Caliby

Caliby webserver overview

Caliby 0.1 designs protein sequences conditioned on one structure or an aligned conformational ensemble. It also scores the sequence in a structure, packs sidechains, generates backbone ensembles with Protpardelle-1c, and offers explicit structure cleaning. Single-structure and ensemble tasks use separate inputs.

This integration targets the Caliby source and its command-line defaults. Native structures, score tables, configuration files, and logs are retained. Caliby and its published model collection are Apache-2.0. Optional AlphaFold2 parameters are provided by DeepMind under CC-BY-4.0; credit the AlphaFold and AlphaFold-Multimer authors when reporting refolding results.

Pricing

Runs cost 27 credits per minute of measured runtime. The minimum reservation is 27 credits, not a minimum final charge. Completed runs are charged in proportion to elapsed time, rounded up to a whole credit; unused reserved credits are returned. Set a spending limit before submission. Each batch job is metered separately.

Inputs

Input	Formats	Requirement
Structures	PDB, CIF	Single-structure design, scoring, packing, ensemble generation, or cleaning. Multiple independent structures may be supplied.
Structural ensembles	ZIP	Ensemble design or scoring. Each directory contains a named primary structure and its conformers.
Residue constraints	CSV	Optional for design tasks. Uses native chain identifiers and label_seq_id positions.
Input selection	TXT	Optional native filename list, or directory names for ensembles.
Protpardelle sampling configuration	YAML	Optional data-only configuration for ensemble generation with the reviewed cc95 epoch 3490 partial-diffusion model.

Each uploaded file may be up to 50 MiB, subject to the plan's file limits. Runs have a one-hour execution limit. A ZIP can contain several ensemble directories. For example, protein/protein.cif is the primary structure and protein/decoy_1.cif is another conformation. PDB and CIF conformers are supported. Directory and file names are preserved.

Single-structure design, scoring, packing, and cleaning also accept .pdb.gz and .cif.gz. Compressed files retain their original bytes and filenames. Ensemble generation requires uncompressed PDB or CIF files. Ensemble ZIPs may retain auxiliary metadata such as scaffold_info.csv; Caliby selects the PDB/CIF conformers itself.

Selection TXT files list one exact structure filename or ensemble directory name per line. Uploaded filenames can contain Unicode characters but cannot contain directory paths. The default Protpardelle sampling configuration uses the cc95 epoch 3490 model with 150 rewind steps; an optional YAML file can adjust its supported sampling settings.

Selected conformations must have matching chain identifiers, residue ordering, and residue numbering. Mismatches fail without automatic alignment or renumbering. Ensemble scoring uses the sequence of the first selected conformation: the named primary by default, or the first remaining conformer when primary inclusion is disabled. A named primary file is still required in either case. The native conformer ordering and default maximum of 32 conformers apply.

Residue constraints and symmetry

The CSV requires pdb_key, matching the structure stem or ensemble directory name. For compressed structures, the native stem removes only .gz: protein.pdb.gz uses protein.pdb as its pdb_key. Supported optional columns are:

Column	Meaning
fixed_pos_seq	Preserve the sequence at listed positions.
fixed_pos_scn	Preserve sidechains at listed positions, which must also be included in fixed_pos_seq.
fixed_pos_override_seq	Set specified residue identities.
pos_restrict_aatype	Restrict individual positions to specified amino acids.
symmetry_pos	Tie sequence choices across symmetry-related positions.

Constraints are passed directly to Caliby. Use its constraint examples for exact CSV syntax. Distilled models do not support fixed sidechain conditioning and reject those requests.

Settings

Unspecified JSON options retain native defaults. false, 0, and null are passed explicitly when supplied. The source validates scientific combinations and supported ranges.

Parameter	Type	Default	Description
Task	enum	Single-structure design	Selects design, scoring, packing, generation, or cleaning. Ensemble modes have their own inputs.
Model	enum	Native default	caliby for design and scoring; caliby_packer_010 for packing.
Native run options	object	{}	Mode-specific run options below.
Sampling settings	object	{}	Sequence and sidechain sampling options below.
AlphaFold2 settings	object	{}	Optional design refolding settings below.
Job name	string	optional	Label for identifying the run in job history.

Design/scoring models are caliby, soluble_caliby, soluble_caliby_v1, caliby_distill, soluble_caliby_distill, and the published caliby_L checkpoint. Packing models are caliby_packer_000, caliby_packer_010, and caliby_packer_030. Custom checkpoint URLs and unreviewed assets are unavailable.

Native run options

Parameter	Type	Default	Description
seed	integer	0	Native random seed, except cleaning.
num_workers	integer	mode-dependent	2 for design, ensemble scoring, and packing; 4 for single scoring; 8 for generation and cleaning.
max_num_conformers	integer or null	32	Ensemble tasks only. null derives the limit from the first selected ensemble in this version.
include_primary_conformer	boolean	true	Includes the primary structure in ensemble conditioning.
save_local_conditionals	boolean	false	Scoring tasks only. Saves native per-position conditional arrays.
run_self_consistency_eval	boolean	false	Design tasks only. Runs optional AlphaFold2 refolding.
input_cfg.n_subsample	integer or null	null	Native structure subsampling for single-structure tasks.
input_cfg.pdb_name_ext	string or null	empty string	Replaces selection-list filename extensions before loading.
input_cfg.array_id, input_cfg.num_arrays	integer or null	null	Native partitioning of ensemble or generation inputs.
num_samples_per_pdb	integer	32	Ensemble generation only.
batch_size	integer	8	Ensemble generation only.

Sampling settings

These settings apply to design, scoring, and packing. Nested options are JSON objects, for example {"potts_sampling_cfg":{"potts_temperature":0.01}}.

Parameter	Type	Default	Description
batch_size	integer	mode-dependent	4 for design and packing; 16 for scoring.
num_seqs_per_pdb	integer	16 for design	Number of designed sequences per input.
num_workers	integer	mode-dependent	Inherits Native run options num_workers unless overridden here.
verbose	boolean	true	Native sampling diagnostics.
omit_aas	array or null	null	Amino-acid letters to omit from sequence sampling.
gaussian_conformers_cfg.n_conformers	integer	0	Additional noisy conformers.
gaussian_conformers_cfg.noise_std	number	0.0	Gaussian coordinate noise.
potts_sampling_cfg.regularization	string or null	LCP	Native Potts regularization.
potts_sampling_cfg.potts_sweeps	integer	500	Number of sampling sweeps.
potts_sampling_cfg.potts_proposal	enum	dlmc	dlmc or chromatic.
potts_sampling_cfg.potts_temperature	number	0.01	Final sampling temperature.
potts_sampling_cfg.rejection_step	boolean	false	Native rejection step.
potts_sampling_cfg.potts_only_cond	boolean	false	Native conditional-only sampling.
scn_packing_cfg.num_steps	integer	50	Sidechain diffusion steps.
scn_packing_cfg.step_scale	number	1.5	Sidechain diffusion step scale.

ensemble_ignore_res_idx_mismatch may only be false. Alignment checks cannot be disabled.

Optional refolding

AlphaFold2 refolding uses locally installed ColabDesign and model parameters. It does not submit sequences to an external folding or MSA service.

The pinned Caliby refolding path supports single-chain outputs. Multichain structures fail with the native ColabDesign error, including when use_multimer is enabled. That setting selects model parameters; it does not add multichain input support. Multichain sequence design remains available with refolding disabled.

Parameter	Type	Default	Description
num_models	integer	5	Models evaluated per sequence.
sample_models	boolean	true	Native model sampling.
num_recycles	integer	3	Recycling iterations.
save_best	boolean	true	Retained native configuration field. The pinned Caliby implementation always saves its best-pLDDT prediction.
use_multimer	boolean	false	Selects AlphaFold-Multimer with the five official multimer-v3 parameter sets.

Outputs

The Results tab shows the main native table when the task produces one. Refolding contains self-consistency metrics when requested. Files provides downloads of the original structures, tables, arrays, configuration files, and logs.

Task	Native outputs
Design	seq_des_outputs.csv with identifiers, input/designed sequences, U, and structure paths; sampled CIF structures.
Scoring	score_outputs.csv with identifiers, sequences, and U; optional local_conditionals/*.npy arrays.
Packing	packing_metrics.csv, packed CIF structures, and aligned input structures.
Ensemble generation	Protpardelle structures, input metadata, and sampling configuration files.
Cleaning	Explicitly cleaned native structures and logs.
Refolding	Self-consistency metrics, predictions, and aligned structures.
All tasks	Native configuration, standard output/error, and every generated result file.

U is Caliby's native Potts energy. It is not converted to a probability, confidence, or binding affinity. Local conditional arrays preserve the native token order and shape. Packing scn_rmsd is in angstroms; chi-angle mean absolute errors are in degrees and chi-angle accuracies are fractions. Refolding retains native sc_ca_rmsd, avg_ca_plddt, and tmalign_score values.

The pinned executable source does not implement the README's mentioned score_inputs_csv or save_potts_params options. Those options are unavailable in this version.