Multi-task learning for single-cell multi-modality biology

Xin Tang, Jiawei Zhang, Yichun He, Xinhe Zhang, Zuwan Lin, Sebastian Partarrieu, Emma Bou Hanna, Zhaolin Ren, Yuhong Yang, Xiao Wang, Na Li, Jie Ding, Jia LiubioRxiv 2022 · DOI

TL;DR

This paper introduces UnitedNet, a novel multi-task deep learning framework that simultaneously tackles joint group identification and cross-modal prediction for single-cell multi-modality data. By integrating these tasks, UnitedNet significantly improves performance on both, outperforming existing state-of-the-art methods across various biological datasets, including those with spatial information.

What Problem Does This Paper Address?

Modern biotechnologies can now measure multiple high-dimensional "modalities" (like gene expression, DNA accessibility, or electrophysiology) from the same single cell. Analyzing this rich data is crucial for understanding complex biological systems. However, existing methods often focus on a single task, such as identifying cell types (joint group identification) or predicting one modality from another (cross-modal prediction), without leveraging the inherent connections between these tasks. This siloed approach limits the full potential of multi-modal datasets. The authors aim to bridge this gap by developing a unified framework that can perform both tasks concurrently, thereby improving overall analytical power and efficiency.

Methods

UnitedNet is a deep neural network built upon an encoder-decoder-discriminator structure. It employs a multi-task learning strategy, training jointly for group identification and cross-modal prediction.

  • Encoders: For each modality, a dedicated encoder maps the raw data into a low-dimensional, modality-specific latent code.
  • Fusion: These modality-specific latent codes are then fused into a shared latent code, which is central to both tasks.
  • Decoders: Decoders reconstruct data from latent codes, enabling both within-modality (reconstruction) and cross-modality (prediction) tasks.
  • Discriminators: These are used in an adversarial manner to improve the quality of generated data and the accuracy of cross-modal predictions.
  • Multi-task Learning: The core innovation is the simultaneous training of two main tasks:
  • Joint Group Identification: This task aims to cluster cells based on their shared latent representations, leveraging both unsupervised (clustering loss) and supervised (classification loss) approaches.
  • Cross-modal Prediction: This task predicts data from one modality given another, using a prediction loss.
  • Loss Functions: A combination of losses, including clustering, contrastive, reconstruction, prediction, discriminator, and generator losses, are used to train the network. The training procedure alternates between updating the group identification and prediction components.

Datasets

UnitedNet was evaluated on a variety of single-cell multi-modality datasets:

  • Simulated Multi-modality Data: Generated using Dyngen, including DNA, pre-mRNA, mRNA, and protein modalities.
  • Multi-sensing Data: Simultaneously measured transcriptomics and intracellular electrophysiology data.
  • Multi-omics Data: Integratively profiled transcriptomics and DNA accessibility data.
  • Spatial-omics Data: Spatially resolved transcriptomics and proteomics data.
  • Patch-seq Dataset: Characterizing cell morphology, electrophysiology, and transcriptomics.
  • ATAC-seq BMMC Dataset: Measuring gene expression and DNA accessibility from bone marrow mononuclear cells across multiple batches.
  • DBiT-seq Embryo Dataset: Measuring mRNA expression, protein expression, and niche mRNA expression.
  • DLPFC Dataset: Spatially mapping gene expression and H&E staining.
  • MNIST Dataset: Used as a multi-modal input for validation, with original, edge-transformed, invert-transformed, and pure noise modalities.

Key Insights and Ideas

  • Synergy of Tasks: Jointly training group identification and cross-modal prediction significantly boosts the performance of both tasks, suggesting a strong complementary relationship.
  • Shared Latent Space as a Bridge: The shared latent space learned by UnitedNet acts as a crucial intermediary, allowing knowledge transfer and alignment between different modalities and tasks.
  • Spatial Information as a Modality: UnitedNet effectively incorporates spatial information (e.g., cell neighborhood) as a distinct modality, leading to improved tissue region identification and spatially resolved predictions.
  • Robustness to Noise: The framework demonstrates resilience to modality-specific noise, outperforming other methods in simulated noisy datasets.

Key Results

  • Superior Performance: UnitedNet consistently outperformed state-of-the-art methods in both unsupervised and supervised group identification and cross-modal prediction across various datasets.
  • Improved Accuracy: For example, on the ATAC-seq dataset, UnitedNet showed enhanced performance for annotation transfer and cross-modal prediction. On spatial-omics data, it improved tissue region identification accuracy.
  • Effective Noise Suppression: In simulated noisy datasets, UnitedNet was able to identify and suppress noise, leading to better performance.
  • Cross-modal Prediction Fidelity: UnitedNet enabled high-fidelity prediction of measurements across multiple modalities, even for complex datasets like Patch-seq.

Limitations and Future Work

  • Remaining Questions: The authors acknowledge that further research is needed to fully understand why multi-task learning improves performance in this context, whether the network can integrate even more tasks (e.g., identifying cross-modal feature relevance), and how to design customized algorithms for specific datasets.
  • Future Directions: Future mechanistic studies, including ablation experiments and theoretical analysis, are proposed to address these remaining questions. Additionally, improvements could be made by automating network configuration and hyperparameter optimization.

Summary updated

Abstract

Abstract Current biotechnologies can simultaneously measure multi-modality high-dimensional information from the same cell and tissue samples. To analyze the multi-modality data, common tasks such as joint data analysis and cross-modal prediction have been developed. However, current analytical methods are generally designed to process multi-modality data for one specific task without considering the underlying connections between tasks. Here, we present UnitedNet, a multi-task deep neural network that integrates the tasks of joint group identification and cross-modal prediction to analyze multi-modality data. We have found that multi-task learning for joint group identification and cross-modal prediction significantly improves the performance of each task. When applied to various single-cell multi-modality datasets, UnitedNet shows superior performance in each task, achieving better unsupervised and supervised joint group identification and cross-modal prediction performances compared with state-of-the-art methods. Furthermore, by considering the spatial information of cells as one modality, UnitedNet substantially improves the accuracy of tissue region identification and enables spatially resolved cross-modal prediction.
BibTeX
@misc{tang2022multi,
  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 Yuhong Yang and Xiao Wang and Na Li and Jie Ding and Jia Liu},
  title = {{Multi-task learning for single-cell multi-modality biology}},
  year = {2022},
  doi = {10.1101/2022.06.03.494730},
  howpublished = {bioRxiv},
}
Figure