AfDesign

AfDesign webserver overview

AlphaFold is differentiable. AfDesign uses that to design a protein backwards: it starts from a random sequence, asks AlphaFold what it would fold into, and moves the sequence along the gradient of a loss such as high pLDDT, low PAE, many contacts, or a match to a backbone you supply. The result is a sequence and the structure AlphaFold predicts for it.

This page covers the AfDesign notebooks of the ColabDesign project (release v1.1.3) on one page with a task selector. Hallucination generates a protein of a chosen length. Fixed backbone design searches for a sequence that folds into a backbone you provide. Partial hallucination keeps positions of a structure and hallucinates the rest, with optional rewiring of the kept segments.

Disulfide design hallucinates a sequence around a random pattern of cysteine pairs. For the sequence design step that most pipelines use instead of fixed backbone design, see ProteinMPNN, and for cyclic peptides use AfCycDesign.

Pricing

AfDesign is billed by the minute of runtime: 15 credits per minute on an NVIDIA T4 GPU. Completed runs are charged in proportion to elapsed time, rounded up to a whole credit, with a minimum charge of 20 credits, so a run of up to 80 seconds costs 20. A run that fails is not charged, and unused reserved credits are returned. Set a spending limit before submission.

The quote before submission is an estimate from the task, the number of residues and the iteration counts. The final charge follows the measured runtime. The table uses the default iterations of each task: 100 pre-design, 50 soft, 50 temperature and 10 hard for hallucination, 50 soft, 50 temperature and 10 hard for disulfide design, and 300 soft, 100 temperature and 10 hard for fixed backbone design and partial hallucination.

Residues	Hallucination	Fixed backbone	Partial hallucination	Disulfide design
35	31	44	44	24 (default)
50	41	63	63	30
100	124 (default)	212	212	80
150	286	503	503	178
200	537	954	954	329
300	1,338	2,393	2,393	811

Residues are the Length for hallucination and disulfide design, the residues of the selected chain for fixed backbone design, and the Total length for partial hallucination, or the residues of the input structure when that is blank. Cost grows faster than the length, so a 200-residue hallucination costs about four times as much as a 100-residue one.

Each hard iteration costs five times as much as the others because it evaluates five AlphaFold models. Hallucination with Copies is quoted for the residues of all copies together, so two copies of 100 residues quote 537 credits, the same as one chain of 200. The loss weights, the contact settings, the partial hallucination positions and rewiring options, and Random seed do not change the quote.

Inputs

Input	Tasks	Accepted	Description
Input structure	Fixed backbone, Partial hallucination	.pdb, .ent, or a PDB ID fetched from RCSB, up to 50 MiB	The backbone to design for, or the structure whose positions are kept. Only the first model of the file is read.

Hallucination and Disulfide design take no structure.

Settings

Design

Parameter	Type	Default	Description
Task	enum	Hallucination	Hallucination, Fixed backbone, Partial hallucination or Disulfide design. Chosen with the Task cards above the inputs.
Chain	string	blank	Fixed backbone and partial hallucination. Chain to use, such as A. Blank uses every chain.
Length	integer	100	Hallucination. Number of residues.
Copies	integer	1	Hallucination. Number of identical copies of the sequence; more than 1 designs a homo-oligomer.
Positions to constrain	string	blank	Partial hallucination. Positions of the input structure to keep, in its numbering, such as 4-9,13-26. Blank keeps every position.
Total length	integer	blank	Partial hallucination. Length of the designed protein when it differs from the input.
Fix the sequence	boolean	off	Partial hallucination. Keeps the sequence of the constrained positions.
Use templates	boolean	off	Partial hallucination. Constrains the positions with the template input.
Segment order	string	blank	Partial hallucination. Zero-based order of the kept segments after rewiring, such as 1,2,3,0,7,4,5,6.
Loop lengths	string	blank	Partial hallucination. Loop length between each pair of neighbouring segments, such as 3,3,3,4,3,3,4.
N-terminal offset	integer	blank	Partial hallucination. Loop length at the N terminus.
Start from the input sequence	boolean	off	Partial hallucination. Initialises the kept positions with the input sequence.
Length	integer	35	Disulfide design. Number of residues.
Disulfide pairs	integer	3	Disulfide design. Cysteine pairs placed at random, more than five residues apart.
Random seed	string	blank	Whole number from 0 to 4294967295. Blank draws a new seed each run.

Rewiring happens only when Segment order, Loop lengths or N-terminal offset is set. Leaving all three blank keeps the segments where the input structure has them.

Hallucination loss

Parameter	Type	Default	Description
PAE weight	number	0.1	One of 0.01, 0.1, 0.5, 1.0.
pLDDT weight	number	0.1	One of 0.01, 0.1, 0.5, 1.0.
Contact weight	number	1.0	One of 0.01, 0.1, 0.5, 1.0.
Minimum sequence separation	integer	9	One of 0, 5, 9. Only contacts between residues at least this far apart in sequence count.
Contact cutoff	enum	14 Å	8 Å, 14 Å or Maximum (21.6875 Å), measured between Cβ atoms.
Contacts per position	enum	2	1, 2, 4, 8 or Maximum (the sequence length).
Binary contacts	boolean	off	Counts a contact as present or absent.

Minimising PAE or maximising pLDDT alone often produces a single helix, which is why the notebook optimises a number of contacts per position instead.

Partial hallucination weights

Parameter	Type	Default	Description
Restraint weight	number	1.0	Weight of the distogram restraint on the kept positions.
Hallucination weight	number	1.0	Weight of the contact loss that hallucinates the rest. The notebook example sets 0.

Optimization

Parameter	Type	Default	Description
Pre-design iterations	integer	100	Hallucination. Softmax pre-design from a Gumbel-initialised sequence.
Soft iterations	integer	50 for hallucination and disulfide design, 300 for fixed backbone and partial hallucination	Iterations that turn the sequence logits into a softmax.
Temperature iterations	integer	50 for hallucination and disulfide design, 100 for fixed backbone and partial hallucination	Iterations that sharpen the softmax toward one-hot.
Hard iterations	integer	10	Iterations with one-hot sequences. The result comes from the lowest-loss iteration of this stage.

Outputs

File	Description
hallucination.pdb, fixbb.pdb or partial.pdb	The best structure, with pLDDT in the B-factor column. Disulfide design writes hallucination.pdb.
sequences.fasta	The designed sequence.
scores.json	The model's log for the structure you received, with the designed sequence.
trajectory.json	The log of every optimisation step.
trajectory.png	Fixed backbone and partial hallucination. RMSD to the input and sequence identity over the run.
animation.html	The trajectory as a video, except for disulfide design.
disulfide_pattern.json	Disulfide design. The cysteine pairs chosen at random (zero-based) and the sequence pattern that seeded the design.
provenance.json, run.log	Settings, source release, library versions and the run log.

Understanding results

Reading the scores

The scores in scores.json keep the library's units.

• plddt runs from 0 to 1. Multiply by 100 for the usual pLDDT scale.
• pae is the predicted aligned error as a fraction of AlphaFold's 31 angstrom maximum. Multiply by 31 for angstroms, and see PAE for how to read it.
• ptm is the predicted TM-score, a score for the arrangement of the whole chain.
• con is the contact loss, which the optimiser lowers to pull residues into contact.
• loss is the weighted sum the run minimised. It changes with the task and the weights, so compare losses only between runs with the same settings.
• rmsd and dgram_cce appear for fixed backbone design and partial hallucination. They measure how closely AlphaFold's output follows the structure you supplied.

pLDDT is a local score, while pae and ptm describe how the parts of the chain sit relative to each other. A design needs good values in all three. The standard pLDDT bands apply: above 90 (0.9 in scores.json) is very high confidence and 70 to 90 is confident. A hallucinated sequence is optimised to raise pLDDT, so a high value there is expected and is weaker evidence than the same value on a sequence AlphaFold was not tuned for.

Worked example

Hallucination of 100 residues with default settings and random seed 0 returns a protein with plddt 0.92, pae 0.137, which is 4.2 angstroms, ptm 0.72 and loss 1.66. With the default weights, the loss is 0.1 * pae + 0.1 * (1 - plddt) + 1.0 * con. The contact term con is 1.63 and makes up almost all of the loss, while the pLDDT and PAE terms add 0.02 between them. The pLDDT is in the very high band, and the mean PAE of 4.2 angstroms and pTM of 0.72 say that the arrangement of the chain is less certain than its local geometry.

Common misreadings

• A hallucinated protein has no target, so a high plddt and a low pae only say that AlphaFold is confident in it, not that it folds.
• The notebooks warn that with AlphaFold as the only loss, the search can find adversarial sequences that AlphaFold likes and that do not fold. Run the designed sequence through an independent predictor such as ESMFold before ordering it.
• For fixed backbone design and partial hallucination, a low rmsd shows that AlphaFold reproduces your structure from the designed sequence, and it does not require the sequence to resemble the native one.
• Disulfide design draws its cysteine pattern with the notebook's own random choice, which the random seed does not control. The seed fixes the starting sequence and the sampling of the optimisation, and disulfide_pattern.json records the pairs each run used. The notebook's optional PyRosetta side-chain refinement is not part of this tool.

Sources

• ColabDesign, release v1.1.3, notebooks design.ipynb, hallucination.ipynb, partial_hallucination_rewire.ipynb and disulfide_design.ipynb. Code under the Beer-Ware licence.
• Jumper J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583 (2021). The AlphaFold parameters are released under CC BY 4.0.