A multi-head attention-based neural network for cancer-specific identification of
- 1Department of Applied Computer Science, The University of Winnipeg, Winnipeg, MB R3B 2E9, Canada; Department of Computer Science, Abdul Wali Khan University, Mardan, KP 23200, Pakistan
- 2Department of Applied Computer Science, The University of Winnipeg, Winnipeg, MB R3B 2E9, Canada
- 3Department of Applied Computer Science, The University of Winnipeg, Winnipeg, MB R3B 2E9, Canada; Department of Biochemistry and Medical Genetics, University of Manitoba, Winnipeg, MB R3E 0J9, Canada; Corresponding author at: Department of Applied Computer Science, The University of Winnipeg, Winnipeg, MB R3B 2E9, Canada.
Abstract
Objective: N 7 -methylguanosine ( m 7 G ) is one of the most prevalent and biologically important ribonucleic acid modifications. While several computational models exist for m 7 G site prediction in human and non-human species, none have been specifically designed to detect m 7 G modification sites in cancers. This study addresses this gap by developing a cancer-specific predictive framework for m 7 G site identification. Methods: We curated four m 7 G benchmark datasets from two cancer cell lines (human osteosarcoma and human hepatocellular carcinoma) from the Gene Expression Omnibus repository and included an additional benchmark dataset from m 7 G Hub v2.0. We developed a multi-head attention-based convolutional neural network (CNN) model using two feature encoding methods—trinucleotide composition and one-hot encoding. The model’s performance was compared to support vector machine, random forest, extreme gradient boosting, and a CNN across all five datasets. Robustness and generalizability were evaluated using 5-fold cross-validation and independent testing. Model interpretability was examined with two explanation tools. Results: The proposed model achieved the highest mean accuracy, Matthews correlation coefficient, and area under the receiver operating characteristic curve (AUROC) among the five evaluated models on each independent test dataset. Its mean accuracy ranged from 0.8397 to 0.8833 and its mean AUROC ranged from 0.9163 to 0.9489. Sensitivity was not consistently highest, and sensitivity and specificity were not uniformly superior across all evaluation settings. Feature attribution and visualization analyses characterized model representations. Conclusions: This study presents one of the first cancer-specific computational frameworks for m 7 G site prediction. By combining multi-head attention with CNNs and complementary feature encoding, our approach enhances predictive power and interpretability. The framework offers a valuable tool for elucidating disease mechanisms, mapping epitranscriptomic signatures, and accelerating biomarker and therapeutic discovery in cancer research. The source code and curated dataset are publicly available at https://github.com/malikmtahir/m7G .
Related
- Comments on "SpatialDG: a novel spatial domain identification method for spatially resolved transcriptomics data based on dual-graph neural network".
- Identification of substrates for the skeletal muscle-specific lysine methyltransferases METTL21C and METTL21E
- Attention-based hierarchical graph autoencoder for dose-specific single-cell resistance dynamics
- SFOA-based optimized SeResNet-BiLSTM network with an attention mechanism for the detection of exons.
- Prediction of bacteria-specific drug metabolism using a cross-attention-based model.
BibTeX
@article{tahir2026multihead,
title = {{A multi-head attention-based neural network for cancer-specific identification of}},
author = {Muhammad Tahir and Sheela Ramanna and Qian Liu},
journal = {Intelligent Oncology},
year = {2026},
month = oct,
doi = {10.1016/j.intonc.2026.100089},
url = {https://doi.org/10.1016/j.intonc.2026.100089},
note = {CC BY},
}
RIS
TY - JOUR
TI - A multi-head attention-based neural network for cancer-specific identification of
AU - Muhammad Tahir
AU - Sheela Ramanna
AU - Qian Liu
PY - 2026
DA - 2026/10/25
JO - Intelligent Oncology
DO - 10.1016/j.intonc.2026.100089
UR - https://doi.org/10.1016/j.intonc.2026.100089
AB - Objective: N 7 -methylguanosine ( m 7 G ) is one of the most prevalent and biologically important ribonucleic acid modifications. While several computational models exist for m 7 G site prediction in human and non-human species, none have been specifically designed to detect m 7 G modification sites in cancers. This study addresses this gap by developing a cancer-specific predictive framework for m 7 G site identification. Methods: We curated four m 7 G benchmark datasets from two cancer cell lines (human osteosarcoma and human hepatocellular carcinoma) from the Gene Expression Omnibus repository and included an additional benchmark dataset from m 7 G Hub v2.0. We developed a multi-head attention-based convolutional neural network (CNN) model using two feature encoding methods—trinucleotide composition and one-hot encoding. The model’s performance was compared to support vector machine, random forest, extreme gradient boosting, and a CNN across all five datasets. Robustness and generalizability were evaluated using 5-fold cross-validation and independent testing. Model interpretability was examined with two explanation tools. Results: The proposed model achieved the highest mean accuracy, Matthews correlation coefficient, and area under the receiver operating characteristic curve (AUROC) among the five evaluated models on each independent test dataset. Its mean accuracy ranged from 0.8397 to 0.8833 and its mean AUROC ranged from 0.9163 to 0.9489. Sensitivity was not consistently highest, and sensitivity and specificity were not uniformly superior across all evaluation settings. Feature attribution and visualization analyses characterized model representations. Conclusions: This study presents one of the first cancer-specific computational frameworks for m 7 G site prediction. By combining multi-head attention with CNNs and complementary feature encoding, our approach enhances predictive power and interpretability. The framework offers a valuable tool for elucidating disease mechanisms, mapping epitranscriptomic signatures, and accelerating biomarker and therapeutic discovery in cancer research. The source code and curated dataset are publicly available at https://github.com/malikmtahir/m7G .
SN - 2950-2616
SN - 2097-7131
ER -