We selected 20 complexes from the Protein-Protein Docking Benchmark 5.0 dataset based on structure resolution and parameterization difficulty. For each complex, we conducted a standard 1 µs-long molecular dynamics (MD) simulation in the NPT ensemble (at 1 atm and 300 K, following a 2 ns NVT equilibration) for the bound receptor, unbound receptor, bound ligand and unbound ligand. We set up all systems using Amber ff14SB and its recommended TIP3P water model, running MD simulations with Amber 16. For the 80 (single chain structure) MD, we sampled 500 frames for each simulation and computed the average prediction confidence.