Developing a breast cancer detection model an‎d prescribing appropriate treatment using genetic data an‎d machine learning algorithms

10/6/2026 12:00:00 AM

Breast cancer is one of the most preva‎lent types of cancer in women. Its early diagnosis an‎d treatment require advanced computational methods an‎d precise analysis of biological data. The main contribution of this dissertation lies in three innovations: first, the design of a novel hybrid algorithm BGWO_SA_Ens along with an autoencoder-based feature selec‎tion model to identify key genes; second, the enhancement of deep learning models for drug–protein interaction prediction by employing optimized hyperparameters that jointly account for both predictive accuracy an‎d uncertainty; an‎d third, the application of conformal prediction models to add a layer of uncertainty quantification, transforming model outputs into confidence intervals an‎d providing the foundations for a decision-support system for clinical experts. In the first stage, gene expression datasets were collected from reputable public repositories, an‎d a limited number of highly discriminative genes were identified through the proposed feature selec‎tion algorithms. Subsequently, biological analyses such as pathway enrichment an‎d protein-protein interaction networks were conducted to eva‎luate the biological an‎d clinical relevance of the selec‎ted genes. Related drugs were then retrieved from existing drug-gene databases, an‎d drug–target interaction prediction was performed using deep learning models. To further enhance accuracy an‎d robustness, hyperparameter tuning was carried out using multi-objective optimization algorithms NSGA-II an‎d MOPSO. Finally, conformal prediction was employed to calculate confidence intervals for the predictions. This final step, as an important innovation, enabled the presentation of results together with their quantified uncertainty, which is highly valuable in clinical practice. Experimental results demonstrated that the proposed framework effectively identifies key genes an‎d potential drugs an‎d achieves highly accurate interaction predictions with quantified uncertainty. In total, by integrating differentially expressed genes with the output of the BGWO_SA_Ens algorithm, 35 top-ranked genes were identified. Among them, TOP2A, AKR1C3, EZH2, MMP1, EDNRB, S100B, an‎d SPP1 played particularly important roles in breast cancer prediction. Enrichment analysis using GO an‎d KEGG confirmed that these genes are actively involved in critical cancer-related pathways such as AMPK, Adipocytokine, an‎d PPAR. The final predictive model, based on the selec‎ted gene set, achieved excellent performance with an F1-score of 0.984, indicating its high accuracy an‎d reliability. In summary, by emphasizing the integration of metaheuristic optimization, deep learning, an‎d uncertainty quantification, this dissertation makes an effective contribution toward the development of precision medicine in breast cancer.

سرطان پستان , كشف نشانگرهاي زيستي , يادگيري ماشين , الگوريتم‌هاي فراابتكاري تركيبي , كمي‌سازي عدم‌قطعيت , پيش‌بيني تعاملات دارو-هدف

Breast cancer , Biomarker discovery , Machine Learning , Hybrid metaheuristic algorithm , Uncertainty quantification , Drug-Target interaction prediction

Morteza Rakhshaninejad

Mohammad Fathian