Can I run Proteina-Complexa online without a GPU cluster?

Yes. Ranomics Tools runs Proteina-Complexa as a wallet-funded full-size run of independent search shards on dedicated A100-80GB GPUs. Upload your protein target (or pick a curated benchmark task), choose how many designs you want, and the run is split into pieces that run on our GPUs and bill as they finish. You only pay for compute that runs, and the run pauses if your balance runs low.

Can Proteina-Complexa design binders against my own small molecule?

No. The only target you can supply is a protein structure (.pdb or .cif), on the protein-binder variant, scored by AlphaFold2 confidence. The ligand-binder and motif variants design against benchmark tasks bundled with the model rather than anything you upload, so a molecule of your own is refused before any GPU runs.

Proteina-Complexa

Proteina-Complexa is a no-install online hard-target binder design tool you can run through tools.ranomics.com on a dedicated GPU. Built for the protein targets the standard design tools stall on — a recessed pocket, or a site spanning two chains: every candidate the search generates is re-folded against your target before the search builds on it.

Example run — A two-chain human secreted protein, about 210 residues per chain, from a solved crystal structure. See what came back →

Design inputs

Design variant

Which model variant to run. Only the protein binder can design against a structure you upload; the ligand and motif variants run curated benchmark tasks.

Target structure

Upload the target as .pdb, .cif, or .mmcif. Leave empty to design against a curated benchmark task instead.

Target task

A curated benchmark task. It comes with its own target and its own design settings, so the fields below do not apply. Blank uses the variant's default task.

Target chain

The chain to design against. For several chains, use the target region field below instead.

Target region

Which chains and residues to design against. One chain (A1-150) or several (A12-157,B12-157,C12-157) for a multi-chain target. Blank uses every residue of the target chain above.

Hotspot residues

Optional. The numbered residues on the patch of your target you want the binder to sit on, written in your PDB’s own numbering. A plain number is read as the target chain; when the run targets more than one chain it is refused, because 241 cannot say which protomer it means — prefix the chain instead (A113,C73). Every one is checked against your structure before any GPU runs, so a typo is refused rather than silently ignored. Click residues in the viewer to toggle them.

Not sure which residues? Score your target's surface with Epitope Scout first.

3D residue picker

Surface

Upload a target structure above to enable click-to-pick. Typed entry in the Hotspot residues field also works.

Binder length (residues)

How long the new binder should be, in residues. Every design draws its own length from this window, so a wide window buys variety and a narrow one buys consistency. Match it to the face you are aiming at: 50 to 80 for a small, fairly flat patch, 100 to 200 for a broad one or a site spanning two chains, where a short binder cannot reach across. Longer binders also use more GPU time per design and are harder to express at the bench. Leaving both blank is not unconstrained — it means 60 to 120, the model's own default. 20 to 300 allowed.

Number of designs

One search shard yields up to 8 designs on a single A100-80GB. Larger requests become a wallet-bounded full-size run of independent shards that run on our GPUs and bill as they finish, with a global top-K across all of them.

Group

Group related runs on your jobs page. Leave blank to file under Uncategorized.

You pay for the GPU time your run uses. Cancelling does not refund GPU time already used.

About Proteina-Complexa

What it is

Designs binders for the targets the standard tools find hard: a recessed pocket, or a site spanning two chains. Rather than generating candidates and hoping, it searches — it generates, re-folds every candidate and scores how well it grips your target, keeps what scores well and generates again from there. Which model does that scoring follows the target: a protein target is scored by an AlphaFold2 refold, a benchmark ligand or motif task by RoseTTAFold3, with a physics force field added where it applies. The run splits into independent shards across as many GPUs as your balance funds — each shard is a separately seeded search — and the designs are ranked globally across all of them into one pooled table. Proteina-Complexa, Didi et al., ICLR 2026.

When to use

You want to aim a binder at one specific patch, including a recessed or partly shielded one, by naming residues on it.
The site you care about spans more than one chain of your target.
You would rather pay for a search that filters as it goes than for raw generation you have to filter afterwards.
You want to scale the search across many GPUs, with your prepaid balance as the only ceiling.

Prerequisites

A target: your own structure (.pdb/.cif) for the protein-binder variant, or a curated benchmark task for any variant.
For your own target, the chain ID — or a chain/residue range such as A1-150, or A12-157,B12-157,C12-157 for a multi-chain target.
Optionally, hotspot residues to aim the binder at a specific epitope.
A funded wallet that covers at least the first wave of shards.

Input glossary

Design variant. protein_binder for a protein target (AF2 reward) and the only variant that can take a target of yours, ligand_binder for a bundled benchmark ligand task (RF3 reward), motif_ame for motif scaffolding / enzyme active sites, or validate for a free config check before spending GPU.
Target. Your own structure, or a curated benchmark task whose target is baked in — the two are mutually exclusive. Uploading your own is available on the protein-binder variant; the ligand and motif variants run curated tasks (their tasks resolve from separate upstream registries).
Target region. Which chains and residues to design against, e.g. A1-150, or A12-157,B12-157,C12-157 for a multi-chain target. Blank uses the whole target chain.
Hotspot residues. Optional. Residues the binder should contact, in original PDB numbering — plain numbers use the target chain, or prefix the chain (A113 C73) for a multi-chain region. Every hotspot is checked against your structure before any GPU runs, so a residue that is not there is refused rather than quietly ignored.
Binder length. The range each design's length is drawn from. Defaults to 60-120 residues.
Number of designs. How many designs to search for. This scales the number of independent search shards rather than the width of one: each shard is a fixed 8-design search on its own GPU and returns up to 8, and the hub picks the global top set across every shard.

Ranked designs with reward scores (AF2 pLDDT / ipTM for protein, RF3 score for ligand / motif, force-field energy where applicable), a self-consistency re-fold RMSD, and downloadable structures. The ligand and motif variants score on RF3 only.

References

Trial run: one shard, 8 designs

Run the smallest complete unit of a Proteina search against your own target and see how the designs score against it.

What you need

A structure file for your target and its chain ID — or a chain and residue range such as A1-150. Hotspot residues are optional, and are what aims the binder at one specific face.

What it costs

You pay for the GPU time your run uses.

Then what

8 designs is one shard on one GPU, and a shard is a whole container — asking for fewer returns less from the same container. Raise the count and it becomes a full-size run bounded by your wallet, split into pieces that run on our GPUs and bill as they finish, with a single ranked list pooled across every shard. Widen the length window once you know the target is workable.

Load these settings →

A past run we actually did

The target

A two-chain human secreted protein, about 210 residues per chain, from a solved crystal structure. Eighteen hotspot residues, nine on each chain, across the two-fold interface where the chains meet.

The site sits on a symmetry axis, so a binder has to reach both chains at once rather than settle on either one. It is also a target whose own sequence the generator already knows well, which turns out to be what this run is worth reading for.

What we put in the form

Target structure — the two-chain crystal structure, chains A and B: Both chains staged and both named as target chains. Give one chain and the model designs against half a site that does not exist on its own.
Hotspot residues — 18 residues, 9 per chain, each written with its chain letter: The chain letter is required once a run names more than one target chain, and the form refuses a bare 241 there. It has to: the model matches hotspots literally as chain plus number, and on a homodimer whose two protomers share one numbering a bare token would resolve to chain A alone — a run that looks symmetric and is not. Writing A241 and B241 is what makes the ask symmetric.
Binder length (residues) — 60 to 69: A window, not a number — the generator draws a length per design. Leaving the field empty gives you 60 to 120; this is the bottom of that default, where binders are cheaper to fold and easier to order later.
Number of designs — 64: Type 64 here and the run splits into eight shards of 8 — not one shard of 64. Every shard the form can launch is a fixed 8-design search and this box buys shards, so the ranked list comes back pooled across all eight. The run below is a single 64-design container, because it came from the internal sweep driver, which widens one shard to 16 samples with 4 replicas each; the form exposes neither control. That width is also why how long it ran says nothing about what an 8-design shard launched from this form does. The settings are reproducible; the single-container shape of the table below is not.
Design variant — Protein binder (de novo, vs a protein target): The only variant that designs against a structure you upload; the others run curated ligand and motif benchmarks. It is also why the table below has no rf3_score column: RF3 is a second scoring stack those other variants need, a protein binder run does not, and it is not free. That is a consequence of this choice rather than a switch of its own — there is no RF3 control on the form.

What came back

64 designs, of which 12 passed — ipTM at or above 0.80 with the re-folded complex landing within 5 Å. They are ranks 1 to 11 and 13, so on this shard the ranking and the filter agree: the top of the table is the answer. Rank 12 is the one row you cannot settle from the table by eye. It prints the same 0.80 as rank 13 and re-folds just as tightly (1.26 Å against 1.24 Å), and it is out all the same: the column rounds to two places, and the 0.80 line is drawn on the stored ipTM — rank 12 falls just short of it where rank 13 just clears. The best scored ipTM 0.89 at pLDDT 88.50, re-folding 1.32 Å from where the generator put it. Of the 52 that did not pass, 30 failed on the re-fold. Twelve of those 30 are the model handing the target's own sequence back as the binder — literal copies of a stretch of the chain it was asked to bind.

Candidates (64)

#	Reward ?	af2_iptm	af2_plddt	binder_scrmsd ?	3D	Structure
1 Top	-0.18	0.89	88.50	1.32	—	—

2	-0.19	0.88	85.69	1.50	—	—

3	-0.19	0.87	86.24	1.03	—	—

4	-0.21	0.86	84.35	1.40	—	—

5	-0.22	0.84	84.50	1.07	—	—

6	-0.23	0.83	85.38	1.94	—	—

7	-0.23	0.82	82.47	1.67	—	—

8	-0.24	0.83	83.55	1.62	—	—

9	-0.25	0.83	80.74	1.37	—	—

10	-0.27	0.82	80.21	0.80	—	—

11	-0.27	0.83	76.31	2.20	—	—

12	-0.27	0.80	78.79	1.26	—	—

13	-0.28	0.80	74.35	1.24	—	—

14	-0.29	0.78	76.51	3.19	—	—

15	-0.29	0.76	76.79	2.02	—	—

16	-0.29	0.77	78.92	0.86	—	—

17	-0.31	0.75	72.79	2.43	—	—

18	-0.31	0.76	71.39	1.82	—	—

19	-0.33	0.73	73.44	2.21	—	—

20	-0.33	0.73	74.22	1.23	—	—

21	-0.33	0.72	74.32	1.78	—	—

22	-0.34	0.72	72.09	0.74	—	—

23	-0.36	0.70	73.69	1.36	—	—

24	-0.37	0.67	71.30	2.00	—	—

25	-0.38	0.69	67.00	1.66	—	—

26	-0.38	0.68	71.32	1.44	—	—

27	-0.38	0.68	69.30	1.99	—	—

28	-0.40	0.66	71.41	20.48	—	—

29	-0.43	0.61	63.28	1.51	—	—

30	-0.43	0.65	68.90	1.67	—	—

31	-0.44	0.60	67.77	1.40	—	—

32	-0.53	0.48	59.28	1.07	—	—

33	-0.55	0.46	53.49	3.23	—	—

34	-0.59	0.42	51.29	2.18	—	—

35	-0.61	0.41	53.93	2.93	—	—

36	-0.88	0.15	58.88	45.94	—	—

37	-0.89	0.15	64.01	42.70	—	—

38	-0.89	0.13	51.77	42.34	—	—

39	-0.90	0.13	58.48	10.15	—	—

40	-0.90	0.13	55.85	39.33	—	—

41	-0.91	0.12	57.55	48.32	—	—

42	-0.91	0.12	61.31	46.87	—	—

43	-0.92	0.12	59.24	42.08	—	—

44	-0.92	0.11	52.74	47.65	—	—

45	-0.93	0.11	60.92	45.10	—	—

46	-0.93	0.11	59.19	45.94	—	—

47	-0.93	0.11	55.24	21.08	—	—

48	-0.93	0.10	48.50	41.63	—	—

49	-0.93	0.11	58.71	45.67	—	—

50	-0.94	0.11	53.83	45.87	—	—

51	-0.94	0.10	74.24	38.48	—	—

52	-0.94	0.10	74.04	43.10	—	—

53	-0.94	0.09	75.46	42.75	—	—

54	-0.94	0.10	74.28	40.35	—	—

55	-0.94	0.10	58.38	32.37	—	—

56	-0.94	0.10	74.89	40.16	—	—

57	-0.95	0.09	74.58	44.05	—	—

58	-0.95	0.10	75.65	43.74	—	—

59	-0.95	0.09	71.14	38.64	—	—

60	-0.95	0.10	42.38	27.68	—	—

61	-0.95	0.09	70.79	32.12	—	—

62	-0.95	0.09	72.76	43.61	—	—

63	-0.95	0.09	71.42	39.10	—	—

64	-0.96	0.09	73.82	43.06	—	—

Validating in the lab is optional Designs from your own run are yours to download and take anywhere. If you do want them made and measured, Ranomics runs yeast display, mammalian display, BLI and DMS on the same data loop.

Test binders in the lab AI Binder Sprint

How to read that

Read binder_scrmsd before you trust af2_plddt. The twelve copies score pLDDT 70.79 to 75.65 — unremarkable, nothing that reads as broken — while their binder_scrmsd is 32 to 44 Å against a median of 2.0 Å for everything else. That column is the generator's backbone compared against an independent re-fold of its own sequence, and a target fragment re-folds into the target's shape rather than the shape it was drawn as. ipTM agrees (the twelve print 0.09 or 0.10, against a 0.67 median for the rest) and total_reward sends all twelve to ranks 51 to 64, so reading the top of the table is safe. Reading the whole table is what needs this paragraph. The scores shortlist; they do not diagnose. Thirteen rows here carry that profile — re-fold past 30 Å at ipTM under 0.10 before rounding — and twelve of them are copies. The thirteenth, at rank 55, is an ordinary de-novo sequence that shares 4% of itself with the target and simply failed to re-fold. Its pLDDT is 58.38, below the band the twelve sit in, but that gap is one shard's worth of evidence and not a rule. We know which twelve because we compared each sequence against the target's, which takes a substring search and no GPU. If a row of yours looks like this, that comparison is the check that settles it.

What happened next

Scaled the same settings out to 17,024 designs across four rounds. Target self-copies ran at 29% of everything generated and not one of them ever cleared ipTM 0.80, so the score gate that was already there kept them out of every shortlist. That is the reassuring version: on this target the failure mode is loud and it is caught. It is worth a look at your own output anyway, because nothing in the tool is checking for it. One note on price. A run settles on the GPU seconds it actually used, not on a flat per-run figure, so you pay for the GPU time your run uses.

Recorded from that one past run. That is what it did, not a prediction for yours.

What good looks like

Each tool reports a subset of these depending on whether it does design, sequence recovery, or structure prediction.

ipTM: Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better: 0.65 to 0.75 depending on the tool. Individual tools set their own pass bar a little either side of that. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's.
pLDDT: Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
i_pAE and pAE: Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
ProteinMPNN recovery: Fraction of native residues recovered when ProteinMPNN redesigns a known sequence on its native backbone — the folded shape on its own, before a sequence is chosen for it. Higher is better: > 0.4 usable; > 0.6 excellent.

Questions people ask about Proteina-Complexa

Can I run Proteina-Complexa online without a GPU cluster?: Yes. Ranomics Tools runs Proteina-Complexa as a wallet-funded full-size run of independent search shards on dedicated A100-80GB GPUs. Upload your protein target (or pick a curated benchmark task), choose how many designs you want, and the run is split into pieces that run on our GPUs and bill as they finish. You only pay for compute that runs, and the run pauses if your balance runs low.
Can Proteina-Complexa design binders against my own small molecule?: No. The only target you can supply is a protein structure (.pdb or .cif), on the protein-binder variant, scored by AlphaFold2 confidence. The ligand-binder and motif variants design against benchmark tasks bundled with the model rather than anything you upload, so a molecule of your own is refused before any GPU runs.
How are Proteina-Complexa designs scored and ranked?: Every candidate is re-folded and scored against your target as it is generated, and which model does that scoring follows the target: a protein target is scored by an AlphaFold2 refold, a benchmark ligand or motif task by RoseTTAFold3, with a physics force field added where it applies. Each shard keeps what scores well, and the hub then ranks across every shard at once, so one pooled table shows the best-scoring designs from the whole run rather than one table per GPU.