Training And Running Predictions¶
After the 1000 Genomes dataset is available, you can train neural genotype models or run classical SNP-based superpopulation prediction.
Run the commands below in order for the neural path. Use the classical SNP section when you want a VCF/SNP baseline instead of AlphaGenome neural tensors.
Neural Genotype Models¶
Prepare or validate the processed tensor cache:
Train a model:
Evaluate a saved checkpoint:
genomics genotype evaluate \
configs/predictors/genotype_based/genes_1000_all_3ontologies.yaml \
--checkpoint best_accuracy \
--split val
Run final held-out test evaluation after model selection:
genomics genotype test \
configs/predictors/genotype_based/genes_1000_all_3ontologies.yaml \
--checkpoint best_accuracy
Enabling DITA¶
DITA is the aligned tensor representation for indel-aware neural models. Enable it when you want to address coordinate shifts caused by indels in dense AlphaGenome tracks.
Set these fields in the genotype config:
dataset_input:
tensor_layout: "haplotype_channels"
alignment_mapping: "bcftools_chain"
feature_mode: "signals_and_masks"
indel_include_valid_mask: true
indel_include_snp_mask: true
Validate required artifacts before training:
genomics audit-data \
--dataset-id 1kg_high_coverage \
--check-bcftools-chain \
--sample-limit 3 \
--fail-on-missing
If artifacts exist in a separate tree, preview syncing them with:
genomics genotype sync-bcftools-artifacts \
--source-dir /path/to/consensus_dataset \
--target-dir /path/to/canonical_dataset \
--link-mode hardlink
Add --apply only after reviewing the preview.
Classical SNP Superpopulation Path¶
If you do not need AlphaGenome neural tensors, run the integrated SNP ancestry pipeline:
For FROGAncestryCalc likelihood comparisons, use the third-party tool directly:
cd third_party/FROGAncestryCalc
python3 tools/vcf_to_frog.py sample.vcf.gz tools/aisnps_55_list.txt input/sample_data.txt
./run.sh
More Details¶
| Topic | Read More |
|---|---|
| Genotype model inputs, caches, and outputs | Genotype Predictor |
| Why DITA exists and what masks mean | Dynamic Indel Tensor Alignment |
| Split policy, preprocessing, stability, and leakage | Preprocessing And Leakage |
| SNP ancestry math and outputs | SNP Ancestry |
| FROGAncestryCalc and native/third-party tools | Native And Third-Party |
| All CLI flags | CLI Reference |