FAQs

Inputs and coordinates

Which genome assemblies are supported?

Human assembly hg38 (GRCh38.p13.genome.fa) and mouse assembly mm10 (GRCm38.p6.genome.fa). To convert coordinates from other assemblies (such as hg19), use UCSC LiftOver.

What coordinate convention do Interval and Variant use?

Interval uses half-open 0-based coordinates ([start, end)), matching standard Python slicing. For compatibility with VCF and dbSNP, Variant takes a 1-based position and exposes 0-based bounds via start (position - 1) and end.

Which sequence lengths are supported, and how does resize work?

AlphaGenome 2 accepts four input widths: SEQUENCE_LENGTH_16KB (16,384 bp), SEQUENCE_LENGTH_100KB (131,072 bp), SEQUENCE_LENGTH_500KB (524,288 bp), and SEQUENCE_LENGTH_1MB (1,048,576 bp; recommended for best accuracy); predict_interval_gradients() and predict_interval_gradients_batch() accept only SEQUENCE_LENGTH_100KB and SEQUENCE_LENGTH_1MB. Calling resize() expands or crops around the interval center using real flanking reference genome sequence rather than zero-padding.

Can I make predictions for arbitrary or non-human/mouse DNA sequences?

Yes, predict_sequence() accepts any sequence composed of ACGTN at a supported length. However, the models were trained on the human and mouse reference genomes, and predictions for other species, synthetic constructs, large structural variants, or heavily N-padded sequences are not benchmarked.

Conditions and outputs

Should I use a reference condition or expression conditioning?

Start with a reference condition (condition=). Reference conditions in dna_model.supported_conditions() span a broad reference across human and mouse tissues, primary cell types, and cell lines, and cover most use cases. supported_conditions() lists human conditions by default; pass organism=dna_client.Organism.MUS_MUSCULUS for mouse.

Expression conditioning (cell_conditioning= plus cell_condition_name=) lets you apply the model to your own human scRNA-seq data without explicitly fine-tuning it on your dataset: the model encodes up to 256 cells and predicts for that cell state. It generalizes across many cell types, but if your data differ substantially from the training distribution, the model may fail to capture that variation.

See Choosing a condition and Conditioning on your own single-cell data.

What do observed, imputed_supported, and imputed_exploratory mean?

Every predicted track carries an imputation_mode label in .metadata: observed (measured in training), imputed_supported (not measured, but imputable from a measured modality in that condition), or imputed_exploratory (not benchmarked; treat as a hypothesis). See Navigating imputed tracks.

How do I compare predictions or scores across multiple conditions?

Each call to predict_interval, predict_variant, or score_variant takes one condition (condition= or cell_conditioning=). To compare reference conditions or expression-conditioning inputs across items, pass matching sequences (conditions= or cell_conditionings= plus cell_condition_names=) to predict_variants(), score_variants(), or the other batch methods.

AlphaGenome 1 vs. AlphaGenome 2: which model should I use?

Use AlphaGenome 2 (alphagenome.models.v2) for new analyses. Use AlphaGenome 1 (alphagenome.models.v1) to reproduce results from the AlphaGenome paper [Avsec et al., 2026], score variants across all tissues and cell types at once without choosing a condition, or predict PRO-cap. See Getting started and Migrating from AlphaGenome 1.

How are DNA strands labeled?

Track metadata records strand in strand as + (forward), - (reverse), or . (unstranded; for example, ATAC-seq, DNase-seq, and unstranded GTEx RNA-seq).

How do I save outputs?

Flatten variant score AnnData objects with tidy_scores() and export with DataFrame.to_csv(...), export GeneExpressionData with .to_anndata(), or save TrackData.values NumPy arrays with np.savez_compressed(...).

Variant scoring

Are insertions and deletions (indels) supported, and must reference_bases match the genome?

Yes. For scoring, the ALT prediction is aligned to REF coordinates following SpliceAI [Jaganathan et al., 2019] (see How variant scoring works). Variant does not normalize indels; use normalize_variant() to left-align them. In predict_variant(), reference_bases and alternate_bases substitute directly into the reference sequence without requiring reference_bases to match the hg38/mm10 allele.

How should I score splicing variants?

Combine the maximum absolute scores across SPLICE_SITES, SPLICE_SITE_USAGE, and SPLICE_JUNCTIONS (/ 5) on conditions with observed splicing or full-length RNA-seq data. See Analyzing a splicing variant for the full walkthrough.

Does AlphaGenome 2 report quantile scores?

No. AlphaGenome 2 scores report raw_score only; quantile scores calibrated against common variants are available in AlphaGenome 1. See Migrating from AlphaGenome 1.

Usage, licensing, and support

What terms of use apply, and is commercial access available?

AlphaGenome is governed by the AlphaGenome Terms of Service: direct API outputs and AlphaGenome Atlas scores are for non-commercial research use only, must not be used to train machine learning models, and must not be used for clinical decision-making. For commercial and enterprise workflows, access AlphaGenome on Google Cloud via Model Garden on Gemini Enterprise Agent Platform or contact alphagenome-cloud-external@google.com.

How should I cite AlphaGenome?

If you use AlphaGenome in your research, please cite:

@article{alphagenome,
  title={Advancing regulatory variant effect prediction with {AlphaGenome}},
  author={Avsec, {\v Z}iga and Latysheva, Natasha and Cheng, Jun and Novati, Guido and Taylor, Kyle R. and Ward, Tom and Bycroft, Clare and Nicolaisen, Lauren and Arvaniti, Eirini and Pan, Joshua and Thomas, Raina and Dutordoir, Vincent and Perino, Matteo and De, Soham and Karollus, Alexander and Gayoso, Adam and Sargeant, Toby and Mottram, Anne and Wong, Lai Hong and Drot{\'a}r, Pavol and Kosiorek, Adam and Senior, Andrew and Tanburn, Richard and Applebaum, Taylor and Basu, Souradeep and Hassabis, Demis and Kohli, Pushmeet},
  journal={Nature},
  volume={649},
  number={8099},
  pages={1206--1218},
  year={2026},
  doi={10.1038/s41586-025-10014-0},
  publisher={Nature Publishing Group UK London}
}

If you use the AlphaGenome Atlas website or scores, please cite:

@article{alphagenome_atlas,
  title={{AlphaGenome} Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants},
  author={Cheng, Jun and Taylor, Kyle R. and Nicolaisen, Lauren and Pan, Joshua and Bycroft, Clare and Perino, Matteo and Ward, Tom and Hawkes, Gareth and Covil, Laura E. and Weilert, Melanie and Thomas, Raina W. and Latysheva, Natasha and Hirschmann, J. Maile and Chen, Xi Dawn and Beaumont, Robin N. and Chundru, V Kartik and Weedon, Michael N. and Bourdareau, Simon and Chu, Hoyin and Hariharan, Dhavanthi and Kagohara, Thais and Ten\'{o}rio, Lucas and Ushigome, Yosuke and Shearer, Courtney A. and Ikica, Barbara and Fang, Ada and Naciri, Mouad and Johnston, Victoria and Green, Richard and Wong, Lai Hong and Dutordoir, Vincent and Mottram, Anne and Gayoso, Adam and Arvaniti, Eirini and Rehm, Heidi L. and Chen, Fei and Lareau, Caleb A. and Wright, Caroline F. and O'Donnell-Luria, Anne and Zeitlinger, Julia and Kohli, Pushmeet and Avsec, {\v Z}iga},
  year = {2026},
  doi = {10.64898/2026.09.16.26363192},
  publisher = {Cold Spring Harbor Laboratory Press},
  journal = {medRxiv}
}

Where do I report issues or ask questions?

Report software issues on GitHub and post usage questions on the AlphaGenome community forum. For direct inquiries, email alphagenome@google.com.