ppi_c3 v2: Code, Results, and Figures for "Reducing the Cold-Start Generalization Gap in Sequence-Based PPI Prediction"

This Zenodo deposit accompanies the manuscript "Reducing the Cold-StartGeneralization Gap in Sequence-Based Protein-Protein Interaction Prediction:A Multi-Objective Search over Frozen Protein-Language-Model Embeddings" byMuhammed Tekin and Murat Gök, submitted to IEEE/ACM Transactions onComputational Biology and Bioinformatics (2026). Contents:- ppi_c3_v2_patch.zip: code-level patch over v1 fixing three audit-discovered defects in the curriculum implementation (the beta-gene algebraic identity, the unused hardness quantile, and the hard-coded logistic-gate shape), including six new regression tests and three replication scripts.- results/: ten JSON files containing the multi-seed (n=15) evaluation outputs that reproduce Tables I-VI of the manuscript. Includes the baselines-multiseed comparison (Ours vs B1/B2/B3), the six-configuration loss-term ablation (L0-L5), the curriculum v2 validation, the NSGA-III Pareto front (269 chromosomes), per-chromosome multi-seed re-evaluation, and pretrain/finetune budget sweeps.- figures/: the six vector-PDF figures from the paper.- README.md: detailed reproduction instructions, file map, and citation. Key findings:- The searched configuration (Ours, rank 5) reduces the in-domain-to-cold C3 generalization gap by 94% relative to a manuscript-style baseline (paired t = 6.75, p ~ 1e-5, n = 15).- A loss-term ablation reveals that pretraining with InfoNCE over the MoCo queue alone (L5, without supervised contrastive or batch-hard triplet terms) reaches C3 AUPR = 0.6320 +/- 0.0119 and wins 15/15 paired seeds against every other configuration tested (paired t = 13-20).- An audit-driven repair of the v1 curriculum produces a clean empirical null on this benchmark (delta C3 = +0.0008, p = 0.67, 8/15 wins). The pipeline runs end-to-end on a single NVIDIA A100 80 GB GPU inapproximately 12 hours. All scripts pass a 318-test unit-test suite(312 original + 6 audit-regression tests).