Protein-Level Representations Outperform Nucleotide Foundation Models for Virulence-Factor Screening Under Organism-Level Holdout

This work presents a computational benchmark for evaluating learned models in virulence-factor screening under strict organism-level holdout. The study compares nucleotide-level DNABERT-2, protein-level ESM-2, and k-mer logistic-regression baselines on annotated coding sequences from 13 organisms, using gene-level virulence annotations and leakage-controlled cross-validation. The results show that DNABERT-2 fails to generalize across held-out organisms, performing near chance and offering no clear improvement over k-mer features, while ESM-2 achieves substantially stronger cross-organism performance. The paper also evaluates BLASTx robustness under simulated synonymous and conservative amino-acid substitutions, finding that translated homology search remains a strong baseline for many perturbation settings. Overall, the benchmark supports protein-level learned representations as a promising complementary direction for future DNA synthesis-screening research, while emphasizing that they are not a replacement for existing screening infrastructure or operational validation.