Enhancing robustness in protein function prediction via missing modality imputation and adaptive multimodal fusion.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein function prediction is crucial in genomics.
• Current methods struggle with incomplete data and fusion strategies.
• Integrating multiple biological data types shows promise.

Purpose of the Study:

• To develop a novel framework, ProMIAF, for robust protein function prediction.
• To address challenges of missing modalities and suboptimal fusion.
• To improve the accuracy and robustness of predicting protein functions.

Main Methods:

• ProMIAF uses Missing Modality Imputation and Adaptive Multimodal Fusion.
• Advanced protein generation techniques recover missing structural and textual data.
• Modality-specific encoders, gated attention, and network propagation are employed.

Main Results:

• ProMIAF significantly outperforms state-of-the-art methods in predicting novel functional annotations.
• Achieved improvements of 8.21% (AUPR), 8.90% (Smin), and 2.61% (Fmax) on the biological process branch.
• Demonstrated robustness and accuracy even without structural and textual modalities.

Conclusions:

• ProMIAF offers a robust solution for protein function prediction challenges.
• The framework effectively leverages cross-modal complementarity.
• ProMIAF advances the field by improving predictive performance with incomplete data.