Distinct Structural Flexibility within SARS-CoV-2 Spike Protein Reveals Potential Therapeutic Targets

Serena H. Chen; M. Todd Young; John Gounley; Christopher Stanley; Debsindhu Bhowmik

1) Data preparation: We used translation and rotation invariant input data for the DL networks. We represented each MD structure by a distance matrix using the C ↵ atoms of the protein and generated two input datasets for the CVAEs. Input 1 included the matrices of the protomer structures of SARS-CoV-2, SARS-CoV-1, MERS-CoV, and HCoV-HKU1. Input 2 included the matrices of the protomer structure and the chain A of the trimer structure of SARS-CoV-2. For Input 1, as the proteins are different in length, we first performed a multiple sequence alignment of the protomer structures of SARS-CoV-2, SARS-CoV-1, MERS-CoV, and HCoV-HKU1 by Clustal Omega [29] . Based on the alignment, we inserted gaps into the corresponding aligned residue location in the distance matrix and set the distance between gaps to be 0Å. The size of each distance matrix after the alignment was 1,342 ⇥ 1,342. To reduce the matrix size, we applied a convolution layer with padding of 2 and a 14 ⇥ 14 filter with strides of size 7 in both the x and y directions. The size of each resulting matrix became 191 ⇥ 191. Finally, after alignment and size reduction we merged a total of 20,000 distance matrices of the four S proteins. An example of a distance matrix following the alignment and size reduction is represented in Fig. S1 . For Input 2, as the proteins are of the same length, no alignment was required. The size of each matrix was 959 ⇥ 959. We again applied a convolution layer with padding of 1 and a 10 ⇥ 10 filter with strides of size 5 in both the x and y directions to reduce the matrix size. The size of each resulting matrix was also 191 ⇥ 191, and we merged a total of 10,000 distance matrices of the protomer and trimer of SARS-CoV-2 S protein.