Evaluate state-of-the-art PPI predictors on de-biased splits

Determine how state-of-the-art deep-learning-based protein–protein interaction prediction models perform on the de-biased data splits produced by the similarity-aware splitting and bias-minimizing negative-sampling pipeline, and establish whether they significantly outperform the Random Forest baseline evaluated in the study.

Background

The study focuses on identifying and mitigating dataset shortcuts rather than benchmarking advanced protein–protein interaction predictors. Its Random Forest baseline, trained on pooled protein-language-model embeddings, is designed as a low-complexity diagnostic and generally falls to chance-level performance after the proposed splitting and negative-sampling procedures, except where residual self-interaction or functional-relatedness biases remain.

The unresolved follow-up problem is therefore to assess modern deep-learning PPI models under these more rigorously de-biased conditions. Such evaluation would establish whether their reported predictive ability reflects genuine interaction-related signal and whether they offer a meaningful improvement over the diagnostic Random Forest baseline.

References

Natural follow-up question are how state-of-the-art deep learning-based PPI prediction models perform on these de-biased data splits and if they significantly outperform the Random Forest baseline utilized here. We leave such questions to future work.

— Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets  (2609.10193 - Bernett et al., 9 Sep 2026) in Discussion, paragraph beginning “An important scope restriction is that this work is not a benchmarking paper.”