CNAsim: improved simulation of single-cell copy number profiles and DNA-seq data from tumors

Abstract

Summary CNAsim is a software package for improved simulation of single-cell copy number alteration (CNA) data from tumors. CNAsim can be used to efficiently generate single-cell copy number profiles for thousands of simulated tumor cells under a more realistic error model and a broader range of possible copy number alteration mechanisms compared to existing simulators. The error model implemented in CNAsim accounts for the specific biases of single-cell sequencing that lead to read count fluctuation and poor resolution of CNA detection. For improved realism over existing simulators, CNAsim can (i) generate WGD, whole-chromosomal CNAs, and chromsome-arm CNAs, (ii) simulate subclonal population structure defined by the accumulation of chromosomal CNAs, and (iii) dilute the sampled cell population with both normal diploid cells and pseudo-diploid cells. The software can also generate DNA-seq data for sampled cells. Availability CNAsim is written in Python and is freely available open-source from https://github.com/samsonweiner/CNAsim. Supplementary information Supplementary data are available at Bioinformatics online.

Assessing the merits: an opinion on the effectiveness of simulation techniques in tumor subclonal reconstruction

“…For the former, a list of germline and/or somatic alterations was bioinformatically spiked-in to the reference sequence. The reference sequence was then used to generate reads with a dedicated package, such as ART or Wessim ( Huang et al 2012 , Kim et al 2013 , Garvin et al 2015 , Chu et al 2017 , Xia et al 2017 , Yang et al 2019 , Cmero et al 2020 , Ricketts et al 2020 , Zaccaria and Raphael 2020 , Mallory and Nakhleh 2022 ; Weiner and Bansal 2023 ), or was iteratively fragmented ( Srivatsa et al 2023 ). If a germline BAM file was the starting point, somatic alterations were simulated by editing the reads ( Salcedo et al 2020 ).…”

“…Simulation methods began with a model of an evolutionary process, which included a tree topology and a distribution of somatic alterations on the nodes or edges of the tree. For methods that assumed a coalescent process ( Yang et al 2019 , Posada 2020 , Srivatsa et al 2023 , Weiner and Bansal 2023 ), the topology was generated first, followed by the distribution of alterations. Alternatively, if a branching process was assumed ( Popic et al 2015 , Qin et al 2015 , Xia et al 2017 , Ricketts et al 2020 , Andersson et al 2021 , Kang et al 2022 ), the topology was generated in parallel with the alterations.…”

“…Three of the simulation methods generated single-cell reads but did not incorporate any of these errors ( Yang et al 2019 , Srivatsa et al 2023 ). Two of the methods that generated reads incorporated genome non-uniformity ( Mallory and Nakhleh 2022 , Weiner and Bansal 2023 ). Many methods incorporated multiple sources of error, but they generated cell-by-mutation matrix profiles rather than actual sequence data ( Table 1 , Supplementary Table S1 ).…”

“…We identified two major categories of simulators: (1) self-contained simulators that were designed for general use and were the subject of an independent publication; (2) custom simulators that were developed to benchmark a single group's reconstruction algorithm. We evaluated nine self-contained simulators: BAMSurgeon, MosaicSim, PSiTE, Pysubsim-tree, SCNVsim, Pysim-sv, CellCoal, SimSCSnTree, and CNAsim ( Qin et al 2015 , Chu et al 2017 , Xia et al 2017 , Yang et al 2019 , Posada 2020 , Salcedo et al 2020 , Mallory and Nakhleh 2022 , Srivatsa et al 2023 , Weiner and Bansal 2023 ) ( Table 1 ). Surprisingly, self-contained simulators were used for benchmarking by only two out of 42 published subclonal reconstruction algorithms ( Supplementary Table S1 ).…”