ATP_mCNN: Predicting ATP binding sites through pretrained language models and multi-window neural networks

Area of Science:

• Computational biology
• Bioinformatics
• Protein structure prediction

Background:

• Adenosine triphosphate (ATP) is crucial for cellular energy and function, interacting with numerous proteins.
• Identifying ATP-binding sites in proteins is essential for understanding biological processes.
• Experimental methods for identifying ATP-binding residues are challenging and time-consuming.

Purpose of the Study:

• To develop a novel computational method for accurately predicting adenosine triphosphate-binding residues in protein sequences.
• To leverage pre-trained protein language models and a multi-window convolutional neural network architecture for improved prediction accuracy.

Main Methods:

• Utilized pre-trained protein language model embeddings as input features.
• Developed a multi-window convolutional neural network (CNN) architecture with parallel convolutional layers for motif scanning at various window sizes.
• Employed max pooling to extract salient features, concatenated across windows for a multi-scale representation.
• Performed residue-level classification for ATP-binding site prediction.

Main Results:

• The developed model achieved an Area Under the ROC Curve (AUC) of 0.95 on benchmark datasets.
• Significantly improved prediction performance compared to previous sequence-based models.
• Outperformed existing convolutional neural network (CNN) baseline models.

Conclusions:

• Pre-trained language models and multi-window CNNs are effective for sequence-based prediction of ATP-binding residues.
• The proposed approach offers a promising new direction for predicting binding sites and elucidating molecular interactions from primary protein sequences.
• This computational tool can aid in understanding protein function and designing new therapeutics.