PhADS: A Bilingual Multimodal Model Based on prostT5 for Phage Anti-Defense System Annotation

PhADS is an innovative bilingual multimodal model developed by George-nsn, built on the prostT5 protein language model, specifically designed to identify and annotate phage anti-defense systems. The project was released on May 31, 2026, and its source code is hosted on GitHub (link: https://github.com/George-nsn/PhADS). PhADS addresses challenges faced by traditional bioinformatics methods in annotating phage anti-defense systems, such as data sparsity and insufficient cross-species generalization capabilities, providing new tools for virology research and biotechnological applications.

Phages are viruses that infect bacteria and play important roles in ecosystems and biotechnology. Especially with the prominent issue of antibiotic resistance, phage therapy has become a research hotspot. There is a complex 'arms race' between phages and their host bacteria: bacteria evolve defense systems to resist infection, while phages develop anti-defense systems to break through these lines. Accurately identifying anti-defense systems in phage genomes is crucial for understanding phage-host interactions, developing phage therapies, and creating synthetic biology tools. However, traditional methods face challenges such as data sparsity and insufficient cross-species generalization capabilities.

PhADS (Phage Anti-Defense System annotator) is a bilingual multimodal deep learning model specifically designed for annotating phage anti-defense systems. Its core innovation lies in combining the prostT5 protein language model with a multimodal learning framework to achieve high-precision identification and annotation. prostT5 is a protein language model based on the Transformer architecture, which can capture evolutionary information and functional patterns of sequences. PhADS optimizes this model for the characteristics of phage anti-defense systems through fine-tuning.

Bilingual Model Design

PhADS adopts a bilingual architecture that can process both protein sequence information and text annotation information, recognizing sequence features while understanding biological functions and classifications.

Multimodal Fusion

Integrate three types of biological data:

1. Sequence Modality: Processes nucleotide and protein sequences of phage genomes
2. Structural Modality: Leverages prostT5's implicit encoding capability for protein structures
3. Annotation Modality: Integrates existing functional annotations and classification information

Representation Learning Based on prostT5

prostT5 acquires evolutionary information from millions of protein sequences through self-supervised learning. PhADS starts with its pre-trained weights and transfers general protein knowledge to the anti-defense system task via transfer learning, reducing reliance on labeled data and improving generalization capabilities.

Phage Genome Annotation

Automatically annotates anti-defense systems in newly sequenced phage genomes, helping researchers quickly identify key genes and supporting phage classification, functional research, and evolutionary analysis.

Phage Therapy Development

Guides the selection and optimization of phage strains, predicts therapeutic effects and host ranges, which is crucial for the development of phage therapy.

Synthetic Biology Design

Helps design artificial phages or plasmids with specific anti-defense capabilities, applicable in fields such as gene therapy and biological control.

PhADS represents an important application direction of AI in virology research. By applying large protein language models to specific virological problems, it demonstrates the potential of deep learning in bioinformatics and provides methodological references for similar studies. Its bilingual multimodal design can be extended to tasks such as antibiotic resistance gene identification and virulence factor prediction, accelerating the digitalization of biomedical research.

Future development directions for PhADS include:

• Integrating more experimentally validated data to improve prediction reliability
• Developing interactive visualization tools to help understand model prediction results
• Expanding to research on other virus-host interactions
• Combining with laboratory automation systems to achieve a closed loop from computational prediction to experimental validation PhADS is expected to promote the transformation of phage research from an experiment-driven to a data-driven paradigm.