Ivanovic Stefan, El-Kebir Mohammed
Department of Computer Science, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA.
Cancer Center Illinois, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA.
Genome Biol. 2025 Apr 7;26(1):87. doi: 10.1186/s13059-025-03553-2.
Low-pass single-cell DNA sequencing technologies and algorithmic advancements have enabled haplotype-specific copy number calling on thousands of cells within tumors. However, measurement uncertainty may result in spurious CNAs inconsistent with realistic evolutionary constraints. We introduce evolution-aware copy number calling via deep reinforcement learning (CNRein). Our simulations demonstrate CNRein infers more accurate copy-number profiles and better recapitulates ground truth clonal structure than existing methods. On sequencing data of breast and ovarian cancer, CNRein produces more parsimonious solutions than existing methods while maintaining agreement with single-nucleotide variants. Additionally, CNRein shows consistency on a breast cancer patient sequenced with distinct low-pass technologies.
低通量单细胞DNA测序技术和算法的进步，使得在肿瘤内数千个细胞上进行单倍型特异性拷贝数检测成为可能。然而，测量的不确定性可能导致与实际进化约束不一致的假拷贝数改变（CNA）。我们通过深度强化学习（CNRein）引入了进化感知拷贝数检测方法。我们的模拟表明，与现有方法相比，CNRein能推断出更准确的拷贝数图谱，更好地重现真实的克隆结构。在乳腺癌和卵巢癌的测序数据上，CNRein比现有方法产生更简洁的解决方案，同时与单核苷酸变异保持一致。此外，CNRein在用不同低通量技术测序的乳腺癌患者样本中表现出一致性。
