MMLoc: A Multi-instance Multi-label Learning Approach for Predicting Protein Subcellular Localization from Immunofluorescence Images

This is the authors' abstract. We don't add key points for this paper.

The emergence and development of immunofluorescence (IF) imaging techniques offers the opportunity to observe protein distribution and localization at the cellular level, which is critical for revealing cellular mechanisms and disease pathology. Deep learning-based artificial intelligence technologies have played an important role in advancing the field of IF image recognition by harnessing these techniques. However, using deep neural networks to directly predict protein localization (PSL) in these complex images remains extremely challenging, especially when dealing with proteins that exhibit multiple localizations across heterogeneous cell populations within the same sample. To address this challenge, we formulate the PSL modeling problem as a Multi-Instance Multi-Label (MIML) learning task and propose a novel PSL model named MMLoc. Specifically, considering an IF image contains multiple cells, we utilize U-Net to segment the IF image and treat each cell as an instance, making this method especially suitable for scenarios where the image possesses multiple labels simultaneously and exhibits single-cell variation. On this basis, we employ a pre-trained ResNet50 model to extract features from instances, integrating a self-attention mechanism to focus on the most discriminative features. After fusing these instances, we connect them through a traditional fully connected classifier for predicting the protein localization class. The experimental results demonstrate that our model achieves competitive performance on a challenging HPA dataset.