Bilkent University
Department of Computer Engineering

---

M.S.THESIS PRESENTATION

 

DOMAIN-ADAPTIVE FINE-TUNING OF EXACTCN FOR LOCUS-SPECIFIC AND SOMATIC COPY NUMBER CALLING

 

Ozan Müjde
Master Student
(Supervisor: Assoc.Prof.Ercüment Çiçek )

Computer Engineering Department
Bilkent University

Abstract: Copy number variation (CNV) underlies a range of genetic diseases and cancer. Calling CNVs from whole-exome sequencing (WES) is harder than whole-genome sequencing, because the exome is captured rather than sequenced end to end: targets are sparse and coverage across them is uneven. ExactCN, a deep learning model for exome CNV calling, predicts a continuous per-exon copy number, but it had been validated only on genome-wide germline exome data, leaving open whether it transfers to clinically significant loci with high sequence homology and structural complexity, whether it can be moved into the somatic domain, and which of its architectural choices carry its performance. This thesis takes up all three. Locus-specific fine-tuning is applied at four such loci, SMN, AMY2A, the complement C4 and the Fc-gamma receptor FCGR3, where the fine-tuned model leads every established caller on F1macro at all four. ExactCNSomatic fine- tunes the model on a 166-gene bladder cancer dataset and reaches F1macro 0.450 against 0.416 for the best standard caller, rising to 0.481 when the regression head is replaced by gradient-boosted trees on the frozen encoder. A systematic ablation of the loss, the input representation, the gene embedding and the network capacity finds no change that improves on the baseline, which identifies training- set rather than model capacity as the binding constraint.

 

DATE: September 9, Wednesday @ 11:00

Place: EA 516