Figure 1:

Scrambler Architecture and Masking Operator (a) High-level architecture. Inclusion: Maximize the entropy of the PSSM predicted by the Scrambler and minimize the prediction error of samples drawn from it. Occlusion: Minimize PSSM entropy and maximize sample prediction error. (b) The temperature-based masking operation. The Scrambler predicts sampling temperatures (‘importance scores’) for the PSSM. (c) The Inclusion and Occlusion scrambling operations. Individual nucleotides (or amino acid residues) are perturbed by raising the sampling temperature of its corresponding categorical feature distribution.