K-Fold

An openly released model developed in Korea that predicts complex structures and interface confidence.

API/CLI tool ID: kfold

The input fields depend on the value of inputFormat.

Understanding inputs and outputs

Apo refers to the structure of a component protein apart from its binding partner. In stage, apo selects this preparation step, complex selects complex prediction, and all performs both. A computationally prepared apo structure is not an experimentally verified unbound structure. Colons in sequence separate chains within one complex.

ranking_score orders candidates within this model. pLDDT assesses local structure, pTM the overall arrangement, and ipTM interfaces between components. pde is a distance error in ångström, not a probability. has_clash flags atoms overlapping too closely; the absence of clashes does not establish actual binding or function.

Input settings

SettingFormatRequiredDefaultDescription and constraints
inputFormattaskOptional"sequence"Input format: sequence separates chains with ':', query is a native K-Fold query file — Allowed values sequence · query
sequencesequenceConditionally required—Separate components (chains) with ':' — Characters ACDEFGHIKLMNPQRSTVWYX:, Applies to sequence
queryFilefileConditionally required—K-Fold query file (YAML/JSON). Specify apo structure paths inside the query. — Extensions yaml · yml · json, Applies to query
seednumberOptional—Random seed for reproducibility (blank: random) — Minimum 0
stagedropdownOptional"all"all = apo preparation + full complex prediction — Allowed values all · apo · complex
kerneldropdownOptional"auto"Backend for AtlasFold apo preparation and complex prediction — Allowed values auto · triton · cuequiv · torch

Output metrics

ColumnDisplay nameRangeUnitDescription
ranking_scoreRanking score——K-Fold's own ranking score (combined across seeds and samples)
plddtpLDDT0~100——
ptmpTM0~1——
iptmipTM0~1—Inter-component interface confidence
pdepDE—ÅPredicted distance error. This metric is not normalized to 0–1.
has_clashClashes———

See Reading results for metric definitions and interpretation. A range shows possible values; whether higher or lower is better depends on the metric.

Execution availability

Open an execution plan in your project to see available tool versions and costs. Review and approve the inputs, settings, and terms before running.

License

Code, model weights, and accompanying data have their own terms. The table summarizes terms from official sources; check your execution plan for the configuration you will use. License texts are available in Third-party notices.

ComponentLicenseCommercial useSourceNotes
CodeApache-2.0Allowed, subject to notice requirementsLicense textK-Fold code is distributed under the Apache 2.0 license.
CodeMITAllowed, subject to notice requirementsLicense textAtlasFold code used for sequence representation is licensed under MIT.
CodeMITAllowed, subject to notice requirementsLicense textTriProRep code for structural representations uses the MIT license.
CodeMITAllowed, subject to notice requirementsLicense textbio-prior is an MIT-licensed training dependency and is not included in the current inference configuration.
Model weightsApache-2.0Allowed, subject to notice requirementsWeight distribution metadataThe distributor’s model card lists the K-Fold weights under Apache 2.0.
Model weightsApache-2.0Allowed, subject to notice requirementsWeight distribution metadataThe asset repository has the same terms
Model weightsMITAllowed, subject to notice requirementsWeight distribution metadataThis entry covers AtlasLM weights distributed under MIT. Other distributions may have different terms.
CodeMPL-2.0Conditional; see notesLicense textgemmi is licensed under MPL-2.0. Distributing an image to customers requires providing the source for the covered files.

atlasfold, TriProRep, and the training-only bioprior use MIT terms. gemmi uses MPL-2.0; redistribution in a container requires providing the corresponding covered source files.

Checked on 2026-10-22.

References

On this page