GAC-BiT-CNN: A Hybrid Generative Adversarial Capsule–Transformer–CNN Model with Fused Protein Language Model Embeddings for Methyltransferase Prediction
5 affiliations
- 1Department of Computer Science, Bahria University, Islamabad 44000, Pakistan
- 2Department of Information Systems, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia
- 3Department of Computer Science, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia
- 4Department of Information Systems and Technology, College of Computer Science and Engineering, University of Jeddah, Jeddah 23890, Saudi Arabia
- 5Department of Information Technology, College of Computer, Qassim University, Buraydah 52571, Saudi Arabia
Abstract
Methyltransferases (MTs) are a functionally diverse class of enzymes that catalyse methyl group transfer from S-adenosyl-L-methionine (SAM) to DNA, RNA, and proteins, thereby regulating epigenetic programming, gene expression, and numerous disease processes. Dysregulation of these enzymes is implicated in a wide spectrum of human cancers and neurodegenerative disorders, making their accurate computational identification a priority in both basic research and drug discovery. We present GAC-BiT-CNN, a novel hybrid deep learning framework integrating Generative Adversarial Capsule Networks (GAC), Bidirectional Transformers (BiT), and Convolutional Neural Networks (CNN). The model leverages high-dimensional protein embeddings from four protein language models, including ProtGen, ESM-2, ProtT5-BFD, and ProtT5-XL, used individually, together with a fused feature set formed by concatenating the ProtT5-BFD and ProtT5-XL embeddings. All architectural components were optimised via systematic grid search hyperparameter tuning, and SHapley Additive exPlanations (SHAP)-based analysis was applied to the fused embeddings to identify the most discriminative feature dimensions. GAC-BiT-CNN with the fused feature set achieved 98.34% accuracy on the training set and 95.50% accuracy on the held-out test set, with an area under the receiver operating characteristic curve (AUC) of 98.70% and a Matthews correlation coefficient (MCC) of 0.901, surpassing all baseline models by a substantial margin. GAC-BiT-CNN provides a scalable, alignment-free framework with representation-level SHAP attribution for high-throughput methyltransferase identification, with direct applications in epigenomics, drug discovery, and cancer biology.
Related
- RNAGEN: A Generative Adversarial Network‐Based Model to Generate Synthetic RNA Sequences to Target Proteins
- Representation learning for low-data protein fitness prediction using sequence–structure embeddings and graph neural networks
- Multi-tissue segmentation of knee MRI based on U-Net and conditional generative adversarial networks
- CBAT-Net: a unified CNN–BiLSTM–MultiHead transformer architecture for joint Q3/Q8 protein secondary structure prediction
- A novel LLM-based GPT-neo transformer with machine learning architecture for promoters prediction in Agrobacterium
BibTeX
@article{muazzam2026gacbitcnn,
title = {{GAC-BiT-CNN: A Hybrid Generative Adversarial Capsule–Transformer–CNN Model with Fused Protein Language Model Embeddings for Methyltransferase Prediction}},
author = {Muhammad Muazzam and Farman Ali and Raed Alsini and Naif Waheb Rajkhan and Omar Alghushairy and Tamim Alkhalifah},
journal = {International Journal of Molecular Sciences},
year = {2026},
month = nov,
doi = {10.3390/ijms27219967},
url = {https://www.mdpi.com/1422-0067/27/21/9967},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - GAC-BiT-CNN: A Hybrid Generative Adversarial Capsule–Transformer–CNN Model with Fused Protein Language Model Embeddings for Methyltransferase Prediction
AU - Muhammad Muazzam
AU - Farman Ali
AU - Raed Alsini
AU - Naif Waheb Rajkhan
AU - Omar Alghushairy
AU - Tamim Alkhalifah
PY - 2026
DA - 2026/11/06
JO - International Journal of Molecular Sciences
DO - 10.3390/ijms27219967
UR - https://www.mdpi.com/1422-0067/27/21/9967
AB - Methyltransferases (MTs) are a functionally diverse class of enzymes that catalyse methyl group transfer from S-adenosyl-L-methionine (SAM) to DNA, RNA, and proteins, thereby regulating epigenetic programming, gene expression, and numerous disease processes. Dysregulation of these enzymes is implicated in a wide spectrum of human cancers and neurodegenerative disorders, making their accurate computational identification a priority in both basic research and drug discovery. We present GAC-BiT-CNN, a novel hybrid deep learning framework integrating Generative Adversarial Capsule Networks (GAC), Bidirectional Transformers (BiT), and Convolutional Neural Networks (CNN). The model leverages high-dimensional protein embeddings from four protein language models, including ProtGen, ESM-2, ProtT5-BFD, and ProtT5-XL, used individually, together with a fused feature set formed by concatenating the ProtT5-BFD and ProtT5-XL embeddings. All architectural components were optimised via systematic grid search hyperparameter tuning, and SHapley Additive exPlanations (SHAP)-based analysis was applied to the fused embeddings to identify the most discriminative feature dimensions. GAC-BiT-CNN with the fused feature set achieved 98.34% accuracy on the training set and 95.50% accuracy on the held-out test set, with an area under the receiver operating characteristic curve (AUC) of 98.70% and a Matthews correlation coefficient (MCC) of 0.901, surpassing all baseline models by a substantial margin. GAC-BiT-CNN provides a scalable, alignment-free framework with representation-level SHAP attribution for high-throughput methyltransferase identification, with direct applications in epigenomics, drug discovery, and cancer biology.
SN - 1422-0067
ER -