Fig. 2

The system diagram of MetaBinG2. MetaBinG2 first loads the reference database and copy it into GPUs as a reference matrix. Next, the short query sequences are converted into k-mer vectors in CPUs, and vectors will be loaded to GPUs as query matrix. Then, the query matrix will be multiplied to the reference matrix in GPUs by CUDA CUBLAS functions and adjusted with the weights, with a similarity score matrix as the output. The source genomes with minimum similarity scores will be selected. The weights are updated according to the latest proportions after all sequences are classified. If the BC distances of the weights before and after the update are less than the cutoff, the final similarity scores together with the annotated taxonomy information will be output