Fig 3. Analysis path.

We randomly selected 26 out of 27 positive samples, trained the classifier on these 26 positive samples, and then tested whether it can select between the one remaining positive sample versus one randomly selected negative sample. Chance performance at this task is 50%. We repeated this process 500 times, each time testing a different sample, in order to obtain a true-positive rate. Finally, because no selection of 500 is privileged, we repeated this entire process 100 times in order to obtain a distribution of possible true-positive rates.