Dataset and results associated with the manuscript 'Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets'

Content of the zip files: data.zip: Datasets employed during this study: high_conf_neg.csv: Pool of human high-confidence negativesHIPPIE-current.csvintact.csvpdb_interactions.csv with sequence, species, and GO annotations pinder.csv with sequence, species, and GO annotationsstring.csv and the subsets string_900.csv, string_experimental.csv, string_database.csv, string_textmining.csvresults_esm2.zip: All results; the baseline and bias analysis results are computed from ESM-2 embeddingsresults_prott5.zip: All results but the baseline and bias analysis results are computed from ProtT5 embeddingsThe folders have the following structure:_shared: stores the embeddings{Dataset}-{Setting}: folder names with Dataset = {Hippie, Intact, PDB_dimers, Pinder, String, String900, StringDB, StringExp, StringText}, and Setting = {Leakage, regular, ILP-split, ILP-split-and-sampling, ILP-hcn}. "Leakage" is the maximally biased setting, "regular" is the similarity-reduced setting, "ILP-split" is the setting where we split the positives using the ILP but sample negatives in a degree-preserving way, "ILP-split-and-sampling" is the setting where we use the ILP both for splitting the positives and for sampling the negatives from the complement, and "ILP-hcn" is the setting where we sample the negatives from the pool of high-confidence negativesPer folder: train.csv, val.csv, test_realistic.csv, test_balanced.csv: the resulting splitsdata/sequences.fasta, data/go_annotations.tsv, data/species.tsv: the metadata fetched from UniProtsimilarities/all_vs_all.tsv: the BLAST resultssimilarities/node_mapping.tsv, similarities/partitioned_proteome.txt, similarities/similarity.graph: the KaHIP results multiqc/{bias_name}_bias.mqc.tsv: the results of the bias analysismultiqc/similarity_heatmap_mqc.html: a sanity check similarity heatmapmultiqc_*: the MultiQC data and report for the respective run(s). Also contains the results of the baseline classifier and the number of discarded PPIs. For ESM-2, we had 4 separate runs (non_struc, structural, PDB with updated data, and hcn), and for ProtT5, we ran everything together.