Потенциал других коронавирусных протечек

Interpretation of misclassified instances in the training set, especially for the models that correctly classified the test set sequences, show several interesting patterns. First, proximal phylogenetic ‘near-neighbors’ of known coronaviruses are also proximal in terms of class probability. For instance, WIV16(27), which shares >96% sequence identity to SARS-CoV-1,has a class probability of 0.78 while the civet SARS examples like HC/GZ/32/03271have class probabilities of 0.89(Supplementary Data). This trend continues with late-SARS isolates such as WHU having a predicted class probability of 0.95. Of course, this relationship would be expected for the training data, but this relationship is maintained in the new SARSCoV-2-related sequences published since the beginning of the pandemic. We used the model to predict the class probabilities of these, as well as other novel coronavirus sequences published throughout2020 and2021.These results are shown in Table 4.Bat coronaviruses with proximity in sequence identity to SARS-CoV-2(28,29), such as RmYN02, RpYN06 and RaTG13, exhibit human pathogen class probabilities that are proximal to the class probability of SARS-CoV-2.
Another interesting pattern observed in the training set was a group of Bat SARS-like and MERS-like viruses that were routinely classified as human pathogens–specifically, members of Jinning mine group of viruses such as Rs4231 and Rs4874, as well as the MERS-likes NL13845  and NL140422 sampled from a cave in Guangdong (30,31). These class designations seem to be  supported by serological evidence of positivity to SARS-likes reported in the area surrounding  the Jinning cave from which these SARS-like viruses were sampled (30). Finally, human enteric  coronavirus 4408 was classified as a non-human pathogen in 35 of the 45 trained models,  including those that were 100% accurate on the test set.
The models also appear to describe a human-pathogen class definition that only includes viruses that can readily transmit between adults. There are now a series of coronaviruses that appear to have the capability to cause clinical illness in children, butthe children act as terminal hosts for the virus. This list now includes Canine Alphacoronaviruses observed in Thailand in 2007 (43)and Malaysia in 2018(33), Murine HepatitisVirus detected in SRA datasets from children with febrile illness(44), Porcine Deltacoronaviruses in children in Haiti in 2014 and 2015(45), as well as human enteric coronavirus 4408(32).