Precise detection of Acrs in prokaryotes using only six features

Abstract

Anti-CRISPR proteins (Acrs) can suppress the activity of CRISPR-Cas systems. Some viruses depend on Acrs to expand their genetic materials into the host genome which can promote species diversity. Therefore, the identification and determination of Acrs are of vital importance. In this work we developed a random forest tree-based tool, AcrDetector, to identify Acrs in the whole genomescale using merely six features. AcrDetector can achieve a mean accuracy of 99.65%, a mean recall of 75.84%, a mean precision of 99.24% and a mean F1 score of 85.97%; in multi-round, 5-fold crossvalidation (30 different random states). To demonstrate that AcrDetector can identify real Acrs precisely at the whole genome-scale we performed a cross-species validation which resulted in 71.43% of real Acrs being ranked in the top 10. We applied AcrDetector to detect Acrs in the latest data. It can accurately identify 3 Acrs, which have previously been verified experimentally. A standalone version of AcrDetector is available at https://github.com/RiversDong/AcrDetector. Additionally, our result showed that most of the Acrs are transferred into their host genomes in a recent stage rather than early.

Anti-CRISPRdb v2.2: an online repository of anti-CRISPR proteins including information on inhibitory mechanisms, activities and neighbors of curated anti-CRISPR proteins

“…Approximately 86.9% (272/313) of the 313 Acrs simultaneously predicted by the three methods were annotated as a ‘hypothetical protein’ or ‘Uncharacterized protein’ according to NCBI annotation. The proteins of unknown function were employed to predict Acrs ( 10 , 41 ), because if a protein has a validated functional annotation, it is less likely to perform an Acr function.…”

AcrNET: Predicting Anti-CRISPR with Deep Learning

“…The mathematical expressions for calculating these metrics are provided in Appendix C. Since other methods require additional information other than the sequence alone, such as gene location on chromosome [Dong et al, 2020], we compare DeepAcr with three recently proposed methods, namely PaCRISPR [Wang et al, 2020], AcRanker [Eitzinger et al, 2020], and the one from Gussow et al [2020]. All of them have shown strong Acr predicting capacities and efficiency.…”

“…By using the criteria for selecting the negative sample proteins as introduced in Wang et al [2020], we obtain 902 negative samples from the negative training dataset and 260 negative samples from the negative testing dataset for cross-dataset test. The 1094 non-redundant positive samples from Anti-CRISPRdb, and the 902 negative samples from the negative training dataset, and 260 negative samples from the cross-dataset dataset of Wang et al [2020], are combined as the whole experiments dataset, where the sequence similarity between positive samples and negative samples is also less than 40%. In the cross-dataset test, the Acrs of types I-F, II-C, and I-D are selected as positive test samples, and the resting Acrs are used as positive training samples.…”

“…We also report TP (true positive), FN (false negative), FP (false positive), TP (true positive) and specificity for ambiguous performance for further analysis. Since other methods require additional information other than the sequence alone, such as gene location on chromosome [10], we compare AcrNET with three recently proposed methods with the same dataset we collected, namely PaCRISPR [40], AcRanker [12], and DeepAcr [38]. All of them have shown strong Acr predicting capacities and efficiency.…”

“…A machine learning approach consisting of a random forest with 1000 trees was built to predict comprehensive Acrs, which showed strong forecasting capability for the unseen test set [Gussow et al, 2020]. Also built upon the random forest algorithm, AcrDetector utilized merely six features to identify Acrs from the whole genome scale but still maintained the prediction precision [Dong et al, 2020]. An integrative database named AcrHub was developed by incorporating three state-of-the-art predictors, including homology network analysis, phylogenetic analysis, and similarity analysis [Wang et al, 2021].…”