A Biologically-Informed Multi-Layer Explainable AI Framework for Trustworthy Computational Genomics
International Journal of Experimental Research and Review · 30.04.2026 · 50 · 73-83
DOI: 10.52756/ijerr.2026.v50.006 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Experimental Research and Review, online erschienen am 30.04.2026. Verfasst von Saurabh Sharma und 3 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Artificial intelligence (AI) has substantially transformed computational genomics, enabling high-precision modelling of complex biological systems including variant calling, gene expression modelling, disease risk prediction, and population structure inference. Deep learning architectures, particularly convolutional neural networks (CNNs) and transformer-based models, have achieved 10–20% higher classification accuracy than classical statistical machine learning approaches on standard genomic benchmarks. However, the predominant use of opaque "black-box" models limits their adoption in high-stakes domains such as clinical genomics, precision medicine, and evolutionary biology, where understanding the biological basis of predictions is as critical as predictive accuracy itself. As the primary novelty of this work, we propose and formalise BIMLEX (Biologically-Informed Multi-Layer EXplainability), a unified five-layer framework that is the first in the computational genomics literature to jointly (i) embed structured biological prior knowledge — gene ontology, pathway databases, and molecular interaction networks — directly into the data representation layer; (ii) orchestrate model-specific, model-agnostic and causal explanation modules through a mathematically-grounded, multi-objective optimisation objective; and (iii) enforce clinical and regulatory compliance constraints within the same end-to-end pipeline. Unlike prior surveys that catalogue XAI methods post-hoc, BIMLEX provides a formal, unified optimisation objective that simultaneously maximises explanation quality (fidelity, stability, biological plausibility and reproducibility) subject to constraints on predictive performance and privacy preservation. A comprehensive comparative evaluation of SHAP, LIME, saliency mapping, attention mechanisms, Grad-CAM, and graph-based explanation methods demonstrates superior biological plausibility and regulatory alignment. The framework directly addresses four critical open challenges: the absence of standardised evaluation benchmarks; the lack of experimental biological validation; insufficient scalability to whole-genome data; and the tension between interpretability and data privacy. These results establish BIMLEX as a foundational architecture for the next generation of trustworthy, transparent and clinically deployable genomic AI systems.
Zitieren
APA
Sharma, S., Paranjape, V., Verma, S., & Hasan, Z. (2026). A Biologically-Informed Multi-Layer Explainable AI Framework for Trustworthy Computational Genomics. International Journal of Experimental Research and Review, 50, 73-83. https://doi.org/10.52756/ijerr.2026.v50.006
Vancouver
Sharma S, Paranjape V, Verma S, Hasan Z. A Biologically-Informed Multi-Layer Explainable AI Framework for Trustworthy Computational Genomics. International Journal of Experimental Research and Review. 2026;50:73-83. doi:10.52756/ijerr.2026.v50.006
BibTeX
@article{Sharma2026,
author = {Sharma, Saurabh and Paranjape, Vishal and Verma, Saurabh and Hasan, Zohaib},
title = {A Biologically-Informed Multi-Layer Explainable AI Framework for Trustworthy Computational Genomics},
journal = {International Journal of Experimental Research and Review},
year = {2026},
volume = {50},
pages = {73--83},
doi = {10.52756/ijerr.2026.v50.006}
}
EUVIMED