FusionEncoder

FusionEncoder predicts intrinsically disordered regions (IDRs) in proteins by integrating traditional biological features and protein language model features through an LSTM-based fusion network followed by a transformer-based encoder.

Key Features:

• Multi-semantic feature integration: FusionEncoder treats traditional biological features and pre-trained protein language model (PPLMs)-based features as two distinct semantic input types within a multi-semantic space.
• LSTM variant fusion network: Traditional biological features are input into the cell state of a modified long short-term memory (LSTM) network, while PPLMs-based features are fed into the input part, with a fusion cell combining both feature types.
• Context-aware sequence encoding: The LSTM architecture leverages its capability to encode long sequences for enhanced context-aware semantic learning of amino acid sequences.
• Transformer-based prediction: A transformer-based encoder layer performs the final prediction of IDRs from the fused representations.

Scientific Applications:

• IDR identification: Accurate prediction of intrinsically disordered regions in proteins for analyzing protein structure and function.
• Amino acid feature representation: Improved computational representation of amino acid features as a foundational process for predictive model development.

Methodology:

FusionEncoder uses a variant of long short-term memory (LSTM) where traditional biological features enter the cell state and PPLMs-based features enter the input part; a fusion cell fuses these features, followed by a transformer-based encoder layer for IDR prediction. Evaluation was performed on four independent test datasets.