Data-driven recombination detection in viral genomes

Abstract

Recombination is a key molecular mechanism for the evolution and adaptation of viruses. The first recombinant SARS-CoV-2 genomes were recognized in 2021; as of today, more than ninety SARS-CoV-2 lineages are designated as recombinant. In the wake of the COVID-19 pandemic, several methods for detecting recombination in SARS-CoV-2 have been proposed; however, none could faithfully confirm manual analyses by experts in the field. We hereby present RecombinHunt, an original data-driven method for the identification of recombinant genomes, capable of recognizing recombinant SARS-CoV-2 genomes (or lineages) with one or two breakpoints with high accuracy and within reduced turn-around times. ReconbinHunt shows high specificity and sensitivity, compares favorably with other state-of-the-art methods, and faithfully confirms manual analyses by experts. RecombinHunt identifies recombinant viral genomes from the recent monkeypox epidemic in high concordance with manually curated analyses by experts, suggesting that our approach is robust and can be applied to any epidemic/pandemic virus.

Fig. 1. RecombinHunt has three possible outcomes: no recombination, 1 breakpoint recombination, or 2 breakpoints recombination.

a Example of likelihood ratio profile for a non-recombined genome with N1 = 66 mutations, only featuring one donor lineage corresponding to BA.2.3.13. b Example of likelihood ratio profile for a recombined genome assigned to XBE Pango lineage with N2 = 72 mutations, breakpoint at the 63rd mutation, donor lineage BA.5.2.6 (from 5'-end to 63rd mutation), and acceptor lineage BE.4 (from 64th mutation to 3'-end). c Example of likelihood ratio profile for a recombined genome assigned to XD Pango lineage with N3 = 67 mutations, two breakpoints at the 23rd and 52nd mutations, donor lineage AY.4 (from 5'-end to 23rd mutation and from 53rd mutation to 3'-end), and acceptor lineage BA.1.22 (from 24th to 52nd mutation).

Fig. 2. Overall RecombinHunt workflow.

An input viral sequence is considered. The donor lineage is searched based on the cumulative likelihood ratio. Then, three branches are considered: non-recombinant model, one-breakpoint recombination model, and two-breakpoint recombination model. The preferred model is chosen using statistical testing, based on the Akaike information criterion.

Fig. 3. RecombinHunt recognizes recombination events in one and two-breakpoint cases.

a Search result on 75% consensus-genomes of 63 high-quality sequences assigned to XBE Pango lineage. RecombinHunt selects BA.5.2.6 (child of BA.5.2, see ground truth) as L1 candidate, starting from the 5'-end of the genome; maximum likelihood ratio (LR) 61.025 is reached at mutation 63 (max-L1). Then, RecombinHunt selects BE.4 as L2 candidate in positions (64-72). BA.5.2.6 (left table) and BE.4 (right table) are compared with the following candidate lineages, ranked by their maximum likelihood ratios. Tables report the number of sequences, breakpoint, maximum likelihood ratio. The next columns illustrate the comparison of the candidate with the first of the table: value of one-sided AIC comparison between recombination model and non-recombination model (lower values are when row candidate is similar to the first candidate); p-value of AIC -- without multiple comparison corrections; and three conditions: (C1) marked if p-value is ≥10⁻⁵; (C2) marked if row breakpoint is at most one mutation apart from the one of the first candidate; (C3) marked if candidate belongs to the same phylogenetic branch as the first one. Candidates with three marks are incorporated into groups, resulting in BA.5.2.6 and CP.3 (candidates for L1) and BE.4, CQ.2, BE.4.1.1, BE.4.1, CQ.1.1, CQ.1 (for L2). b Search result on 75% consensus-genome of 14 high-quality sequences assigned to XD Pango lineage. For the 5'-end portion of the genome, RecombinHunt selects AY.4 as L1 candidate (child of B.1.617.2 in the ground truth), with maximum likelihood ratio at the 23rd mutation; AY.4 is also selected for the 3'-end portion of the genome with maximum likelihood ratio at the 53rd position. Then, RecombinHunt identifies the first L2 candidate BA.1.22 (consistent with the ground truth BA.1*) in the 24-52 mutation interval. AY.4 (left table) and BA.1.22 (right table) are compared with the following candidates; no one conjunctively meets the three conditions.

Fig. 4. Barplots of RecombinHunt (RH) outputs for breakpoints positions.

Label colors reflect the four groups in Table 3. On the x-axis, the consensus-genome mutations; on the y-axis, the count of single sequences with a breakpoint detected on the x position. A light blue stripe indicates the RH 1BP position; two light orange stripes indicate the RH 2BP positions (see Table 2). Blue bar plots count the sequences whose 1BP is located at a given mutation; orange bar plots count the sequences whose 2BPs are located at given positions (without distinction between the first and the second one).