New Deep Learning Model AVP-Fusion Achieves 95% Accuracy in Antiviral Peptide Prediction

Global: AVP-Fusion Deep Learning Framework Boosts Antiviral Peptide Prediction Accuracy

A new two-stage deep learning framework called AVP-Fusion was introduced in a December 2025 preprint on arXiv, aiming to improve the identification of antiviral peptides (AVPs) for drug discovery. The authors designed the system to capture both local sequence motifs and long-range dependencies while addressing ambiguous, hard-to-classify samples.

Adaptive Feature Fusion

The first stage constructs a panoramic feature space by integrating ten distinct sequence descriptors. An Adaptive Gating Mechanism dynamically weights contributions from convolutional neural networks (CNNs) that extract local motifs and bidirectional LSTM networks (BiLSTMs) that model global dependencies, allowing the model to adjust its focus based on the specific context of each peptide sequence.

Contrastive Learning and Data Augmentation

To sharpen decision boundaries, the framework incorporates contrastive learning driven by Online Hard Example Mining (OHEM). Additionally, the authors augment training data using BLOSUM62-based substitutions, which introduces realistic sequence variations and helps the model generalize across diverse viral families.

Benchmark Performance

When evaluated on the benchmark Set 1 dataset, AVP-Fusion achieved an accuracy of 0.9531 and a Matthews correlation coefficient (MCC) of 0.9064, surpassing previously reported state‑of‑the‑art methods. These metrics indicate a substantial improvement in both predictive power and robustness.

Subclass Prediction via Transfer Learning

In a second stage, the pretrained model is fine‑tuned to predict AVP subclasses for six viral families and eight individual viruses. Transfer learning enables accurate subclass identification even when training samples are scarce, demonstrating the framework’s flexibility for targeted antiviral screening.

Interpretability and Applications

The adaptive gating provides insight into which sequence features drive predictions, offering a degree of interpretability uncommon in deep‑learning‑based bioinformatics tools. The authors suggest that AVP-Fusion could accelerate high‑throughput screening pipelines for antiviral drug candidates.

Future Outlook

Further validation on larger, experimentally curated datasets and integration with structural modeling are proposed as next steps to confirm the model’s utility across a broader spectrum of viruses.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.