Fig. 5
From: Sequence-based prediction of protein binding regions and drug–target interactions
Histograms and Gumbel distributions of maximum attention on compounds for BR/non-BR. We collected the maximum attention values of compound tokens in the first transformer for BR/non-BR from each complex. We fitted the collected maximum attention values to a Gumbel distribution, showing the differences in the distribution of attention on BR and non-BR protein grids. The blue histogram plots the maximum attention values for BR, while the blue line shows the fitted Gumbel distribution. The gray histogram and line analogously represent the non-BR case. a Histograms and fitted Gumbel distributions of the COACH dataset. b Histograms and fitted Gumbel distributions of the HOLO4K dataset