ProDiPA: Offline Multi-Teacher Distillation and Pareto Alignment for Protein Generation

Abstract

Overview. This Zenodo record is the compact public-release view of the UniRef50-MultiProp (U50-MP) dataset: a homology-partitioned protein sequence collection with predicted structural confidence, solubility, and thermophilic propensity annotations. It does not rebuild, re-filter, re-cluster, re-split, or re-score any sequence. Dataset naming. UniRef50-MultiProp (U50-MP): the complete data resource.U50-MP15: 15,000 Train sequences with ESMFold, Protein-Sol, and TemBERTureCLS predictions.U50-FT25: 25,000 Train sequences with Fold and Thermo predictions; Sol covers only the nested 15K.U50-PP41: 40,998 strict Pareto preference pairs, including 32,930 development pairs and 8,068 evaluation pairs. Cohorts. train_15k_complete.tsv.gz (U50-MP15): 15,000 Train sequences with ESMFold, Protein-Sol, and TemBERTureCLS predictions.train_25k_fold_thermo.tsv.gz (U50-FT25): the nested 25,000-sequence Train cohort. All records have Fold and Thermo predictions; only the original 15k has Protein-Sol. The other 10k uses NA and has_sol=FALSE, not a low score.validation_1000_complete.tsv.gz: 1,000 independent Validation sequences with all three predictions.train.fasta.gz, validation.fasta.gz, test_h.fasta.gz: complete frozen sequence splits.split_membership.tsv.gz: split and homology-group membership for all 59,178 sequences.teacher_train_membership.tsv.gz: membership of the frozen Fold/Sol/Thermo teacher exports.pareto_preferences.tsv.gz (U50-PP41): strict, length-matched, role-capped Pareto preference dataset. It is suitable for unconditional preference analysis only because the 15k lacks verified four-superfamily condition labels. Provenance and interpretation. Sequences originate from UniRef50 release 2026_03 (CC BY 4.0 according to the frozen source manifest). The frozen split contains 47,342 Train, 5,918 Validation, and 5,918 Test-H sequences. Homology isolation means passing the executed project protocol; it is not proof against every remote homolog or pretrained-model overlap. All property values are computational predictions, not experimental measurements. Mean pLDDT is structural confidence, Protein-Sol is a predicted solubility score, and TemBERTureCLS is a thermophilicity classifier score rather than measured melting temperature. Test-H has no property labels in this release and was not used for teacher selection. Teacher-selection thresholds (80 / 0.60 / 0.40) are not applied to the general 15k/25k release tables. Use teacher_train_membership.tsv.gz for the actual frozen teacher subsets. Missing values are serialized as NA and must not be imputed as zero. See DATA_DICTIONARY.tsv and RELEASE_SUMMARY.tsv. License. This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).