Third: Prediction of key mutations. When SARITA is trained on sequences encompassing the first part of the pandemic (Alpha, Beta, and Gamma variants), it generates the key mutations characterizing the Spike protein evolution (including, Delta, Omicron and Iota).

5/6

Furthermore, other approaches are more likely to overmutate the generated sequences, introducing up to hundreds of mutations that are not matched by best alignments in the test set.

4/6

Second: Viral evolution prediction. When trained on sequences collected until Feb 2021, SARITA outperforms all other approaches in predicting test set sequences (March 2021 onward). We consider the test set best match for each generated sequence via Levenshtein distance (LD).

3/6

We compare SARITA against other approaches. First, quality: Do the generated sequences have the right length? Do they contain uncharacterized amino acids? Are they close the original Wuhan strain S1, i.e., are the models generating sequences that resemble the S1 subunit?

2/6

Ever thought about running #metagenomics on your phone with ONT MinION? I presented at #AMIA OCTOPUS: Disk-based, Multiplatform, Mobile-friendly Metagenomics Classifier