Uploaded July 2026 | Updated November 2026, 57 minutes ago
Speaker:
Burcu Bakir Gungor, Associate Professor at the Department of Computer Engineering at Abdullah Gul University, CEO and founder of SmartTech.
Abstract:
Identifying dysregulated biomolecules is vital for understanding disease mechanisms. While traditional computational feature selection methods are widely used to analyze -omics datasets, they typically rank features individually based on statistics. This neglects biological domain knowledge and feature dependencies, often yielding redundant results and limited interpretability. Consequently, incorporating biological context into machine learning has become essential. Along this line, our group proposed the Grouping–Scoring–Modeling (G-S-M) approach to evaluate groups of features (e.g., genes, microRNAs, microorganisms) rather than individual markers.
The feature groups can be either generated by utilizing pre-existing domain knowledge stored in a biological database or through a fully data-driven approach using statistical measures. These groups are generated using pre-existing domain knowledge or fully data-driven statistical measures. This approach has driven the development of numerous bioinformatics tools, including miRcorrNet, PriPath, miRModuleNet, 3Mint, GeNetOntology, mirDisNet, microBiomeGSM, and AMP-GSM.
In this talk, Burcu Bakir Gungor first details domain-knowledge-based G-S-M approaches and data-driven methods—specifically based on Recursive Cluster Elimination—that they have recently developed and applied across various single and multi-omics datasets. Next, she presents extensions of their established G-S-M framework to operate on multi-layer networks, systematically identifying dysregulated modules as robust biomarker signatures. To demonstrate the practical utility of this extended methodology, she highlights their newly launched work on colorectal cancer (CRC) metagenomics data analysis. They model the CRC microbiome through a multi-layer network representing microbial species, encoded enzymes, and metabolic pathways.
By applying CROSSIM multiplex embedding and extending their G-S-M framework, they holistically analyze structural and functional dysbiosis to drive early CRC detection and personalized therapies. To develop clinically explainable, personalized probiotic and dietary recommendations, they develop LLMs with a GraphRAG architecture.
In parallel, she discusses their initiative on computational anticancer peptide prediction. They overcome traditional sequence model limitations by using a multi-layer network that captures diverse physicochemical, evolutionary, and structural peptide properties simultaneously. To efficiently analyze these complex structures and mitigate oversmoothing, they leverage Arnoldi Graph Convolutional Networks, integrating multiplex network theory with advanced deep learning to accelerate targeted drug discovery.
Ultimately, her group aims to elucidate the molecular mechanisms behind disease development using advanced machine learning. These insights can be exploited for early clinical diagnostics or as druggable targets to help physicians design personalized medical treatment plans.
More information here:
bsc.es/research-and-development/research-seminars/sorswomeninbsc-data-driven-and-domain-knowledge-based-feature-grouping-and-scoring-omics-data
Speaker:
Burcu Bakir Gungor, Associate Professor at the Department of Computer Engineering at Abdullah Gul University, CEO and founder of SmartTech.
Abstract:
Identifying dysregulated biomolecules is vital for understanding disease mechanisms. While traditional computational feature selection methods are widely used to analyze -omics datasets, they typically rank features individually based on statistics. This neglects biological domain knowledge and feature dependencies, often yielding redundant results and limited interpretability. Consequently, incorporating biological context into machine learning has become essential. Along this line, our group proposed the Grouping–Scoring–Modeling (G-S-M) approach to evaluate groups of features (e.g., genes, microRNAs, microorganisms) rather than individual markers.
The feature groups can be either generated by utilizing pre-existing domain knowledge stored in a biological database or through a fully data-driven approach using statistical measures. These groups are generated using pre-existing domain knowledge or fully data-driven statistical measures. This approach has driven the development of numerous bioinformatics tools, including miRcorrNet, PriPath, miRModuleNet, 3Mint, GeNetOntology, mirDisNet, microBiomeGSM, and AMP-GSM.
In this talk, Burcu Bakir Gungor first details domain-knowledge-based G-S-M approaches and data-driven methods—specifically based on Recursive Cluster Elimination—that they have recently developed and applied across various single and multi-omics datasets. Next, she presents extensions of their established G-S-M framework to operate on multi-layer networks, systematically identifying dysregulated modules as robust biomarker signatures. To demonstrate the practical utility of this extended methodology, she highlights their newly launched work on colorectal cancer (CRC) metagenomics data analysis. They model the CRC microbiome through a multi-layer network representing microbial species, encoded enzymes, and metabolic pathways.
By applying CROSSIM multiplex embedding and extending their G-S-M framework, they holistically analyze structural and functional dysbiosis to drive early CRC detection and personalized therapies. To develop clinically explainable, personalized probiotic and dietary recommendations, they develop LLMs with a GraphRAG architecture.
In parallel, she discusses their initiative on computational anticancer peptide prediction. They overcome traditional sequence model limitations by using a multi-layer network that captures diverse physicochemical, evolutionary, and structural peptide properties simultaneously. To efficiently analyze these complex structures and mitigate oversmoothing, they leverage Arnoldi Graph Convolutional Networks, integrating multiplex network theory with advanced deep learning to accelerate targeted drug discovery.
Ultimately, her group aims to elucidate the molecular mechanisms behind disease development using advanced machine learning. These insights can be exploited for early clinical diagnostics or as druggable targets to help physicians design personalized medical treatment plans.
More information here:
bsc.es/research-and-development/research-seminars/sorswomeninbsc-data-driven-and-domain-knowledge-based-feature-grouping-and-scoring-omics-data










