Generalizable Cross-Species Splice Site Prediction Through Multitask CNN Learning for Biomedical and Medical Applications
Interpersona An International Journal on Personal Relationships · 13.11.2026 · 20 (4) · e2057
DOI: 10.54899/iij.v20i4.2057 ↗Kurz gefasst
Fachartikel in der Zeitschrift Interpersona An International Journal on Personal Relationships, online erschienen am 13.11.2026. Verfasst von Ali Salem Bin Sama und 9 weiteren Autor:innen.
Abstract (Originalsprache)
Accurate identification of splice sites is a fundamental problem in computational genomics, as RNA splicing plays a crucial role in gene regulation, transcript diversity, and evolutionary conservation across species [11]. While recent deep learning-based approaches have demonstrated high accuracy for splice-site prediction within individual species [8, 9], their performance often degrades in cross-species settings due to species-specific sequence biases and limited generalization capability [12]. To address this challenge, this work proposes a Multi-Task Convolutional Neural Network (Multi-Task CNN) framework for robust cross-species splice-site prediction. The proposed model operates on one-hot encoded DNA sequence windows and employs a shared convolutional backbone composed of standard and dilated one-dimensional convolutional layers, augmented with a self-attention mechanism to capture both local splice-site motifs and long-range contextual dependencies [3, 6]. Splice-site classification is treated as the primary task, while species prediction is incorporated as an auxiliary task to encourage the learning of biologically meaningful and evolutionarily conserved representations. The model is trained using a weighted multi-task loss function that jointly optimizes both objectives, following established principles in multi-task and representation learning [1, 4]. Comprehensive experimental evaluation demonstrates consistently high performance across standard classification metrics, including accuracy, precision, recall, and F1-score. Analysis of confusion matrices, training stability, and confidence-based predictions confirms the robustness and strong generalization capability of the proposed approach on unseen genomic data. Overall, this study establishes multi-task learning as an effective strategy for mitigating species-specific overfitting and improving cross-species splice-site prediction.
Zitieren
APA
Sama, A. S. B., Kumar, A., Varman, V. M., Dar, S. A., Rather, A. A., Bashir, R., Osman, A. A. F., Ataelfadiel, M., Abdullah, A., & Idress, O. E. E. (2026). Generalizable Cross-Species Splice Site Prediction Through Multitask CNN Learning for Biomedical and Medical Applications. Interpersona An International Journal on Personal Relationships, 20(4), e2057. https://doi.org/10.54899/iij.v20i4.2057
Vancouver
Sama ASB, Kumar A, Varman VM, Dar SA, Rather AA, Bashir R, et al. Generalizable Cross-Species Splice Site Prediction Through Multitask CNN Learning for Biomedical and Medical Applications. Interpersona An International Journal on Personal Relationships. 2026;20(4):e2057. doi:10.54899/iij.v20i4.2057
BibTeX
@article{Sama2026,
author = {Sama, Ali Salem Bin and Kumar, Abhay and Varman, Vijaya Mahindra and Dar, Showkat A. and Rather, Aafaq A. and Bashir, Raeesa and Osman, Ahmed A. F. and Ataelfadiel, Mohammed and Abdullah, Ala and Idress, Osman Elwasila Elshiekh},
title = {Generalizable Cross-Species Splice Site Prediction Through Multitask CNN Learning for Biomedical and Medical Applications},
journal = {Interpersona An International Journal on Personal Relationships},
year = {2026},
volume = {20},
number = {4},
pages = {e2057},
doi = {10.54899/iij.v20i4.2057}
}
EUVIMED