To test the performance of our model, we used the mixed human and mouse reads with different ratios. Reads are drawn from sequencing output without any selection or tuning. Reads were further split randomly into a 50% training set and a 50% testing set. We trained the same network with five training sets and tested the performance of each trained network using all five testing sets. The final performance was evaluated by accuracy after six epochs. Accuracies of model trained by either 0% or 100% human showed a strict linear relation with percentage of human reads in the testing sets (Fig. 3A). The 10% and 50% human training sets behaved with similar accuracies across all testing sets (Fig. 3A). With 50% or less human reads in the testing set, models that are trained by training sets with 10% and 50% human reads outperformed models trained by other training sets. With 90% or more human reads in the testing sets, models trained by 90% and 100% human reads performed similarly. The total accuracy positively correlated with the ratio in testing sets.
