Unveiling HIV-1 U Sequences: Shedding Light Through Transfer Learning on Genomic Spectrograms
摘要
The human immunodeficiency virus type 1 (HIV-1) has caused one of the most pernicious pandemics in recent history. The current difficulty with effective treatments and vaccines is largely attributed to its high mutation rate and multiple complex genetic recombination events, resulting in various subtypes. Some fragments and complete sequences do not match any subtype (referred to as U, unclassified or unclassifiable). Further understanding of these sequences and fragments may provide insights into the phylogeny of HIV-1 and its viral evolution. By applying transfer learning to a pre-trained CNN (VGG-16) using all the genomic spectrograms of all complete HIV-1 genomic sequences, we achieved a total hit rate of 98.27%, with the CNN primarily focusing on detecting the recombinant features, mainly in the 5’LTR and 3’LTR regions. We also generated a phylogenetic tree of all complete HIV-1 sequences. Subsequently, we compared subtype U interpretability patterns with phylogenetic distance, yielding valuable information not only about subtype U but also about intersubtype parentage and their common and distinct genomic components. Interpretability patterns coincided with multiple microbiological evidences in 60.870% of the sequences, were inconclusive in 34.783%, and only one controversial sequence was detected (4.348%). Compared to multiple sequence alignment algorithms, our methodology is faster, requires low computing power, and does not necessitate manual adjustments.