Fig. 5. Habitat-specific models outperform SemiBin(global) which outperforms Metabat2 across 10 environments.

We tested every model on the 10 environments (10 testing samples, no overlap with the training samples). We also trained a model from all environments (training from all samples used to generate the pre-trained model for the 10 environments). We termed this model as SemiBin(global). Shown in each cell is the number of high-quality bins obtained from the testing samples, while the color indicates the performance relative to using the model trained in the same environment (a pseudo-count of 10 was added to the raw numbers to smooth the estimates). In most environments, the pre-trained model from the same environment as the testing environment returns the best results. When transferring a model to a different environment, the model can also get good results, and in most situations, it still performs better than Metabat2. In particular, in dog gut, cat gut, and human oral, SemiBin(pretrain) performed better than Metabat2 when training from any environment.