MSc Thesis Defense - Computer Science: Jingfeng Guo

Please join the Department of Computer Science for the upcoming thesis defense:

Presenter: Jingfeng Guo

Thesis title: Pathway-Aware Robust Multi-Omics Survival Modeling: A LIHC-Centered Study with Cross-Cancer Controls

Abstract: Background. Multi-omics survival models may shrink lower-dimensional copy number variation (CNV) and somatic mutation (MUT) blocks to zero when combined with transcriptomic data. Such weak modality suppression does not by itself show that useful information was discarded: coefficient participation, fitted-model dependence, and incremental prediction beyond a strong clinical and RNA baseline are distinct questions.

Research Objectives. This study characterized weak-modality suppression in hepatocellular carcinoma (LIHC), breast invasive carcinoma (BRCA), and lower-grade glioma (LGG); evaluated pathway representation and grouped penalization under matched comparisons; tested conditional CNV/MUT value beyond clinical and RNA information; and assessed robustness to missing modalities.

Methods. The Cancer Genome Atlas provided RNA-seq, GISTIC2 CNV, MC3 mutation, clinical, and survival data. Complete-case cohorts comprised 346 patients for LIHC progressionfree interval (PFI), 345 for LIHC overall survival (OS), 778 each for BRCA PFI and OS, and 503 for LGG OS. RNA was mapped to MSigDB Hallmark activity by GSVA; each pathway also received seven CNV and four MUT summaries. Pathway-grouped composite MCP (cMCP) was compared with Elastic Net using identical features, patients, and five frozen outer splits.

Conditional CNV/MUT utility was tested against an enhanced clinical-plus-Hallmark-RNA baseline using direct-addition, biology-weighted, and gated BioResidual models on the same outer folds. This analysis retained continuous CNV values on their observed scale and incorporated nonredundant mutation burdens, prespecified cancer-specific drivers and molecular axes, and fold-specific CNV–RNA interactions. BioResidual used an out-of-fold baseline offset and an inner-fold exact-zero gate, allowing unsupported corrections to vanish. Evidence included 2,000 patient-level paired bootstrap resamples, a permuted-weak negative control, whole-block permutation, and factorial Shapley attribution. Matched five-member ensembles trained with and without modality dropout were evaluated under synthetic MCAR missingness and natural incompleteness. Outcomes included C-index, three-year Uno C-index, time-dependent AUC, Brier score, integrated Brier score, and calibration. For BRCA PFI, PathwayGAT-Cox was also compared with an architecture-matched no-graph control and five degree-preserving random graphs.

Results. Elastic Net suppressed at least one low-dimensional molecular block in all five cancer–endpoint combinations. Relative to same-feature Elastic Net, pathway-grouped cMCP improved mean C-index by 0.0374 for BRCA PFI and 0.0658 for BRCA OS; both paired bootstrap intervals were above zero. Uno C-index and time-dependent AUC preserved this ordering, whereas Brier score and IBS did not improve consistently. This disagreement emphasized that discrimination and absolute risk error answer different evaluation questions.

For LGG OS, BioResidual improved on the clinical plus Hallmark RNA baseline (paired ΔC = +0.0164, 95% CI [0.0051, 0.0300]); the negative control, block permutation, and supplementary metrics were directionally consistent. Weak-modality residual corrections were gated to exactly zero for every LIHC and BRCA outer split.

Under 10%–70% MCAR missingness, modality dropout increased mean C-index for LIHC PFI, LIHC OS, and LGG OS but decreased it for both BRCA endpoints; natural-incompleteness results remained endpoint-dependent. Observed natural incompleteness did not support a single cross-cancer direction.

PathwayGAT-Cox exceeded Hallmark cMCP with an interval excluding zero only for BRCA PFI (ΔC = +0.0509, 95% CI [0.0068, 0.0963]). However, the biological graph did not outperform the no-graph control (ΔC = −0.0024, 95% CI [−0.0085, 0.0032]) or the five-random-graph mean (ΔC = +0.0013, 95% CI [−0.0029, 0.0053]); the neural-model difference was therefore not attributable to biological topology. Mediation analysis identified 11 indirect associations with FDR below 0.1, without causal or biomarker implications.

Conclusion. Pathway grouping produced compact, biologically annotated models, but benefits depended on cancer, endpoint, and question. BRCA supported structured penalization over same-feature Elastic Net without conditional CNV/MUT value beyond a strong baseline; LGG showed the opposite pattern. The resulting evidence chain separates coefficient participation, fitted-model dependence, attribution, conditional increment, and missingness robustness. Accordingly, biological annotation, model dependence, and predictive improvement should not be treated as interchangeable evidence. Its main contribution is to distinguish sources of apparent predictive benefit rather than claim a universally superior model. Findings remain internally validated and do not establish clinical utility.


Committee Members:
Dr. Abedalrhman Alkhateeb (supervisor, committee chair), Dr. Saad B. Ahmed, Dr. Malek Alsmadi (Electrical Engineering)

Please contact grad.compsci@lakeheadu.ca for the Zoom link. Everyone is welcome.