AI_WinterSchool_2026

Workshops for Day 3 of the 2026 Winter School for AI in Predictive Agriculture @ UQ, Queensland AU. Using DNA foundation models to inform SNP selection for genomic prediction tasks.

Workshop Files

• data/panicle_length_genomatrix_snps.vcf: A sorghum genotype matrix of 1000 SNPs across 527 accessions. SNPs match Chr9_random1000 SNPs.
• data/pheno_PL_filtered.tsv: Phenotype data for panicle length in sorghum across 527 accessions.
• data/DS24-Chr9_random1000_maf05.tsv: 1000 randomly selected SNPs from Sorghum bicolor chromosome 9.
• data/sap_flowering.pheno: Tab-separated values file containing phenotype data ("days to flowering") for the Sorghum Association Panel (SAP) population (Chromosome 1).
• data/snp_matrix_1000.csv.gz: Compressed CSV file containing a genotype matrix of 1,000 randomly selected SNPs from Chromosome 1. The accessions in this matrix correspond to those in sap_flowering.pheno.
• data/260706_df_full_comparison.csv: CSV file containing Plant Caduceus model inference scores (including embedding distance, expression score, and log-likelihood probability score along with ensemble ranks) for the 1,000 selected Chromosome 1 SNPs in sap_flowering_1000.vcf.
• data/sap_flowering_alleles.tsv: TSV file mapping each of the 1,000 Chromosome 1 SNPs to its exact reference genome allele and alternative allele.

---

Sorghum bicolor Minimal Ensembl VEP (Chr09)

A minimal, offline setup of Ensembl Plant VEP (Variant Effect Predictor) tailored for predicting variant effects on Chromosome 9 of Sorghum bicolor. Designed to be loaded and run in seconds inside Google Colab without overriding or breaking your existing Python/ML environment.

The codebase for this minimal setup is located in the sorghum-vep-minimal/ directory.

Features

• Ultra-fast setup: Uses a standalone micromamba binary to install VEP and htslib into an isolated prefix. Takes ~60 seconds to set up.
• Isolated environment: Preserves Colab's default Python environment (including PyTorch, CUDA, and transformers).
• Coordinate-aligned: Automatically resolves naming mismatches between Ensembl GFF3 (which uses 9) and Phytozome FASTA (which uses Chr09).
• Python Integration: Seamlessly write variant DataFrames to VCF, run VEP, and read predicted consequences back into Pandas DataFrames.

---

Quick Start in Google Colab

Run the following cell to bootstrap the environment and download reference data.

1. Installation & Setup

# 1. Clone this repository
!git clone https://github.com/scicrow/AI_WinterSchool_2026.git
%cd AI_WinterSchool_2026/sorghum-vep-minimal

# 2. Bootstrap the isolated VEP conda environment
!bash colab_bootstrap.sh

# 3. Download and filter Chr09 GFF3/FASTA files
!python prepare_data.py

2. Running VEP from Python

Here is how you can integrate VEP into your python workflow:

import pandas as pd
from vep_wrapper import write_snps_to_vcf, run_vep, load_vep_results

# Define some test SNPs on Chromosome 9
snps = [
    {"id": "test_snp_1", "pos": 12682424, "ref": "C", "alt": "G"},
    {"id": "test_snp_2", "pos": 25373090, "ref": "A", "alt": "G"}
]

# 1. Convert to VCF format
write_snps_to_vcf(snps, "input_variants.vcf")

# 2. Run VEP programmatically in offline mode
run_vep("input_variants.vcf", "vep_predictions.txt")

# 3. Load VEP predictions directly into a Pandas DataFrame
df_predictions = load_vep_results("vep_predictions.txt")
display(df_predictions.head())

---

Local Installation

If running locally on a system with Conda/Mamba installed:

# Navigate to the tool directory
cd sorghum-vep-minimal

# Create the environment
conda env create -f environment.yml

# Activate the environment
conda activate sorghum-vep

# Download and index reference data
python prepare_data.py

sorghum-vep-minimal Structure

• environment.yml: Conda environment configuration.
• colab_bootstrap.sh: Automates isolated conda setup in Colab.
• prepare_data.py: Handles downloading, renaming, and indexing chromosome 9 files.
• vep_wrapper.py: Provides Python APIs to generate VCFs, invoke VEP, and load predictions.