Fig. 2: Performance comparison and analysis of DMFF-DTA.

a, b Polar bar charts depicting DMFF-DTA against other SOTA models on three evaluation metrics for the Davis (a) and KIBA (b) datasets. The height of each bar signifies the metric value for the corresponding model. In each polar plot, a dashed line of the same color as the best-performing model extends at the level of that model’s metric value across the chart, thereby facilitating an intuitive visual comparison of the best model against the others. c A step line chart overlaid with a sized scatter plot showing the relationship between protein sequence length and model training runtime. The x-axis denotes the binding site length interval, while the y-axis indicates the runtime per training epoch. For each interval length, the color of the scatter points represents the final MSE on the test set, as depicted in the accompanying color bar. Additionally, the size of the scatter points indicates the GPU memory consumption during training for that interval length. d Bar charts illustrating the performance of three graph construction techniques across three evaluation metrics, emphasizing the superior precision of our AF2-based contact map generation.