LucaVirusCress

LucaVirusCress: A Deep Learning Framework for CRESS Virus Identification

LucaVirusCress is a specialized binary classification model for identifying CRESS (Circular Rep-Encoding Single-Stranded) DNA viruses by targeting REP (Replication-associated protein). Built on LucaProt, it supports high-throughput viral sequence screening in large metagenomic datasets.

Contents

• Project Overview
• Training Performance
• Dataset Details
• Technical Specifications
• Usage
• Model Weights
• License

Project Overview

• Task: Binary classification (CRESS REP vs. non-CRESS REP)
• Target: CRESS virus REP proteins
• Architecture: SequenceAndStructureFusionNetwork (SEFN), a 4-layer Transformer-based fusion network
• Application: Discovery of novel CRESS viruses from environmental and clinical omics data

Training Performance

The model reached optimal convergence at 4,000 global steps.

The training process used weighted loss to handle extreme class imbalance.

Dataset Details

To ensure high specificity, the model was trained on a highly imbalanced dataset (1:40 ratio) with carefully selected hard negatives.

• Positive samples: Verified CRESS REP protein sequences
• Negative samples:
  1. Non-CRESS proteins with high sequence similarity to REP
  2. Proteins with structural motifs similar to the target class
  3. Non-target proteins in the CRESS viral genome (for example, Cap proteins)

Technical Specifications

Usage

1. Environment Setup

git clone https://github.com/lpyhhh/LucaVirusCress.git
cd LucaVirusCress

conda env create -f environment.yaml
conda activate lucaviruscress

# Download model files from Hugging Face
git clone https://huggingface.co/Daxiao123/LucaVirusCress hf_model

# Copy model files into checkpoint directory
mkdir -p models/cress/protein/binary_class/sefn/20230201140320/checkpoint-4000
cp -r hf_model/* models/cress/protein/binary_class/sefn/20230201140320/checkpoint-4000/

# Clean temporary directory
rm -rf hf_model

2. Single Sequence Prediction

The embedding matrix is generated in real time.

cd src

python predict_one_sample.py \
  --protein_id protein_1 \
  --sequence "MTTSTAFT...(your sequence)" \
  --emb_dir ../emb \
  --dataset_name cress \
  --dataset_type protein \
  --task_type binary_class \
  --model_type sefn \
  --time_str 20260321140320 \
  --step 4000 \
  --threshold 0.5 \
  --gpu_id 0

3. Batch Prediction (FASTA)

For large-scale screening of metagenomic assemblies:

cd src

python predict_many_samples.py \
  --fasta_file ../data/input.fasta \
  --save_file ../result/prediction_results.csv \
  --emb_dir ../emb \
  --dataset_name cress \
  --model_type sefn \
  --time_str 20260321140320 \
  --step 4000 \
  --gpu_id 0

Key Parameters

• --truncation_seq_length: Maximum sequence length (default: 4096)
• --pos_weight: Handled internally (configured as 40 in config.json for imbalance)
• --threshold: Classification threshold (default: 0.5)

Model Weights

Pre-trained weights and configuration are hosted on Hugging Face:

https://huggingface.co/Daxiao123/LucaVirusCress

License

This project is licensed under the MIT License. See LICENSE for details.

Acknowledgment

Thanks to everyone who supported this project.