Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
LYCEUM: LEARNING TO CALL COPY NUMBER VARIANTS ON LOW-COVERAGE ANCIENT GENOMES
Ahmet Arda Ceylan
Master Student
(Supervisor:Assoc.Prof.Dr.Ercüment Çiçek)
Computer Engineering Department
Bilkent University
Abstract: Copy number variants (CNVs) are key drivers of phenotypic variation and disease susceptibility, yet their analysis in ancient DNA (aDNA) remains constrained compared to single nucleotide polymorphism studies. Detecting CNVs in aDNA is inherently challenging because samples are degraded, contaminated, and typically sequenced at low coverage, which causes conventional read-depth-based callers to underperform. This thesis introduces LYCEUM, the first machine learning-based CNV caller optimized specifically for low-coverage aDNA. LYCEUM utilizes convolutional and transformer encoder blocks paired with chromosome-specific classification tokens to categorize exon-level deletions, duplications, and no-call events from read depth signals. A two-stage training strategy first pre-trains the model on 550 whole-genome sequencing samples from the 1000 Genomes Project using DRAGEN-based labels, and subsequently fine-tunes it on high-coverage aDNA samples alongside simulated ancient genomes using down-sampled read depth as input. On real ancient samples down-sampled to 0.05× coverage, LYCEUM im- proves gene-level F1 scores by 20% for deletions and 39% for duplications over the next-best method. On simulated ancient data at 0.05×, these improvements reach 60% and 68%, respectively. Segmental deletion calls effectively recover geographic population structure and exhibit allele frequency spectra consistent with negative selection. These results demonstrate that transfer learning enables accurate CNV calling in noisy, ultra-low-coverage ancient genomes, thereby supporting downstream evolutionary and paleogenomic analyses
DATE: July 13, Monday @ 11:00
Place: EA 516