Abstract: Despite advances in cancer genomics, pancreatic cancer remains one of the most lethal malignancies, with most patients diagnosed at advanced, inoperable stages. This study integrated clinical and somatic mutation data from TCGA and ICGC, yielding a harmonised dataset of 796 patients (discovery: n=597; validation: n=199). Of 43,871 genes, 24 with a mutation prevalence of at least 5% were selected for analysis. Mutation impact was assessed using SIFT, PolyPhen, and VEP, and missing data were imputed. Cox proportional hazards regression and twelve machine learning (ML) algorithms were applied to predict survival at multiple time points (6, 12, 24, and 36 months), with SMOTE used to address class imbalance. TP53 and HMCN1 mutations, particularly high-impact variants, were significantly associated with survival in multivariate hazard analysis, with HMCN1 reported here for the first time as an independent prognostic factor in pancreatic cancer. The best hazard model achieved AUROC values ranging from 0.51 to 0.63, while the top ML model, Multinomial Naïve Bayes, achieved an AUROC of 0.671 and an AUPR of 0.606. Both approaches struggled at shorter and longer intervals due to limited sample sizes and data imbalance. These results indicate that mutation data alone are insufficient for reliable survival prediction in pancreatic cancer. Incorporating multi-omics profiles, treatment history, and larger patient cohorts may substantially improve model accuracy in future work.

 

Presenter: Al Imtiaz (Student ID: 0419054001)

Venue: Graduate Seminar Room