How collaborativebioinformatics/cnvision works
Project type and purpose
• CNVision is a multi‐module, data‐driven machine learning pipeline for copy number variant (CNV) calling and filtering.
• It combines:
• Data preprocessing (VCF/CRAM → BAM → ground truth VCF)
• Sequence‐based classification (DNABERT-2 fine-tuning)
• Image‐based classification (CNN/ViT on CIGAR-encoded RGB images)
• Filtering & evaluation against truth sets (Truvari, benchmarking metrics)
• It is organized as standalone modules, each handling one stage of the pipeline, and designed for batch/cloud execution (e.g., on DNANexus, AWS S3).
File structure
• Top-level modules (~parallel functional stages):
• Data_Module – raw data download, genome indexing, SV/CNV calling, visualization
• LLM_Module – sequence tokenization, embedding, DNABERT-2 fine-tuning
• Image_Encoding_Module – CIGAR → 4-channel tensors → 3-channel RGB images
• Image_Embedding_Module – DINOv3 metadata prep, LoRA fine-tuning for ViTs
• Filtering_and_Evaluation_Module – Truvari benchmarking, VCF merging/refinement
• DNANexus_Module – pipeline glue & cloud orchestration scripts
• Helpers/config: LICENSE, root README.md, per-module README.md, build scripts (conda, pip requirements).
README for additional insights
• The root README contains flowcharts for each module, technology stack (Python 3.8, Conda, samtools, bcftools, R), and external data sources (AWS S3, NCBI FTP).
• It explicitly describes two parallel ML branches (sequence vs. image) that converge in the filtering stage.