Having tested partitioning of cyanovirin-N and impurities in 46 ATPS, the three best potential phase systems generated by the programme were then tested on the CPC.
To reduce the risks associated with both over- and under-parameterization, we tested partitioning schemes most likely to represent groupings of genes that may have evolved under similar conditions (Table 1).
In contrast, we test partitions of EFMs.
For the AA data, we tested two partitioning strategies: 1) no partitioning, and 2) partitioning by gene (78 partitions).
bFor this method the results are averages over 5 random train-test partitions.
In addition, we examined the architecture at several train-test partitions.
Since the train and test partitions in these works differ slightly from our experiments, an exact comparison is not feasible.
We report average results over 20 random train-test partitions for MSRC-9 and 5 for MSRC-21.
We follow [2], training on 60 images and testing on the rest, reporting average performances over 10 random training-test partitions.
