Skip to content

FAQ

Tutorial notes: Core usage · Models, downloads and integrations · Bioinformatics · API parameters · Catalog scope.

How can pip install the genomics computing dependencies together?

From this checkout, run python -m pip install -e ".[genomics]" for the Python libraries used by genotype PCA, ANI, prokaryotic CDS prediction and BAM statistics. After publication, use python -m pip install "dnakit[genomics]". These APIs do not require separate PLINK, FastANI or corresponding scientific executables.

See the genomics tutorial for examples and the 43-analysis migration notes for the scope of each external adapter. Reference databases, model weights and remaining professional backends are still task-specific dependencies.

Where do the Skills come from, and how are they cited and installed?

DNAKit curates DNA/genomics workflows from bioSkills, ToolUniverse, ClawBio, SciAgent, K-Dense, NVIDIA BioNeMo, DeepMind, PKU-EMBL, omics-skills, ENCODE Toolkit, and other original sources. Task/stage and input-condition equivalence rules choose one implementation per duplicate group. Each skill has its own directory with supporting files, original licenses, and UPSTREAM.md; portability changes are documented.

With a DNAKit version containing this feature:

dnakit skills search "variant annotation" --limit 5
dnakit skills install dnakit --agent codex
dnakit skills install variant-annotation --agent codex
dnakit skills cite variant-annotation

Install the dnakit discovery skill to describe tasks in natural language and let the Agent find an appropriate workflow. MCP clients can use search_dnakit_skills and read_dnakit_skill for selective reading. See the Skills tutorial for paths, Claude Code/Cursor, offline examples, and dependencies. Install the current checkout with python -m pip install -e "."; PyPI installation requires a published release containing this feature.

Use dnakit skills stats and skills/README.md for current counts. Pinned sources, deduplication decisions, license-review exclusions, and a publication inventory are in registry/. A small set of supporting skills completes dependencies of the included composite workflows.

Third-party content retains its MIT, Apache-2.0, CC-BY-4.0, or AGPL-3.0-only terms and original authorship. Candidates without established content permissions, with current use restrictions, or missing essential resources remain in the review records. See the discovery report for search scope and exclusions, and third-party notices for license scope.

Credit DNAKit's curation and integration separately from the original workflows and cite the actual tools, models, datasets, and methods used. Installing instructions does not install analysis dependencies, and catalog inclusion does not establish scientific validation.

All bundled skills now have DNAKit integration guidance. Collection prefixes are removed where applicable, with previous identifiers retained as aliases. Actual software and model names remain unchanged.

Use dnakit.databases for bundled database adapters, dnakit.encode for ENCODE, and dnakit skills run / dnakit.workflows.run() for packaged scripts. The migrated collection adapters do not require ToolUniverse, ClawBio or ENCODE Toolkit installations. Real scientific libraries, programs, model weights, reference data and credentials remain task-specific dependencies. See each skill's DNAKIT.md and the native interface guide.

What matching strategy is used for sequence function search?

The 10 types of functions in sequence function search are all deterministic rule algorithms: when the algorithm version, input sequence and parameters are the same, the results must be the same. They do not use machine learning models, random sampling, or inference from experimental data. The PWM score is also a matching score calculated according to a fixed formula, not a model predicted probability.

These functions return candidate sequences that match the rules. Hits on promoter motifs, TF motifs, PAMs, or inverted repeats do not necessarily mean that the position must have true promoter activity, transcription factor binding, CRISPR editing efficiency, or hairpin structure.

Universal matching convention

  • Strand direction: Scan the forward strand directly; generate the reverse complementary sequence first and then scan the reverse strand; strand="both" scan both strands.
  • IUPAC Match: Each symbol represents a set of possible bases, such as N=ACGT, R=AG, Y=CT, W=AT. When the base set represented by the target symbol and the rule symbol intersect, the position is compatible; the entire window is hit only when all positions are compatible.
  • Gap: Explicit Gap is a scan boundary and candidates will not cross the Gap.
  • Circular sequence: exact/IUPAC motif, PWM, promoter motif, restriction enzyme site and PAM can span the circular origin under supported conditions; canonical motif, codon and repeat structure scans cannot span the circular origin.
  • Resource upper limit: max_matches Limit the number of returns. When the upper limit is reached, the result will be marked as truncated; when the number of comparisons or scanning units exceeds the corresponding upper limit, an error will be reported and calculations will not be silently omitted.

1. Common motif and PWM

  • exact: Scan bit by bit with a window of the same length as the motif, and the window string must be exactly the same as the motif to hit.
  • IUPAC: Also scan window by window, but each position is judged according to the IUPAC set compatibility rules above.
  • regex: Use restricted DNA regular expression lookups; whether overlapping hits are allowed is controlled by overlapping.
  • PWM: Compute a fixed log2-odds score for each window of the same length as the matrix: sum over positions of log2(PWM probability / background probability). The score is greater than or equal to threshold to hit; the background default A, C, G, T are 0.25 each, and the target window containing ambiguous bases will be skipped.

The motif, canonical, PWM and threshold are provided by the caller and are not built-in biological function prediction models.

2. Start and stop codons

The program starts from the 0th, 1st, and 2nd bases of each selected strand respectively, and advances 3 bp each time to form three reading frames; a total of six reading frames are obtained when scanning both strands. Each triplet is a hit as long as it belongs to the start or stop codon set.

  • NCBI Genetic Code Table 1: The start codon is ATG, and the stop codon is TAA, TAG, TGA.
  • NCBI Genetic Code Table 11: The start codons are ATG, GTG, TTG, and the stop codons are TAA, TAG, TGA.
  • When passing in start_codons or stop_codons, use the collection provided by the caller instead of the corresponding default collection.

This feature only reports individual codon positions, does not require an in-frame stop codon following the start codon, and is not equivalent to full ORF detection.

3. Promoter motif

The program treats each promoter consensus sequence as an IUPAC motif and matches it window by window on the selected strand. The only built-in rules are:

  • Eukaryotic TATA box: TATAWAWR;
  • Bacteria -Zone 10: TATAAT;
  • Bacteria -Zone 35: TTGACA.

When motifs is passed in, the caller-supplied motif mapping is used instead of the built-in mapping. The program does not examine the distance between the -10/-35 regions, the transcription start site, or other regulatory elements and therefore only returns promoter motif candidates and does not predict promoter activity.

4.TF motif

The caller must provide the PWM and threshold of the transcription factor. The program scores windows by windows using the same log2-odds formula as the general-purpose PWM, returning only windows with scores greater than or equal to the threshold. DNAKit does not have built-in TF motif databases such as JASPAR, nor does it interpret scores as true binding strengths or binding probabilities.

5. Restriction enzyme sites

The program matches the restriction enzyme recognition sequence on both front and back strands according to IUPAC rules. After a hit, the cut point is calculated based on the upper and lower strand cutting offsets in the definition of the enzyme. The forward and reverse strand repeat results of palindromic recognition sites are combined.

The built-in mini directories include BamHI, EcoRI, HaeIII, HindIII, NotI, and SmaI; a custom RestrictionEnzyme can also be passed in. This rule does not include full REBASE and does not take into account the impact of methylation on digestion.

6. CRISPR PAM

The program first searches for the PAM bit by bit according to IUPAC rules on the selected chain, and then extracts a fixed-length guide from the adjacent position based on whether the PAM is located on the 3' side or the 5' side:

  • SpCas9: NGG, PAM is located on the 3' side of the guide, the default guide length is 20 bp;
  • SaCas9: NNGRRT, PAM is located on the 3' side of the guide, the default guide length is 21 bp;
  • AsCas12a: TTTV, PAM is located on the 5' side of the guide, the default guide length is 20 bp.

The guide does not hit when it exceeds the sequence boundary; guides containing ambiguous bases are excluded by default. Then do deterministic filtering by min_gc, max_gc and exclude_motifs. Custom PAMRule can also be provided. This feature does not predict editing efficiency and off-target risk.

7. Palindrome sequence

The program enumerates each starting point and each length in the range min_length to max_length, comparing the candidate fragment bit by bit to its own reverse complement; only hits are made if all positions meet IUPAC compatibility rules. maximal_per_start=True, only the longest hit is retained for each starting point.

8. Repeat upside down

The program enumerates the starting point, arm length and loop length of the left arm, and then calculates the reverse complementary sequence of the left arm; it hits when the right arm and this sequence satisfy the IUPAC compatibility rules bit by bit. Currently no mismatch or indel is allowed, so it is only a sequence candidate for potential hairpin structures, no folding or free energy predictions are made.

9. Tandem repetition

The program enumerates repeating units from each starting point in ascending order of unit length, checking whether subsequent adjacent units are exactly the same as the first unit string, and extending to the first different location. After reaching the corresponding minimum number of repetitions, the smallest qualified unit from the starting point and its longest continuous repetition interval are reported.

The exact character equality rule applies here, and the IUPAC ambiguous symbols must also be the same character; no mismatch, indel, or break is allowed. When set to overlapping=False, scanning will continue from the end of the repeat interval after a hit; when set to True, scanning will continue from the next base.

10. Microsatellites

Microsatellite invokes the same tandem repeat algorithm, but fixes the repeat unit length at 1 to 6 bp. The default minimum number of repetitions is:

  • 1 bp unit repeated at least 6 times;
  • The 2-6 bp unit is repeated at least 3 times.

Matches must be consecutive and identical, no breaks, mismatches, or indels allowed. min_repeats_by_unit The threshold can be modified, but a threshold must be provided for each unit length from 1 to 6 bp.

What calculation methods and references are available for Diversity and Novelty?

DNAKit uses the paper's raw Levenshtein-distance formula for Diversity; Novelty offers two methods:

Metric Calculation Optional method
Diversity Σ(i≠j) Levenshtein(xᵢ,xⱼ) / [n(n−1)], equivalent to the mean raw edit distance over all unordered sequence pairs. None
Novelty novelty_calculation="similarity": each query receives 1 - nearest_reference_similarity. novelty_calculation="levenshtein": meanᵢ minₛ Levenshtein(queryᵢ, referenceₛ).

The Levenshtein formulas follow Cherednichenko & Poptsova, Data augmentation with generative models improves detection of Non-B DNA structures, Computers in Biology and Medicine 184 (2025) 109440, DOI 10.1016/j.compbiomed.2024.109440. Section 2.8, equations (20) and (21), specifies Levenshtein distance. The terminology is inherited from Jain et al., Biological Sequence Design with GFlowNets, ICML 2022, PMLR paper page.

The article has an official GitHub repository, with the relevant code in seq_analysis.ipynb. As of 2026-09-02, the repository declares no open-source license, and its notebook uses fixed 100-bp preprocessing, flattened one-hot/KDTree operations, and chunk-local calculations that do not fully match the published equations. DNAKit therefore does not copy that code: it independently implements the published formulas with its bounded Levenshtein routine, performs no implicit padding or truncation, and does not claim exact reproduction of the notebook values in Table 2.

Levenshtein results are unnormalized counts of edit operations. Larger values mean greater diversity or greater distance from the reference library; compare datasets with similar sequence-length distributions.

What are the calculation basis and references for physical and chemical properties?

The functions in Physical and Chemical Properties do not use machine learning models. They are divided into theoretical formulations, public empirical parametric models, DNAKit internal transparency rules, and external Primer3 thermodynamic structure predictions. The table below itemizes the actual basis; internal rules for unsourced papers are clearly marked, without supplementary packaging with non-existent citations.

Function Basis for calculation References or sources
Molecular Weight Anhydrous DNA residue mass summation, end and 5' phosphate correction The current implementation records the standard oligonucleotide mass formula and includes a local Biopython numerical comparison, but this comparison cannot replace a primary scientific citation.
260 nm extinction coefficient Sum of adjacent dinucleotide hypochromicity parameters minus internal single base parameter Warshaw & Tinoco, 1966, DOI 10.1016/0022-2836(66)90115-X; Cantor, Warshaw & Shapiro, 1970, DOI 10.1002/bip.1970.360090909.
Tm (Wallace) 2 × (A+T) + 4 × (G+C) Rule of thumb for short oligonucleotides Wallace et al., 1979, DOI 10.1093/nar/6.11.3543. The current implementation does not write this DOI into the result provenance, and the FAQ is here to supplement the literature relationship.
Tm (nearest-neighbor) Neighbor packing, terminal, symmetry, salt concentration and chain concentration models SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460.
salt concentration correction SantaLucia monovalent salt entropy correction SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460.
local Tm Sliding window repeatedly calls Wallace or nearest-neighbor Tm Does not introduce a new scientific model; references inherit the chosen Wallace 1979 or SantaLucia 1998 method.
Synthesis Risk Five-component equal weighted average of GC, isobase contiguous, tandem repeats, and inverted repeats DNAKit internal transparent heuristic rules; no external paper or vendor rule sets, scores are not experimental synthesis success probabilities.
Thermodynamic Parameters ΔH, ΔS, ΔG and Tm for a perfectly complementary duplex SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460.
Nearest-neighbor Same as thermodynamic parameters, but returns stack-by-stack step details SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460.
Duplex stability Native full complementation path using SantaLucia; optional mismatch/dangling path calling Primer3 SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460; Untergasser et al., 2012, DOI 10.1093/nar/gks596. The stable Boolean criterion Tm > configured temperature is a DNAKit result-interpretation rule.
Base stacking Query ΔH/ΔS parameters and calculate ΔG by adjacent dinucleotide SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460.
Hairpin User-installed Primer3 ntthal Thermodynamic Hairpin Structure Prediction Untergasser et al., 2012, DOI 10.1093/nar/gks596 and Primer3 Official Manual ; specific values also depend on the actual Primer3 version and parameter directory.
Self-dimer User-installed Primer3 ntthal thermodynamic self-dimer prediction Untergasser et al., 2012, DOI 10.1093/nar/gks596 and Primer3 official manual ; specific values also depend on the actual Primer3 version and parameter directory.
Heterodimer User-installed Primer3 ntthal thermodynamic heterodimer prediction Untergasser et al., 2012, DOI 10.1093/nar/gks596 and Primer3 Official Manual ; exact values also depend on the actual Primer3 version and parameter catalog.

These references illustrate the sources of algorithms and parameters and do not imply that any calculated results are equivalent to experimental measurements. Tm, ΔG, and structural results are still affected by salt concentration, chain concentration, temperature, sequence length, chemical modifications, and the actual backend version.

What are the calculation basis and references for double-chain thermodynamic expansion?

Double-stranded thermodynamics extension does not use machine learning. Native double-strand energies are derived from SantaLucia public parameters, other results are standard thermodynamic relations, DNAKit transparent combination rules, or external Primer3 calculations.

Function Actual calculation basis References or sources
1. Perfectly complementary duplex ΔH/ΔS/ΔG/Tm Adjacent stacking, termini, symmetry, monovalent salt and chain concentration formulas SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460. This DOI is recorded in the provenance of native results.
2. Unified double-chain stability native follows item 1; primer3-cli calls the mismatch/dangling-end structural model of ntthal The native path refers to SantaLucia 1998; the Primer3 path refers to Untergasser et al., 2012, DOI 10.1093/nar/gks596 and Primer3 official manual. Tm > configured temperature is a DNAKit result-interpretation rule and does not have a separate paper.
3. Contribution of adjacent base pairs to steps Check the adjacent stacking ΔH/ΔS table and calculate ΔG=ΔH−TΔS/1000 SantaLucia, 1998, DOI 10.1073/pnas.95.4.1460. No additional references are made to the new model.
4. Conditions and Na⁺/K⁺ The condition object validates and records numeric values; the native model corrects ΔS using 0.368(N−1)ln([Na⁺]+[K⁺]) The salt correction refers to SantaLucia 1998. Object validation and merging the Na⁺+K⁺ fields is a DNAKit implementation rule and does not have a separate paper.
5. Ka, Kd and double-strand ratio Ka=exp(−ΔG/RT), Kd=1/Ka, and then solve the ideal two-state mass conservation equation ΔG cited SantaLucia 1998; the equilibrium constant relationship and quadratic equation are standard thermodynamics and DNAKit transparent algebra implementation, not bound to separate papers or fitted data.
6. Theoretical melting curve Repeat items 1 and 5 at each temperature, and do linear interpolation for the intersection point of double-strand ratio 0.5 Reference inheritance SantaLucia 1998; temperature scan and linear interpolation are DNAKit internal deterministic rules, no new paper.
7. 5′/3′ end stability Calculate the nearest neighbor ΔG for two equal-length end windows separately and then compare Energy reference SantaLucia 1998; window truncation and "higher ΔG is the less stable end" are DNAKit internal rules.
8. DMSO/formamide correction Linear empirical summation formula of Primer3 manual Primer3 official manual . The default DMSO factor of 0.6 is quoted in the manual from Musielski et al., 1981; the formamide formula is quoted from Blake & Delcourt, 1996, DOI 10.1093/nar/24.11.2095. This function is not a mechanistic free energy model.
9. Primer3 CLI extension oligotm calculates Tm, ntthal calculates hairpins and dimers; DNAKit only does adapter Untergasser et al., 2012, DOI 10.1093/nar/gks596 and Primer3 official manual . The actual Primer3 version and thermodynamic parameter catalog should also be documented.

What are the calculation basis and references for secondary structure properties?

Secondary structure properties Contains DNAKit internal parsing/formulas and conditional NUPACK adapter. Structural energy and equilibrium results are NUPACK calculations only if items 6 and 7 are run explicitly.

Function Actual calculation basis References or sources
1. Dot-bracket parsing Stack pairing, continuous nested pair merging and fixed structure classification DNAKit internal deterministic parser, no paper parameters or training models. For the symbol convention of Dot-parens-plus, please refer to Zadeh et al., 2011, DOI 10.1002/jcc.21596; the extended bracket parsing of () [] {} <> is implemented by DNAKit.
2. Pairing probability and window accessibility Marginal probability derivation of NUPACK style dense probability matrix For matrix semantics, please refer to NUPACK analysis document and Zadeh et al. 2011. Symmetry/rowsum checks and "arithmetic mean of window edge probabilities" are DNAKit internal rules, not NUPACK joint open probabilities.
3. Normalized ensemble defect The expected proportion of each base that does not adopt the target pairing state Zadeh, Wolfe & Pierce, 2011, DOI 10.1002/jcc.21633. DNAKit only evaluates this definition for the provided probability matrix.
4. Target structure probability exp[−(Gtarget−Gensemble)/(RT)] Boltzmann relation Standard statistical thermodynamic formula, DNAKit is not bound to a separate paper or fitting parameters. If the free energy comes from NUPACK, the actual NUPACK model and version should be quoted.
5. NUPACK passive detection Python package location and metadata check Engineering environment check, does not perform NUPACK, so no scientific paper citations; the results must not be written as NUPACK predictions.
6. NUPACK single complex Partition functions, MFE, pairing probabilities, suboptimal structures, and Boltzmann sampling for external NUPACK 4 DNAKit provenance Current record Zadeh et al. 2011, DOI 10.1002/jcc.21596. The NUPACK 4 official citation page also specifies Fornace, Porubsky & Pierce, 2020, DOI 10.1021/acssynbio.9b00523; citations should be selected according to the actual calling function and version, see NUPACK official citation page .
7. NUPACK tube equilibration Multi-chain complex and tube equilibration analysis of external NUPACK 4 Dirks et al., 2007, DOI 10.1137/060651100; Fornace, Porubsky & Pierce, 2020, DOI 10.1021/acssynbio.9b00523; and NUPACK official citation page . DNAKit's target/non-target ratio is an internal summation of the concentrations returned by the adapter.

What are the calculation basis and references for three-dimensional structure and mechanical properties?

Three-dimensional structure and mechanical propertiesDo not predict three-dimensional structures from DNA sequences. Items 1–4 are deterministic analytical/geometric calculations of explicit coordinates, items 5 and 6 are adapters to external 3DNA/DSSR results.

Function Actual calculation basis References or sources
1. Read a single PDB model legacy PDB MODEL/ATOM fixed columns, DNA residue table and alternative conformation selection wwPDB PDB v3.3 coordinate record specification . This is a file format source, not a scientific prediction paper.
2. Read PDB multi-model Same as item 1, then group by MODEL sequence number wwPDB PDB v3.3 coordinate record specification . Multi-model organizations are format rules and there are no individual predictive models.
3. Explicit Coordinate Geometry Center of mass, radius of gyration, tensor eigenvalues, spherical point sampling SASA, voxel volume, dihedral angles, distance/angle hydrogen bond filtering and global helix axis approximation Current results provenance marks the entire native geometry implementation as a DNAKit internal method, unbound to the unity paper. Among them, SASA is the Shrake–Rupley style spherical point exposure method, please refer to Shrake & Rupley, 1973, DOI 10.1016/0022-2836(73)90011-9; voxel volume, hydrogen bond threshold and C1′ global axis are DNAKit specific rules and should not be written as 3DNA/DSSR standard parameters.
4. NMR/Multi-model RMSF Square root of mean square displacement after centering of common atomic translations DNAKit internal implementation of the standard RMSF mathematical definition, not bound to a separate paper. The implementation does not do a Kabsch rotation fit and does not infer persistence lengths or mechanical moduli.
5. 3DNA bp_step.par Analysis Read the 12 rigid body parameters calculated by 3DNA. DNAKit only finds the average, 360/twist and pitch Lu & Olson, 2003, DOI 10.1093/nar/gkg680 and Official description of 3DNA parameters .
6. DSSR JSON summary Read DSSR identified nucleotide, pair, helix, stem, hairpin, and hydrogen bond counts Lu, Bussemaker & Olson, 2015, DOI 10.1093/nar/gkv716 and DSSR JSON official description . DSSR completes structure recognition, and DNAKit only parses existing JSON.

What are the calculation basis and references for optical and concentration conversion?

Optics and concentration conversion does not use machine learning. It uses the public single-stranded ε260 empirical parameters, Beer–Lambert's law, molecular weight and unit conversions, and double-stranded/modified parameters explicitly provided by the caller.

Function Actual calculation basis References or sources
1. Single-chain molar extinction coefficient at 260 nm Σ adjacent-dinucleotide coefficients − Σ internal single-base coefficients Warshaw & Tinoco, 1966, DOI 10.1016/0022-2836(66)90115-X; Cantor, Warshaw & Shapiro, 1970, DOI 10.1002/bip.1970.360090909. Two articles have been written about the provenance of this result.
2. Theoretical optical properties of single chain/double chain Use item 1 for single chain; use 13200 × bp average formula or (ε1+ε2)(1−h) for double chain; use the sum of anhydrous residues for molecular weight Single chain ε cite the above two papers. 13200 M⁻¹·cm⁻¹/bp is the traditional 1 OD260 ≈ 50 µg/mL double-chain average conversion used in the implementation. The current source code is not bound to the original paper; h, Δε and ΔMW are provided by the caller, and their actual experiments, literature or manufacturer data sheets should be cited. The anhydrous residue molecular weight formula is not yet bound to the original paper.
3. 1 OD260 corresponding to nmol/mass 10⁶/ε and 1000×MW/ε Directly derived from OD260 definition and unit conversion, no new paper; citations should inherit from selected ε260 and MW sources.
4. A260 to concentration and total amount c=A/(εl), plus dilution factor, m=nMW, n=cV and explicit dye subtraction Beer–Lambert standard laws of physics with DNAKit transparent unit conversion, currently implemented unbound to a specific modern paper. ε, MW and dye factors should each cite their actual source.
5. Concentration/quantity/mass interchange of substances m=nMW, n=m/MW, c=n/V, n=cV DNAKit internal dimension conversion, not an empirical model, no separate paper.
6. Dye and modification correction Σcount×Δε, Σcount×ΔMW, and Σlabel-peak×correction-factor DNAKit does not have a built-in dye parameter table and only does deterministic addition and subtraction, so there is no unified paper. The caller must document authentic experimental, literature, or manufacturer sources for each parameter.

The above citations only indicate the sources of algorithms, parameters or external tools. For free energy, hypochromicity, dye correction, or 3DNA/DSSR/NUPACK outputs provided by the caller, the actual parameter table, software version, input conditions, and data source must also be documented.

What are the references for deep-learning property prediction?

The following table covers the primary model papers, task datasets, and training protocols used by all 54 functions integrated in deep-learning property prediction. Examples are organized by task: regulatory tracks, splicing and genomic regions, epigenetic marks and open chromatin, promoters and enhancers, variant effects, taxonomy and molecular function, and molecular relationships and interactions. Multiple output heads share one paper, so the same citation is not repeated for every function.

Functions Reference
RNA-seq, CAGE, PRO-cap, ATAC-seq, DNase-seq, ChIP-seq, splicing, and contact maps Avsec et al., Advancing regulatory variant effect prediction with AlphaGenome, Nature 649, 1206–1218 (2026), DOI 10.1038/s41586-025-10014-0.
Human and mouse regulatory tracks Avsec et al., Effective gene expression prediction from sequence by integrating long-range interactions, Nature Methods 18, 1196–1203 (2021), DOI 10.1038/s41592-021-01252-x.
18 NT Revised classification tasks Backbone: Avsec et al., Effective gene expression prediction from sequence by integrating long-range interactions, Nature Methods 18, 1196–1203 (2021), DOI 10.1038/s41592-021-01252-x; task definitions: Dalla-Torre et al., Nucleotide Transformer: building and evaluating robust foundation models for human genomics, Nature Methods 22, 287–297 (2025), DOI 10.1038/s41592-024-02523-z, and the revised dataset card; full-fine-tuning protocol: Wu et al., GENERator: A Long-Context Generative Genomic Foundation Model, arXiv (2025), arXiv:2502.07272, Appendix C.4.
9 Genomic Benchmarks classification tasks Backbone: Avsec et al., Effective gene expression prediction from sequence by integrating long-range interactions, Nature Methods 18, 1196–1203 (2021), DOI 10.1038/s41592-021-01252-x; task datasets: Grešová et al., Genomic benchmarks: a collection of datasets for genomic sequence classification, BMC Genomic Data 24, 25 (2023), DOI 10.1186/s12863-023-01123-8; full-fine-tuning protocol: Wu et al., GENERator: A Long-Context Generative Genomic Foundation Model, arXiv (2025), arXiv:2502.07272, Appendix C.4.
Foundation encoder for 14-class single-nucleotide genome segmentation Dalla-Torre et al., Nucleotide Transformer: building and evaluating robust foundation models for human genomics, Nature Methods 22, 287–297 (2025), DOI 10.1038/s41592-024-02523-z.
Trained 14-class single-nucleotide genome segmentation head de Almeida et al., Annotating the genome at single-nucleotide resolution with DNA foundation models, Nature Methods (2025), DOI 10.1038/s41592-025-02881-2.
Long-context zero-shot variant effects and exon probability Brixi et al., Genome modelling and design across all domains of life with Evo 2, Nature (2026), DOI 10.1038/s41586-026-10176-5.
Allele-conditional-probability variant effects Wu et al., GENERator: A Long-Context Generative Genomic Foundation Model, arXiv (2025), arXiv:2502.07272; Li et al., GENERator-v2: Reconciling Coarse Tokenization with Single-Nucleotide Resolution in Genomic Language Modeling, bioRxiv (2026), DOI 10.64898/2026.01.27.702015.
Central dogma, taxonomy, species, protein localization/stability, ncRNA family, and pair-interaction tasks He et al., Generalized biological foundation model with unified nucleic acid and protein language, Nature Machine Intelligence 7, 942–953 (2025), DOI 10.1038/s42256-025-01044-4.

Direct inference also depends on the following released code or trained weights:

These citations identify the model, trained task-head, and checkpoint sources. They do not make a prediction equivalent to an experimental measurement or a clinical conclusion.