Confidence estimation relies on more than just the predicted interaction probability from AI models. We also incorporate experimental evidence from multiple sources (UniProt, BioGRID, and STRING) to refine confidence assessments. Because experimental databases differ in accuracy, we benchmarked them to evaluate their reliability (see Figure 3C and 3D in the manuscript). Based on these benchmarks, when high-quality experimental evidence is available as prior information, we can apply a lower threshold to AI-derived scores while still classifying interactions as confident.