Attention: Improved performance from adding structural probing data (SHAPE) and homologous sequences.

Attention Maps were a good indicator for the location of structural contacts. High attention values (red pixels) were almost exclusively found at the “near” distance class, resulting in higher probability scores for the generator’s initial prediction for that class. Incorporation of experimental A or simulated B SHAPE data both resulted in lower false positive rates of attention placement compared to when no SHAPE data C or no SHAPE data and no homolgous sequence information d was used, with false positive rates being, 4.3%, 5.7%, 14.8% and 21.4%, respectively (the false positive rate was calculated as the number of incorrectly placed red, high attention points divided by the total number of red, high attention points, where high attention points were defined by attention scores above 0.01). (PDF)