Mixing Long and Short HIV Sequences Can Sharpen Transmission Cluster Detection, Study Finds

A public health lab that wants to know who is passing HIV to whom usually starts with a short stretch of the virus's genetic code, the pol region, because it is already being sequenced for drug resistance testing in routine clinical care. That convenience has made partial pol sequences the standard raw material for inferring molecular clusters, the groups of infections that appear to be linked. A study published November 25, 2025 in PLOS Computational Biology asks what happens when some of those short sequences are paired with longer ones.

The authors are Guang A, Dunn CW, Novitsky V, Howison M and Kantor R. Their starting point is that near full-length (nFL) HIV-1 genomes, made increasingly available by long-read and next-generation sequencing, have been shown to give more accurate phylogenetic trees and clusters than short pol sequences. But long sequences are hard to get for everyone. Some people diagnosed with HIV are virally suppressed, have moved to another clinic, or do not want to give samples, and resources are limited. Combining sequences of different lengths in one alignment also presents technical problems, so analysts often trim everything down to the shortest length.

The team proposes what they call a T-shaped alignment. If sequences are sorted by completeness, the nFL genomes form the top bar of the T and the pol-only sequences form the stem. Every pol-only sequence is treated as having missing characters for the rest of the genome, while the non-pol data from the long sequences are still used. Standard phylogenetic software can read such an alignment, so no new programs are needed, and sequences of any length covering overlapping parts of the genome could in principle be included. The authors say they know of no earlier study of such mixed HIV-1 alignments.

To test it, they downloaded a filtered alignment of 1196 subtype B nFL sequences published through the end of 2018 from the Los Alamos National Laboratory HIV database. Four were hypermutated, which the authors judged too few to matter, so all were kept. They also built a sensitivity dataset from 116 subtype B genomes from a single study, chosen because it was the largest from one study and included geographic linkage information.

For each dataset they made 100 samples by shuffling the order of sequences. From each sample they made 13 alignments by masking everything outside the protease-reverse transcriptase region with gaps, using HXB2 genome coordinates. Eleven mixtures ran in 10% steps of nFL sequences, plus 95% and 99% mixtures, producing 13,000 alignments in all. Trees were built with IQTree. The working assumption was that an alignment of 100% long sequences gives the most accurate clustering, and the mixtures were measured against it.

In the full dataset, adding long sequences did not help steadily. Below 50% nFL, accuracy did not improve, and the authors say the approach appeared more misleading at low mixture proportions. Above 50%, results moved gradually closer to the all-long benchmark, with more and larger clusters identified. Stricter bootstrap thresholds narrowed the accuracy gaps but also reduced the number of clusters found and their mean size. In the single-study subset, accuracy improved either after a 30% threshold or immediately, depending on the bootstrap choice.

The authors conclude that the needed proportion of long sequences depends on the analysis goal and the dataset. They stress the findings are preliminary and should be validated in regional HIV sequence datasets before they inform public health work. The study was funded by NIH grants to Kantor and an NIGMS Institutional Development Award, and the authors declare no competing interests. The code is public on GitHub.