Light Attention

Light Attention predicts protein subcellular localization from embeddings produced by protein language models (pLMs) and pre-trained Transformers, avoiding reliance on multiple sequence alignments (MSAs).

Key Features:

• Light Attention (LA): softmax-weighted aggregation mechanism providing linear complexity relative to protein sequence length.
• Deep Learning Architecture: lightweight deep neural network that integrates the LA mechanism for sequence-level aggregation.
• Protein Language Models and Transformers: uses embeddings from pLMs and pre-trained Transformers instead of MSA-derived input features.
• MSA-free prediction: bypasses computation of multiple sequence alignments, reducing dependence on MSA-based expert features.
• Performance: demonstrates superior performance across ten subcellular localization classes with approximately an eight percentage point improvement in Q10 accuracy over prior SOTA.
• Evaluation and test data: assessed using a newly compiled test set that highlights limitations of static benchmark datasets.
• Scalability: architecture supports large-scale prediction across entire proteomes, including the human proteome.

Scientific Applications:

• Proteomics: annotation of proteomes by predicting protein subcellular localization to infer protein function.
• Disease Research: identification of mislocalized proteins and localization changes relevant to disease mechanisms.
• Drug Discovery: informing target selection and mechanism-of-action hypotheses by localizing candidate proteins.
• Functional Genomics: integrating localization predictions to interpret gene function and cellular pathways.

Methodology:

Embeddings from pLMs and pre-trained Transformers are input to a lightweight deep neural network that applies a softmax-weighted light attention (LA) aggregation mechanism, delivering linear complexity with sequence length to predict subcellular localization classes.