FusioBERT: A Task-Adapted Peptide Language Modeling Framework for Anti- Fusarium Peptide Discovery

Fusarium pathogens cause substantial crop losses, motivating targeted antifungal peptide discovery. We developed FusioBERT, a 7.52-million-parameter peptide language model incorporating fixed sinusoidal positional encoding, learnable relative position bias, RMSNorm, and SwiGLU. It was pretrained on 19.18 million short sequences and 12,616 candidates mined from 232 Trichoderma genomes. The curated benchmark comprised 210 labeled peptides; evaluation used five balanced 178-sequence subsets with three-seed 5-fold cross-validation. FusioBERT achieved an MCC of 0.948 ± 0.058 and generated continuously differentiated genome-scale rankings, whereas physicochemical classifiers exhibited score saturation. Across nine external peptide tasks, FusioBERT obtained a mean MCC of 0.685, compared with 0.666 for PepBERT-large and 0.654 for ESM2. Fixed-seed residue-order shuffling reduced MCC from 0.947 to 0.874, providing preliminary evidence of sequence-order sensitivity beyond bulk composition. Among the three evaluated prioritization strategies, FusioBERT-only ranking yielded the most favorable descriptive wet-laboratory outcomes, including the strongest broad-spectrum candidates, SR1403–2 and SR1403–3. Their Hill-logistic EC50 values ranged from 1.62 to 7.07 μM; observed MIC end points were 8–32 μM. Physicochemical fusion, docking, and molecular dynamics were treated as auxiliary analyses. FusioBERT supports model-first prioritization but does not directly establish activity probabilities, candidate safety, or universal homology-independent generalization.