Ablation study: impact of PLM embedding configurations on performance across three datasets.

The default Paraplume configuration uses all six embeddings: AbLang2, antiBERTy, IgBert, IgT5, ESM-2, and ProtT5. We report results when each embedding is removed individually or used alone. Bold indicates the best score, and underlined values represent the second-best. Based on performance across all 12 evaluation points, we chose to retain all six embeddings. While choosing different settings for each dataset could yield higher scores, we prioritize robustness and use the same configuration across all three datasets. Paraplume-S is a lightweight variant of Paraplume that uses only ESM-2 embeddings. All results are averaged over 16 random seeds to account for variability. (XLSX)