture of the multi-modal data. Here, we present UnitedNet, an explainable multi-task deep neural network capable of integrating different tasks to analyze single-cell multi-modality data. Applied to various multi-modality datasets (e.g., Patch-seq, multiome ATAC + gene expression, and spatial transcriptomics), UnitedNet demonstrates similar or better accuracy in multi-modal integration and cross-modal prediction compared with state-of-the-art methods. Moreover, by dissecting the trained UnitedNet with the explainable machine learning algorithm, we can directly quantify the relationship between gene expression and other modalities with cell-type specificity. UnitedNet is a comprehensive end-to-end framework that could be broadly applicable to single-cell multi-modality biology. This framework has the potential to facilitate the discovery of cell-type-specific regulation kinetics across transcriptomics and other modalities.
BibTeX
@inproceedings{tang2023explainable,
  author = {Xin Tang and Jiawei Zhang and Yichun He and Xinhe Zhang and Zuwan Lin and Sebastian Partarrieu and Emma Bou Hanna and Zhaolin Ren and Hao Shen and Yuhong Yang and Xiao Wang and Na Li and Jie Ding and Jia Liu},
  title = {{Explainable multi-task learning for multi-modality biological data analysis}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-37477-x},
}
Figure