# SLiMs prediction method based on enhanced attention mechanism and feature fusion

## TL;DR

This paper introduces EMAF_SLiMs, a new method for predicting short linear motifs in proteins using enhanced attention and feature fusion, improving accuracy over existing approaches.

## Contribution

The novel EMAF_SLiMs method combines enhanced attention mechanisms and feature fusion for more accurate SLiMs prediction.

## Key findings

- EMAF_SLiMs outperforms existing methods on independent test sets.
- The enhanced attention model effectively highlights SLiMs characteristics.
- Multi-head attention successfully fuses semantic, physicochemical, and evolutionary features.

## Abstract

Short linear motifs (SLiMs) are functional regions composed of short sequences of specific amino acids. They usually do not have independent 3D three-dimensional structures, but play important roles in biological processes. Traditional detection methods have high cost and heavy workload, therefore it is necessary to seek an accurate detection method for SLiMs.

In this paper, we propose a new SLiMs prediction method, named EMAF_SLiMs, based on enhanced attention mechanism and feature fusion. We calculate three features sets which contain semantic embedding, physicochemical characteristic and evolutionary information. Then, we design the enhanced attention model based on SwiftFormer to highlight the characteristic of SLiMs. In addition, the multi-head attention mechanism is employed to effectively fuse these three feature sets. Finally, we construct an MLP network for prediction. EMAF_SLiMs has better performance on independent test sets, compared to other existing methods.

The source code and sample data are available via a Github project at https://github.com/jdchhh/EMAF_SLiMs/tree/master.