1-Defending Against Genome Reconstruction Attacks in Genomic Data Sharing Beacons
Amir Mohammad Erfan
Master Student
(Supervisor: Assoc. Prof. Ercüment Çiçek)
Computer Engineering Department
Bilkent University
Abstract: The GA4GH Beacon protocol answers allele queries with summary statistics, allowing genomic cohorts to be searched without releasing individual genotypes. However, recent literature has demonstrated that these summaries Genome Reconstruction Attacks (GRA). By exploiting public linkage disequilibrium (LD) data, an adversary can accurately reconstruct the beacon’s participants. Traditional defenses, including response randomization and heuristic-based methods, offer inadequate protection: uniform randomization can compromise the clinical utility of rare variants (minor allele frequency ≤ 0.05), while heuristic rules provide suboptimal protection against complex correlation-based attacks. In this paper, we introduce the first defense mechanism specifically designed to mitigate GRA in beacons. We propose a query-adaptive system driven by a continuous-action Reinforcement Learning (RL) agent utilizing the Twin Delayed Deep Deterministico rithm. Our system continuously monitors the query history to distinguish between an honest researcher and an attacker. Based o level, the RL agent dynamically determines the precise magnitude of noise to inject into the released frequencies, learning an allocation strategy that empirically outperforms heuristics. We evaluated our defense on simulated beacons constructed from the 1000 Genomes Project across admixed American super-populations. In the highly correlated Human Leukocyte Antigen (HLA) region, our system reduces an attacker’s om 0.795 to 0.602 (a 24.3% relative drop) on our primary European benchmark while preserving near-perfect utility for honest users with clinically relevant query sequences, maintaining a pristine mean absolute error (MAE) of 0.004 on the European benchmark. In contrast, a native random-answering baseline (p = 0.5) only reduces the attacker’s score to 0.664 while imposing a MAE of 0.157 for honest users, effectively destroying the accuracy of rare disease queries. Furthermore, our system degrades reconstructions across all evglobal populations, achieving up to a 42.8% relative drop on high average global MAE of just 0.016.
DATE: November 16, Monday @ 14:40 Place: EA 502
2-Reference-free and robust copy number inference in single-cell tumor transcriptomes
Mahmut Mert Gençtürk
Master Student
(Supervisor: Assoc. Prof. Ercüment Çiçek)
Computer Engineering Department
Bilkent University
Abstract: Copy number changes are a hallmark of cancer and reveal which cells are malignant and how tumors evolve. Single-cell RNA sequencing is widely used to them in individual tumor cells. Existing methods do so by normalieference of normal cells. As a result, their output depends on which cells are analyzed together, they often label normal cells uggle in samples without normal cells. Here we present REFCON, a deep-learning model that infers copy number from each cell’s own rence. We evaluated REFCON on more than 410,000 held-out cells from cell lines, organoids and 112 tumors of nine cancer types, profiled with seven sequencing technologies. It agreed more closely with single-cell DNA than existing methods, was unaffected by sample composition, and labeled far fewer normal cells as aneuploid. In patient-derived organoids, which contain no normal cells, it gave consistent profiles across passages and tracked clonal evolution. REFCON addresses long-standing limitations of reference-based inference and enables robust, reference-free copy number calling tomes.
DATE: November 16, Monday @ 15:00 Place: EA 502