DiffAb

DiffAb webserver overview

The DiffAb webserver runs the pinned c3e2966 release to redesign antibody complementarity-determining region (CDR) sequences, structures, or both. It accepts an antibody structure with optional antigen chains and supports independent CDR co-design, coordinated multi-CDR co-design, native-start optimization, fixed-backbone sequence design, and fixed-sequence structure prediction.

Each job returns the generated PDB structures, reference structures, a CDR-sequence table, the exact inference configuration, native metadata, and downloadable files. The sections below cover input preparation and how to read the returned files.

Pricing

DiffAb jobs start at 50 credits. The exact quote is calculated from the selected mode, CDRs, optimization steps, and sample count before submission.

These representative quotes use the default three CDRs (HCDR1, HCDR2, and HCDR3), five samples, and optimization step 4 where applicable:

Design mode	Credits
Co-design selected CDRs independently	120
Co-design selected CDRs together (sequence + structure)	50
Optimize selected CDRs (sequence + structure)	50
Fixed backbone for selected CDRs (sequence only)	120
Predict selected CDR structures (fixed sequence)	120

Coordinated multi-CDR co-design scales with sample count. Independent co-design, fixed-backbone design, and structure prediction scale with sample count and selected CDR count. Optimization also scales with the selected optimization step totals. Chain IDs, seed controls, and renumbering do not change the quote.

Inputs

Input	Type	Required	Description
Antibody Structure or Antibody-Antigen Complex (PDB)	PDB	Yes	One .pdb or .ent file up to 50 MiB, uploaded directly or fetched from RCSB. Multi-model files use the first model. An antibody structure is required; antigen chains are optional.
Job name	string	No	Optional label used to identify the job in history.

Preparing the input

The structure must contain an antibody or nanobody. An antigen-only PDB does not run, because the webserver redesigns CDRs on an existing framework and does not build one from an antigen. Add the antigen chains that form the interface when the redesign should be conditioned on the binding context.

As in the published model (Luo et al. 2022), DiffAb redesigns residues that already exist in the selected CDR, so each loop keeps its length. A partly unresolved HCDR3 can fail detection or give an incomplete design problem. PDB Fixer can add missing atoms and residues, but a rebuilt loop is modeled rather than observed, so record the repair and inspect it before submitting.

Chain IDs are single characters and must match the identifier on the PDB ATOM records, including capitalization. Leave them blank only when the file holds one clear antibody. With several antibodies or symmetry-related copies, enter the heavy and light chains of the Fv that contacts the antigen you want to keep.

For a nanobody, enter the heavy chain, leave the light chain blank, and select only heavy-chain CDRs. A selected loop that is absent from the detected antibody produces no variant, and a job with no matching loop fails with a "no requested CDR variants" message. If the numbering of an unusual variable domain is in doubt, ANARCII can renumber the structure before design.

Settings

Core settings

Parameter	Type	Default	Description
Heavy chain ID	string	Auto-detect	Optional one-character antibody heavy-chain identifier. Leave blank for automatic antibody-chain detection and Chothia renumbering.
Light chain ID	string	Auto-detect	Optional one-character antibody light-chain identifier. Leave blank for automatic detection or for a heavy-chain-only nanobody.
Design mode	enum	Co-design selected CDRs together (sequence + structure)	Select independent sequence-and-structure co-design, coordinated multi-CDR co-design, native-start optimization, fixed-backbone sequence design, or fixed-sequence structure prediction.
CDRs to design	enum	HCDR1, HCDR2, HCDR3	Select one or more of HCDR1, HCDR2, HCDR3, LCDR1, LCDR2, and LCDR3. A selected loop absent from the detected antibody produces no variant for that loop.
Optimization steps	enum	4	Visible only for Optimize selected CDRs (sequence + structure). Select one or more native-start diffusion lengths from 1, 2, 4, 8, 16, 32, and 64; each selected CDR and step combination creates a separate variant.
Number of samples	integer	5	Generates 1 to 10 samples for each native variant.

Advanced settings

Parameter	Type	Default	Description
Use a random seed	boolean	false	Generates a new seed for the job. Enabling the switch hides the reproducible seed field below.
Random seed	integer	2022	Reproducible seed from 0 to 4294967295. Visible when Use a random seed is disabled.
Input is already Chothia-numbered	boolean	false	Skips DiffAb's default AbNumber and ANARCI renumbering. Enable only for a correctly Chothia-numbered structure and provide at least one antibody chain ID because source chain auto-detection is disabled in this mode.

Outputs

Artifact	Format	Description
Generated structures	PDB	Numbered designs such as H_CDR3/0000.pdb, grouped by CDR or optimization variant and shown in the structure viewer.
Prepared inference input	PDB	Exact Chothia-renumbered input used for inference when automatic renumbering is applied.
Reference structures	PDB	Native reference.pdb and variant reference files such as REF1.pdb, kept separate from generated designs.
CDR sequence table	CSV	cdr_sequences.csv with the design name and extracted HCDR1, HCDR2, HCDR3, LCDR1, LCDR2, and LCDR3 sequences.
Generated structure archive	TAR.GZ	generated.tar.gz containing only the numbered generated PDB structures.
Native run metadata	JSON	metadata.json describing the DiffAb variants and native run.
Inference configuration	YAML	diffab_config.yaml containing the exact configuration used for inference.
Execution log	TXT	log.txt when the native DiffAb run produces a log.

Understanding results

The Viewer tab displays generated, prepared-input, and reference PDB files. The CDR sequences tab lists one row per generated structure with columns for the six antibody CDRs, and the Files tab provides every returned artifact for download.

Variant folders identify the CDR or optimization branch that produced a structure. Four-digit sample numbers record generation order only. DiffAb does not return a confidence, affinity, energy, or scientific ranking score.

Samples and reference structures

Only the numbered files, such as H_CDR3/0000.pdb, are generated designs. reference.pdb and the per-variant REF1.pdb preserve the native structures used by the run, so they should not be counted as extra candidates. Multi-CDR output is grouped under its own variant folder, and optimization folders also carry the step count.

Auditing the CDR sequence table

The table lists one row per generated structure, with the sequences read back from that PDB. It has no row for the input antibody, so compare each row with the CDR sequences of your starting structure to see what actually changed.

For independent co-design, fixed-backbone design, and optimization, each variant should change only its named CDR. For fixed-sequence structure prediction, that CDR should stay identical to the input. In coordinated multi-CDR mode, check every selected CDR column. A blank cell means the chain or loop was absent or could not be numbered, not that the loop has no residues.

Files worth keeping

Keep diffab_config.yaml with every shortlisted structure, because it records the exact mode, CDR list, sample count, and seed. metadata.json describes the detected variants, and input_chothia.pdb is the exact renumbered input, present only when renumbering was not skipped. A PDB without its source structure, chain mapping, and configuration is hard to reproduce. generated.tar.gz holds the numbered designs only, and log.txt is the native run log.

Choosing candidates

Samples are unranked, so 0000.pdb is the first sample generated and not the best one. Filter them with separate tools: MolProbity for backbone and clash geometry, BioPhi for humanness, and TAP2 for developability flags. Binding and function need experimental measurement.

Sources

• Luo S, Su Y, Peng X, Wang S, Peng J, Ma J. "Antigen-specific antibody design and optimization with diffusion-based generative models for protein structures." Advances in Neural Information Processing Systems 35 (NeurIPS 2022). proceedings.neurips.cc