How to build a protein–RNA complex

A step-by-step guide: where to get the sequences, where to paste them, what to set in the options and how to tell whether the result can be trusted. AlphaFold 3 does the computing on Google's side — nothing is folded on this site.

Step 1. Get the protein sequence

Three sources, easiest first:

Our own constructs. What is actually expressed in the lab is in the table below — press “copy” and paste.
UniProt — for any known protein. Search by name and organism (ADAR2 human) and take the entry marked Reviewed, which is manually curated Swiss-Prot. Then Download → FASTA (canonical). The Family & Domains section gives the domain boundaries you need to cut out a single dsRBD.
NCBI Protein — when the organism is not in UniProt (that is our case for squid and hydra). Use Send to → File → FASTA.

Strip the FASTA header (the > line), spaces and line breaks before pasting — AlphaFold takes amino-acid letters only, and it will reject X and U (selenocysteine).

Proteins of this project: whole constructs first, then single domains.

What it is	aa

Step 2. Get the RNA sequence

The trap with RNA is grabbing an arbitrary stretch. Single-stranded RNA with no structure folds into a mess and there is nothing to check the prediction against. Take something that has an experimental answer:

PDB — the best source. Find a solved structure of your protein with RNA (for ADAR, see the table at the bottom), open the Sequence tab or download the FASTA: that is exactly the oligonucleotide that was crystallised. It also gives you something to compare the model against afterwards.
Rfam — when you need a family rather than one structure: hairpins, riboswitches, known motifs.
NCBI Nucleotide or a genome browser — when the substrate is a natural one. For ADAR1 those are the double-stranded stretches formed by inverted Alu repeats in 3′ UTRs.

Three rules. Letters A C G U only: turn the thymine of a DNA record into uracil. Replace modified nucleotides (written X in the PDB) with ordinary ones — in our presets 8-azanebularine is already replaced by the adenosine that sits in the edited position. Add a double-stranded RNA as two separate chains, not one string; a hairpin goes in as one chain and folds back on itself.

What it is	nt	source

Step 3. Where to go

Service	What it does	Cost and limits
AlphaFold Server the main route	Proteins, DNA, RNA, ligands and ions in one complex. This is AlphaFold 3 itself.	Free, Google sign-in. 30 jobs per day, complex up to 5000 positions, non-commercial use only.
AlphaFold DB check before you compute	Ready-made models of single proteins for almost all of UniProt.	Free and unlimited. No complexes with RNA — the protein alone.
Boltz-2, Protenix	Open models of the same class, protein together with RNA as well.	Free, but you need a card with 40–48 GB of memory. Worth it when 30 jobs a day is not enough.

How to save quota: look in AlphaFold DB first. If all you need is the fold of a single protein, it has already been computed there. The server is for complexes.

Step 4. What to set in the interface

Continue with Google. A Remaining jobs counter appears in the top right — how many jobs you have left today.
The first entity is already in the form. To the left of the input box is the type selector: Protein, DNA, RNA, Ligand, Ion. Set Protein and paste the sequence from step 1.
Copies — how many copies of this chain are in the complex. For a single dsRBD it is 1. Use 2 only when you are testing a dimer: the complex doubles in size.
Add entity adds a second chain. Choose RNA and paste the sequence from step 2. For a double-stranded duplex press Add entity once more and add the second strand separately.
Continue and preview job. In the preview window:
- Job name — name it like a human would, or you will not find it among a dozen jobs (say hADAR2_dsRBD1_hairpin13).
- Seed — the random seed. The same seed on the same input gives the same result; to check whether the prediction is stable, run the same input again with a different seed.
Confirm and submit job — the job goes into the queue. A domain with a hairpin takes minutes, a large complex tens of minutes.
Residue modifications, ligands and ions are not needed here — leave them alone.

If you would rather not click, the server accepts a ready JSON file (up to 100 jobs per file). You can build one at the bottom of this page.

Step 5. What comes back and how to read it

You download an archive. Inside are five models (…_model_0.cif … …_model_4.cif, number zero being the model's own favourite), a copy of your request (…_job_request.json) and the confidence files.

What to look at	Where	How to read it
pLDDT confidence per residue	the B-factor column of the .cif itself	> 90 — trust it; 70–90 — the fold is broadly right; 50–70 — doubtful; < 50 — usually a disordered stretch the model filled with something.
PAE error in relative position	a separate file, and a plot on the site	A dark square on the diagonal means the domain is folded confidently. A dark off-diagonal block between protein and RNA means their relative placement is confident too. A pale one means the chains are fine individually but the model does not know how they sit together.
ipTM interface quality	the summary in the results	The number that matters for a complex: > 0.8 — the contact is predicted confidently; 0.6–0.8 — a grey zone; < 0.6 — protein and RNA were probably just parked next to each other.
pTM	same place	The same idea for the whole structure, without separating the chains.

What to open it with. Mol* Viewer is the easiest — in the browser: Open Files, then colour by pLDDT. For real work use PyMOL or ChimeraX: that is also where you superpose the model onto an experimental structure (align model, 7zlq in PyMOL) and read the RMSD.

What is worth computing in this project

Job	size	why
One dsRBD + 13 nt hairpin	≈ 80	Start here: minutes to compute, and the answer is in the PDB (7ZLQ).
Two neighbouring dsRBD + 32 nt duplex	≈ 210	The main question of the project: do two domains sit on one RNA molecule, and how far apart.
The same, with domains from hydra and squid	≈ 210	The comparison: do domains from different organisms hold RNA the same way.
A whole N-terminal construct + RNA	400–950	Expensive and murky: long disordered linkers between domains, so the model places the domains wherever — and the PAE will say so honestly.

Before you submit, check: no FASTA header or stray characters in the protein; the RNA is written in ACGU; a double-stranded RNA went in as two chains; the copy number matches your intent; the job has a sensible name.

How many positions will the complex dsRBD1 of human ADAR2 + the 13 nt hairpin take? Get the domain length from the table in step 1.

Hint
Positions are amino acids plus nucleotides added together.
How many jobs per day does the free AlphaFold Server allow?
A model has high pLDDT on both the protein and the RNA, yet those numbers cannot tell you whether the protein sits on the RNA correctly. Which score does?

Hint
You need a score for the interface between chains, not for each chain on its own.

A ready job file instead of typing it in

Builds JSON in the AlphaFold Server format (dialect: alphafoldserver, version 1) — the site can import it instead of you adding entities by hand.

protein copies seed

Solved ADAR structures with nucleic acid — what to compare against

PDB	year	method	what it is	nucleotides