Multi-marginal Contrastive Learning for Multilabel Subcellular Protein Localization

Ziyi Liu, Zengmao Wang, Bo Du

Introduction

Protein Subcellular Localization (PSL) serves as a critical "Google Maps" for cell biology, pinpointing exactly where proteins execute their functions. Accurate localization is essential for understanding disease mechanisms, such as cancer, where protein mislocalization often signals pathological states. However, automating PSL using computer vision faces significant hurdles: the massive scale of immunohistochemistry (IHC) images (often 3000×3000 pixels), high morphological variability across different tissue types (e.g., liver vs. kidney), and the multi-label nature of protein distribution, where approximately 20% of proteins reside in multiple locations simultaneously.

The DeePSLoc Framework

To address these challenges, researchers developed DeePSLoc, a deep learning framework that combines a dual-branch "Global-Local" architecture with a novel multi-marginal contrastive learning strategy.

Dual-Branch Architecture

DeePSLoc processes high-resolution IHC images without losing critical details or exceeding GPU memory limits by splitting the task:

  1. Global Branch: Analyzes a downsampled version of the entire image to capture broad tissue structure and context.
  2. Local Branch: Utilizes activation maps from the global branch to identify and crop informative high-resolution patches (256×256 pixels), where fine-grained protein signatures are most prominent.
Tissue variability and local-global context in protein localization.

Multi-marginal Contrastive Learning

Standard contrastive learning typically employs a fixed distance margin to separate similar and dissimilar features. DeePSLoc introduces a multi-marginal attention mechanism that acknowledges the inherent variability of biological data. By weightly aggregating samples, the model generates a spectrum of "easy" and "hard" positive/negative pairs. This allows the network to learn robust, invariant features across diverse tissue morphologies, effectively navigating the messy distributions found in real-world IHC samples.

Experimental Results

DeePSLoc was evaluated on four datasets from the Human Protein Atlas (HPA), including single-label (HPA-7, HPA-8) and challenging multi-label benchmarks (Multi-HPA, HPA-18).

  • Single-Label Performance: On the HPA-7 dataset, DeePSLoc achieved a remarkable 97.95% accuracy, significantly surpassing the previous state-of-the-art (SOTA) of 89.09%.
  • Multi-Label Performance: On the complex HPA-18 dataset, the model reached 61.2% subset accuracy, outperforming specialized models like ML-GCN (60.3%) and C-Tran (57.9%).
  • Ablation Study: Removing the contrastive learning component led to a sharp decline in performance (e.g., from ~92% to ~71% on HPA-7), confirming that the multi-marginal strategy is the primary driver of the model's success.

Discussion and Future Directions

The success of DeePSLoc underscores the importance of combining global context with local detail and the effectiveness of flexible, attention-based contrastive learning in biological domains. While the current dual-branch approach introduces some computational overhead during inference due to patch processing, it provides a scalable solution for high-resolution medical imaging. Future work may focus on stabilizing end-to-end training and exploring the framework's generalizability to other domains, such as whole-slide pathology imaging.

Summary updated

Abstract

Protein subcellular localization(PSL) is an important task to study human cell functions and cancer pathogenesis. It has attracted great attention in the computer vision community. However, the huge size of immune histochemical (IHC) images, the disorganized location distribution in different tissue images and the limited training images are always the challenges for the PSL to learn a strong generalization model with deep learning. In this paper, we propose a deep protein subcellular localization method with multi-marginal contrastive learning to perceive the same PSLs in different tissue images and different PSLs within the same tissue image. In the proposed method, we learn the representation of an IHC image by fusing the global features from the downsampled images and local features from the selected patches with the activation map to tackle the oversize of an IHC image. Then a multi-marginal attention mechanism is proposed to generate contrastive pairs with different margins and improve the discriminative features of PSL patterns effectively. Finally, the ensemble prediction of each IHC image is obtained with different patches. The results on the benchmark datasets show that the proposed method achieves significant improvements for the PSL task.
BibTeX
@inproceedings{liu2022multi,
  author = {Ziyi Liu and Zengmao Wang and Bo Du},
  title = {{Multi-marginal Contrastive Learning for Multilabel Subcellular Protein Localization}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.1109/cvpr52688.2022.01997},
}
Figure