How google-deepmind/alphagenome_research works
AlphaGenome Research provides a DNA sequence prediction model and a higher-level DNA Model interface for predictions and variant analysis, alongside TFRecord data loading, fine-tuning, and evaluation utilities. The main workflow is user input to DNA Model, model inference, then predictions or variant scores; documented weight providers supply pretrained checkpoints. Internal model components and data/evaluation branches are included. Relationships are limited to README statements and the supplied source index; several subsystem details were not excerpted.
Main entry points: dna_model.py DNA Model API · model.py AlphaGenome model · track_prediction.py Track evaluation · variant_scoring.py Variant scorers
Components 21
Model interface
- DNA Model API
src/alphagenome_research/model/dna_model.py - Sequence predictions
src/alphagenome_research/model/dna_model.py
Variant analysis
- Variant scorers
src/alphagenome_research/model/variant_scoring/variant_scoring.py - Splice scoring
src/alphagenome_research/model/variant_scoring/splice_junction.py - Gene masks
src/alphagenome_research/model/variant_scoring/gene_mask_extractor.py - Interval scoring
src/alphagenome_research/model/interval_scoring/interval_scoring.py
Model engine
- AlphaGenome model
src/alphagenome_research/model/model.py - Prediction heads
src/alphagenome_research/model/heads.py - Attention layers
src/alphagenome_research/model/attention.py - Embeddings
src/alphagenome_research/model/embeddings.py - DNA encoder
src/alphagenome_research/model/one_hot_encoder.py - Track metadata
src/alphagenome_research/model/metadata/metadata.py
Data and training
- TFRecord loader
src/alphagenome_research/io/dataset.py - Track bundles
src/alphagenome_research/io/bundles.py - Fine-tuning dataset
src/alphagenome_research/finetuning/dataset.py - Fine-tuning
src/alphagenome_research/finetuning/finetune.py
Evaluation
- Track evaluation
src/alphagenome_research/evals/track_prediction.py - Regression metrics
src/alphagenome_research/evals/regression_metrics.py
Other components
- Research user
- Kaggle weights
- Hugging Face weights
Connections 20
- Research user → DNA Model API: requests predictions
README.md - DNA Model API → Kaggle weights: downloads weights
- DNA Model API → Hugging Face weights: downloads weights
- DNA Model API → AlphaGenome model: runs inference
src/alphagenome_research/model/dna_model.py - AlphaGenome model → Sequence predictions: produces tracks
README.md - DNA Model API → Variant scorers: scores variants
src/alphagenome_research/model/dna_model.py - DNA Model API → Interval scoring: scores intervals
- Variant scorers → Splice scoring: uses scorer
src/alphagenome_research/model/variant_scoring/splice_junction.py - Variant scorers → Gene masks: extracts masks
src/alphagenome_research/model/variant_scoring/gene_mask_extractor.py - DNA Model API → Track metadata: selects track outputs
src/alphagenome_research/model/dna_model.py - AlphaGenome model → Prediction heads: dispatches heads
src/alphagenome_research/model/model.py - AlphaGenome model → Attention layers: uses attention
src/alphagenome_research/model/model.py - AlphaGenome model → Embeddings: uses embeddings
src/alphagenome_research/model/model.py - DNA Model API → DNA encoder: encodes sequences
src/alphagenome_research/model/dna_model.py - TFRecord loader → Track bundles: parses bundle types
src/alphagenome_research/io/dataset.py - Fine-tuning → Fine-tuning dataset: loads training data
src/alphagenome_research/finetuning/finetune.py - Fine-tuning → AlphaGenome model: trains model
src/alphagenome_research/finetuning/finetune.py - Track evaluation → TFRecord loader: loads evaluation data
src/alphagenome_research/evals/track_prediction.py - Track evaluation → AlphaGenome model: evaluates predictions
src/alphagenome_research/evals/track_prediction.py - Track evaluation → Regression metrics: computes metrics
src/alphagenome_research/evals/track_prediction.py