Fig. 2

Classifying CDRs by sequence similarity. This illustration plots in a schematic low-dimensional space the locations (dots) of CDRs from three different classes (colors). 1-nearest-neighbor classification predicts the class of one CDR from that of the most similar one; we here refine that to require the nearest neighbor to be close enough, within a specific distance threshold. Contour rings show sequence distances of 0.2, 0.3, and 0.4 from three query CDRs (“A”, “B”, and “C”) from those classes. At a threshold of 0.2, only “B” has a close-enough nearest neighbor, “b1”, which is of the same class, so 1-nearest-neighbor classification is correct. At this threshold, “A” and “C” are not predicted. When the threshold is relaxed to 0.3, “A” now has a close-enough nearest neighbor, “a1”, of the same class, so it is also correctly predicted. However, “C” has “b2”, of a different class, as its close-enough nearest neighbor, so it is incorrectly predicted. In this manner, we study trends trading off correct, incorrect, and unidentified, as the threshold is varied