RxCubed Tutorial: Reading a Protein–Peptide Interface

RxCubed Tutorial: Reading a Protein–Peptide Interface

This is one of RxCubed's two guided tutorials, and you can start with either. The Foundations tutorial teaches the platform through one small molecule and one pocket; this one teaches it through the thing that tutorial cannot reach: a protein–peptide interface — what holds it together, how tightly, and what it would take to design a better binder for it.

You don't need the other one first. Everything you need to know about the platform — searching, saving evidence, the viewers, running a tool, Results Hub, writing a report — is taught here as you go. If you have done the Foundations tutorial, steps 1 and 2 will feel familiar; skim them.

Before you start

Open this tutorial from the Learning Hub with Start tutorial. That gives you a learning workspace of its own: a practice project, separate from your research projects and from any other tutorial's workspace, with this walkthrough's checklist on its Dashboard. Each step ticks itself off when the evidence it asks for exists, and you can restart the whole thing at any time.

The scenario

Every step below works toward one real research question:

What holds the p53 peptide onto MDM2, can the platform tell that a known better binder is better, and what would a docking run built on what I found actually show?

This is an educational analysis of two published, experimentally solved complexes — not a discovery, a validation, or a treatment recommendation.

The science, in three paragraphs

p53 is the guardian of the genome, and MDM2 is its off switch. When a cell is damaged, p53 halts the cell cycle or triggers apoptosis. MDM2 binds p53's transactivation helix and marks it for destruction, keeping the brake off in healthy cells. A large fraction of cancers never mutate p53 at all — they amplify MDM2 and suppress a perfectly good p53. Block that one interaction and the patient's own wild-type p53 comes back on. That idea produced the nutlins on the small-molecule side, and stapled peptides that reached clinical trials on the peptide side.

The entire interaction runs through three residues. p53's Phe19, Trp23 and Leu26 each drop into their own hydrophobic subpocket on MDM2, like three pegs into three holes. It is the textbook example of a hotspot — the small part of a large interface that does most of the binding. You are going to find those three residues from the structure alone, without being told where to look.

And there is a built-in answer key. The same MDM2 pocket has been solved with PMI (TSFAEYWNLLSP), a 12-residue peptide engineered to bind far more tightly than the natural p53 sequence — around 135-fold tighter, at Kd 3.3 nM (Pazgier et al., PNAS 2009). Two structures, one pocket, two peptides of known and very different affinity. That lets you ask something the Foundations tutorial could not: does the platform rank them correctly? The answer turns out to be the most useful thing here.

What you'll learn

• How to find a paper, collect structures, run a tool and find its result — the RxCubed workflow, from the first search to a saved report.
• How to turn a complex into a list of contacts, and a list of contacts into a hotspot.
• How to read an interface three ways — geometry, chemistry, and burial — and why no one of them is enough alone.
• Where a binding-affinity predictor stops being trustworthy, how to notice that from its own output, and what to do about it.
• How an interface analysis becomes the restraints a docking run needs, and how to judge a docking result once you have one.
• Which parts of this work RxCubed does locally and which are honestly external.

Links to RxCubed pages open in a new tab so you can keep this tutorial open.

Quick Start

Seven steps make up the tutorial. Start with the action in the table, then use the detailed instructions below when you want more guidance.

Every numbered step runs on RxCubed's own machines in a second or two and costs you nothing. The docking run is a worked example you read rather than run — why.

The tutorial is pinned to these specific records. If a live search doesn't surface them, or ranks something else first, use these directly:

• the open-access paper Structural basis for high-affinity peptide inhibition of p53 interactions with MDM2 and MDMX (Pazgier et al., PNAS 2009), which reports the affinities step 5 checks;
• RCSB PDB entry 1YCR — MDM2 with the natural p53 peptide, downloaded as mmCIF; and
• RCSB PDB entry 3EQS — the same MDM2 pocket with PMI, the engineered peptide, also as mmCIF.

🧭 Step 1 — Find the paper that measured it

Under Find papers, open Literature Search and search for “PMI peptide MDM2 MDMX high affinity”. Add a Quality: “Reports a measured binding affinity for a peptide inhibitor of MDM2”.

A Quality is not a keyword. Keywords decide what is retrieved; qualities describe what would make a retrieved paper good for you, and the assistant scores each result against them. That is why the two boxes are separate. Sort by AI score and open a result's AI summary to see, in a sentence or two, how well it matches the quality you asked for.

Shortlist, then save

Results are never saved directly. Add to shortlist, choose the file types you want, then Save shortlist. A shortlist row is an intention, not a file — nothing is fetched until you save.

Save Structural basis for high-affinity peptide inhibition of p53 interactions with MDM2 and MDMX. It reports the crystal structure of MDM2 with PMI that you will collect next, and the affinities step 5 holds the platform to.

Results Hub

Open Results Hub. Everything in a project lives here, organised as Category → Tool → Run → Evidence: your search is a run, and the paper you saved is evidence under it. That shape never changes — a literature search, a contact map and a report are all runs with outputs.

✦ Ask AI

On the paper's information box in Results Hub, click Ask AI:

"What does this paper claim made PMI bind MDM2 so much more tightly than the natural p53 peptide, and which of those claims could I check from the crystal structures alone?"

Keep the answer in mind — steps 4 to 6 check some of it.

🧬 Step 2 — Collect both complexes

Under Collect files, open RCSB PDB, search for each of these, and save both as Coordinates (mmCIF), the default:

• “1YCR” — MDM2 with the natural p53 peptide. Chain A is the MDM2 N-terminal domain (85 ordered residues); chain B is the p53 transactivation peptide, 13 residues modelled.
• “3EQS” — the same MDM2 pocket with PMI, the engineered peptide. Chain A is MDM2 (84 ordered residues); chain B is PMI, 11 of its 12 residues ordered.

On residue numbering. Every residue number in this tutorial is the one the PDB entry itself uses, and it matches p53's own canonical numbering — so Phe19 here is Phe19 in the literature. You get the same numbers whether you save mmCIF or legacy PDB.

Open 1YCR from Results Hub in the structure viewer. This is close to the smallest genuine peptide–protein complex in the PDB, and that is exactly why it was chosen: every computation below is proportional to it, so each one takes a second or two.

Rotate it until you can see the peptide lying in a groove on the surface of the domain. It is not buried in a pocket — it is a short helix lying along an open channel. That is the whole reason peptides are a separate modality from small molecules: there is far more contact area, and far less of a deep pocket to grip.

✦ Ask AI

"This peptide binds in a shallow surface groove rather than a deep pocket. What does that imply about how hard it would be to find a small molecule that does the same job, and why did anyone bother trying?"

🔗 Step 3 — Map the interface

Under Model, open Tool Selector and choose Contact mapper (geometric). Run it on 1YCR, with:

• Target chain: B — the peptide
• Partner type: Chain
• Partner ID: A — MDM2
• Cutoff: leave at 4.0 Å

Every run's result is written to Results Hub automatically, under the tool that produced it. You get 11 contact residues — every peptide residue within 4 Å of MDM2. Only Asp21 and Lys24 stay clear. That is the answer to "where do these two things touch", and it is most of the peptide.

Note the UniProt accession column: P04637, with each residue's position in human p53's own canonical sequence. The residue numbers here are the same ones the literature uses, not a crystal-specific numbering you would have to translate.

Distance alone does not find the hotspot

It is tempting to sort by Closest distance and call the closest residues the important ones. Try it, and see why that fails:

The shortest contact belongs to Asn29, at the peptide's frayed C-terminal end, through a backbone hydrogen bond. Leu26, which the literature counts among the three residues that carry this interaction, is tenth. A closest distance says two atoms are near each other. It does not say how much of a residue is committed to the interface — one short hydrogen bond from a backbone atom looks better here than a whole side chain packed into a pocket.

Now ask what kind of contacts they are

Run Contact mapper (chemical) on 1YCR with Partner type Chain and Partner ID B. Geometry tells you two atoms are close; chemistry tells you why they are staying close.

Its summary counts the interaction types across the interface: 67 hydrophobic contacts against 12 polar and 4 ionic. This interface is held mainly by greasy side chains packing against greasy surface, plus a handful of hydrogen bonds — Trp23's indole NE1 among them.

Reading the closest rows. The shortest contacts in the table are marked polar and also vdw_clash: two atoms closer than their van der Waals radii would allow. For a polar pair at 2.7–2.9 Å that is a hydrogen bond, not a problem with the structure — read those rows as H-bonds.

Run both, always. Two residues at identical distances can be doing completely different things, and a distance table alone cannot tell you which. But neither one has yet told you which three residues matter — that is the next step.

✦ Ask AI

"I have 11 contact residues across this interface, and the shortest contact is a backbone hydrogen bond from the C-terminal residue. Why is 'closest contact' a poor measure of how much a residue contributes to binding, and what would be a better one?"

🎯 Step 4 — Find the three residues that carry it

Run Surface accessibility (SASA) on 1YCR with Chain set to B.

The surface is still measured on the whole complex, so MDM2 still covers the interface — but the table now lists only the peptide's 13 residues, most buried first. The Relative column is how much of each residue is still exposed to solvent, as a percentage of what that residue type would have fully exposed:

Three residues are almost completely buried, and there is a clear gap to everything else. Those are the hotspots: Phe19, Trp23 and Leu26, each sunk into its own subpocket on MDM2 — and you found them by measuring, not by being told. Compare that with step 3, where Leu26 came tenth on distance.

Geometry said where. Chemistry said what kind. Burial says how much. All three agreeing is what makes the triad a real finding rather than an artefact of a 4 Å cutoff.

Confirm the shape

Run Secondary structure (DSSP) on 1YCR. For chain B you will find one α-helix, residues 19–23.

Read that carefully, because it is narrower than the story usually told: the helix is five residues, not the whole 13-mer, and the ends are extended. Phe19 and Trp23 sit one turn apart on the same face of it, which is how a single short helix can present two of the three anchors to MDM2 at once. This matters for design — a stapled peptide works by holding that short helix in shape.

✦ Ask AI

"DSSP says only residues 19–23 of this peptide are helical, but it is often described as a helical peptide. Why might the ends be disordered in the crystal, and does it matter for designing a stapled version?"

You took an experimental structure and extracted, without being told, the three residues a decade of drug discovery has been trying to imitate. Nutlin-3 — the famous small-molecule MDM2 inhibitor — was built to put three groups into those same three subpockets. You have just seen the original.

⚖️ Step 5 — Score both, and find the edge of the method

This is the step worth doing slowly.

Run Binding affinity (PRODIGY) twice:

1. On 1YCR — Chains, side A: A, Chains, side B: B
2. On 3EQS — the same two chain names

PRODIGY turns the contacts across an interface into a predicted binding free energy. You now have a prediction for the natural peptide and one for the engineered peptide that is known to be 135-fold tighter.

Here is what you get:

PRODIGY ranks them the wrong way round. It calls the engineered peptide slightly weaker, when experimentally PMI binds at Kd 3.3 nM — 135-fold tighter than the natural peptide, not weaker. And in absolute terms the prediction for PMI (4.0 µM) is about 1,200× too weak, roughly three orders of magnitude.

So the tool is wrong twice over here: wrong about the ranking, and wrong about the size. The ranking is the part that would have misled you.

This is not a bug, and it is not a reason to distrust the tool. It is the most important thing in this tutorial: you have just found the edge of a method's domain, using nothing but two public structures and one number you could check.

Why it happens — and the answer is in the table

Look at the contacts column. PRODIGY is, at heart, a linear model over contact counts. 1YCR has 13 ordered peptide residues and 57 intermolecular contacts; 3EQS has 11 ordered residues and 38 contacts. The model is largely measuring how much peptide the crystallographer could resolve, not how well that peptide binds.

PMI's real advantage is a tightened internal hydrogen-bonding network and a helix that is already in the right shape before it ever reaches MDM2 — so it pays less entropy on binding. A contact count cannot see either of those things. PRODIGY was trained on protein–protein complexes, where both partners are large and folded; a 12-residue peptide is at the edge of that training set, and this is what the edge looks like from the inside.

What to actually do with this

• Use PRODIGY for what it is good at: comparing complexes of similar size and type, and getting an order-of-magnitude feel for an interface.
• Do not use it to rank short peptides against each other. Now you know why, and you can say so to a colleague with a number behind it.
• Check a method against a known answer before you trust it on an unknown one. That is the whole discipline, and it is why this tutorial spends two runs on it.

The Foundations tutorial ends on a version of this same lesson: a docking engine can find the right pose and still fail to rank it first. Sampling and scoring are different problems. Here: a predictor can be right about where an interface is and wrong about how tightly it holds.

✦ Ask AI

"PRODIGY predicted the engineered peptide PMI binds slightly weaker than the natural p53 peptide, but experimentally PMI is about 135-fold tighter. Given PRODIGY is a contact-based linear model, what properties of PMI is it structurally unable to account for, and what method would I need instead?"

Finding out that a tool is wrong, understanding why from its own output, and knowing what it is still good for, is a more valuable skill than getting a confirming answer. You now have a concrete, defensible example of it.

🔬 Step 6 — Compare the two peptides

Run Structure Alignment with 1YCR as structure A and 3EQS as structure B.

The result:

It has aligned the two MDM2 domains — and the sequence identity of 1.00 tells you something worth pausing on: the receptor is identical between the two structures. Same protein, same fold, same pocket, 0.79 Å apart.

That makes this pair a genuinely controlled comparison. Any difference in how these two peptides bind is attributable to the peptides, because nothing else differs. Pairs like that are rare and worth recognising when you find one.

What the engineers actually changed

Line the two sequences up by their shared anchors:

The triad is untouched — Phe, Trp and Leu sit at exactly the same spacing (+4 from Phe to Trp, +3 from Trp to Leu) in both. Everything the engineers changed is around it, and you can grade those changes against your own step-3 contact list and step-4 burial table:

• Position 22, Leu→Tyr. Your contact table has residue 22 touching MDM2 at 3.33 Å, and your burial table ranks it fourth, at 9.8% exposed — so this one is at the interface, just not one of the three deep anchors. Swapping a leucine for a tyrosine there adds an aromatic ring and a hydroxyl to a position that was already in contact.
• Position 24, Lys→Asn. Residue 24 is not in your contact list at all. This change is not about gripping MDM2; it is about the peptide itself — removing a charge and letting the helix hold its own shape better.

That is the general lesson of hotspot-guided design in one table: you keep what carries the binding and optimise everything else. PMI did not find a better way to grip MDM2; it found a better scaffold to present the same grip.

A note on the Sequence alignment tool. RxCubed has one, but it takes sequence files, and the RCSB connector saves coordinates — so it is not reachable from what you collected in step 2 without uploading a FASTA yourself. For two peptides this short, the table above is clearer anyway.

✦ Ask AI

"The engineered peptide PMI keeps the same Phe/Trp/Leu triad as p53 at the same spacing but changes the residues around it. Why would changing non-contacting residues improve affinity 135-fold?"

📝 Step 7 — Write it up

Ask the assistant to write a report on the experiment. Detailed instructions help it write the report you actually intend — for example:

"Write a very short report on this analysis: the research question, the paper, the two structures, the three hotspot residues and how burial identified them when distance did not, and PRODIGY ranking PMI below p53 despite PMI binding about 135-fold more tightly — and why."

Click Save as report.

Open Reports. Your new report is there, named Project Report plus the date and time it was saved. Click the pen icon next to its title to rename it, then open it in the viewer. You can edit it right there, and every change is autosaved.

✦ Ask AI

"Given everything in this project, what is the single weakest claim in my report, and what one experiment would strengthen it most?"

That's the tutorial, start to finish: a real research question, two real structures, a hotspot you measured, a method you caught out, and a short report to show for it. Nothing past this point is required. The docking worked example below is worth reading; the optional exercise after it spends your own credits.

🧩 Worked example — Read a docking run built on your restraints

Here is what the hotspots you found in step 4 are actually for. This one is not a numbered step: you read it rather than run it, and nothing on your Dashboard checklist waits on it.

Protein–protein docking (HADDOCK) is different from the small-molecule docking in the Foundations tutorial in one crucial way: it cannot do a blind search. A small-molecule docking run searches a box you draw. HADDOCK needs you to tell it which residues are at the interface — it refines and scores from there. That is a real limitation, and also why steps 3 and 4 mattered: those restraints are the output of your own interface analysis.

Read this one rather than running it

This is a worked example, not an exercise — and the reason is worth knowing, because it tells you something true about the platform.

HADDOCK docks two separate structure files. You cannot hand it 1YCR whole: that would be asking it to dock something that is already docked. So you need MDM2 and the peptide as two files — and RxCubed currently has no tool that splits a structure into its chains. Structure cleanup removes heterogens and adds hydrogens; ligand extraction pulls out ligands, not protein chains. Nothing here will hand you chain A and chain B as separate evidence.

That is a genuine gap, not something you are missing. Until it is filled, Protein–protein docking is usable when your two partners already arrive as separate files — which is the common case for real work, and not the case for a complex you just downloaded.

So read the run below. Reading a docking result and judging it is the part with the value; pressing the button adds a wait, not understanding.

The run

Run parameters — the restraints are exactly what steps 3 and 4 produce:

• Partner A: MDM2, chain A of 1YCR
• Partner B: the p53 peptide, chain B of 1YCR
• Active residues A: 25 26 50 51 54 57 58 61 62 72 73 75 93 94 96 100 104 — MDM2's side of the interface, from Contact mapper (geometric) run the other way round (target chain A, partner ID B)
• Active residues B: 19 23 26 — your triad, from step 4
• Passive residues: blank on both sides; HADDOCK auto-derives them
• Number of models: 20

It completes in about 90 seconds of real compute. When you do run HADDOCK on your own separated partners, expect that order of magnitude for a complex this small, and note that it runs on a shared, strictly serial queue — one job at a time across everyone using the platform.

Reading what comes back

You get 5 models in 3 clusters, scored:

Three things worth reading here, and none of them is the top number:

1. The top three models are all in cluster 1, within 3 points of each other. Agreement between independently sampled models is a better signal than any single score — a lone good score with nothing near it is the shape of a fluke. Here it held up: measured against the crystal, all three cluster-1 models put the peptide within 1.0–1.1 Å of where it really sits, while the cluster-2 model is 4.5 Å off and the cluster-3 model 14.7 Å.
2. A HADDOCK score is not a binding energy. It is a weighted sum of energy terms in arbitrary units. Compare scores within one run; never quote one as an affinity, and never compare it across different systems.
3. This run got the answer it was told to get, partly. The interface was supplied, so HADDOCK was never going to put the peptide somewhere else entirely. What it did determine was the orientation, the side-chain packing and the backbone fit within that interface. Be honest about which part of a result was input and which part was computation — that distinction is most of scientific integrity in this field.

It is also why the optional co-folding exercise below is a stronger test than this one: co-folding is given no interface and has to find it.

✦ Ask AI

"Someone docked a peptide back onto its receptor using restraints derived from the crystal structure of that same complex, and reported a good score. What have they actually demonstrated, what have they not, and what would a fair version of this test look like?"

🤖 Optional — design or fold something new (external)

Everything above ran locally and cost nothing. This step does not: it uses your own Neurosnap account and credits, and it is genuinely optional.

RxCubed has no peptide generator of its own. The de-novo generation tool here is small-molecule and target-blind, so this part of the work is honestly external, and saying so is better than blurring it.

Pick one of these — not all three:

• Co-fold the complex (Boltz-2, Chai-1). Give it the two sequences and let it predict the complex from scratch, then compare to 1YCR. This is the honest first-line approach for peptide work: it needs no restraints, so unlike the docking worked example it is not told where the interface is. That makes it a far stronger test than that docking run.
• Design a new binder (BindCraft, or ESMFold2 Binder Design) against the same MDM2 pocket, then ask the obvious question: did it put a bulky hydrophobic group into the Trp23 subpocket? If it did, it rediscovered the hotspot independently.
• Predict affinity properly with a method that is not a contact count — the honest follow-up to step 5.

Keep it to one job, sized down. Short binder, one or two designs, not a batch. The teaching goal is "this is what pocket-conditioned generation gives you", not a design campaign. Neurosnap exposes no per-job price through its API — only a team credit balance — so RxCubed cannot quote you a figure before you press the button. Check your own balance first, and treat job size as the lever.

Rowan, RxCubed's other external provider, now runs peptide and cyclic- peptide binder design too (BoltzGen). It is not wired into RxCubed for peptide work — the five Rowan workflows available here are small-molecule ones — and GPU design campaigns are billed per GPU-minute, so it is deliberately not part of this tutorial. It is worth knowing about if peptide design becomes your actual job.

Why Drug-likeness and ADMET are not in this tutorial

If you have done the Foundations tutorial you used them there, and you might reasonably reach for them here. Do not point them at a peptide.

Lipinski-style drug-likeness filters and the ADMET models behind them were built for small molecules — roughly, things under 500 daltons made of a few dozen heavy atoms. A 13-residue peptide violates most of those rules by construction, and the models will not tell you that. They will return confident, well-formatted numbers that mean nothing.

Peptide-appropriate property prediction (membrane permeability, protease stability, macrocycle handling) is a genuinely different toolset, and RxCubed does not have it yet. The correct move, until it does, is to leave the question open rather than answer it with the wrong instrument — which is itself worth internalising.

What this tutorial does not cover

• Designing a peptide from nothing. The docking worked example rebuilds a known interface; the optional step points at real generators but they are external.
• Anything about stability, permeability or delivery — the properties that actually decide whether a peptide can be a drug. See above.
• Molecular dynamics. The peptide's flexibility genuinely matters here and a single crystal structure hides it. MD is not currently available on this platform.
• The MDMX paralogue, which the same peptides also bind, and which is why dual MDM2/MDMX inhibitors exist.

Restarting the tutorial

Restart tutorial, under Learning workspace options on the learning-workspace banner, deletes everything you created in this walkthrough — evidence, files, runs, notes, and reports — and returns this tutorial's workspace to empty. It asks for a typed confirmation first, and it never touches your account settings, your API keys, any research project, or the other tutorial's workspace.

Where to go next

• Re-run steps 3 and 4 on 3EQS instead of 1YCR and compare. Does PMI's tighter binding show up as more contacts, or as more buried ones?
• Try the Foundations tutorial for the small-molecule side of the same story: one drug, one pocket, and a docking run you can check against the crystal.
• Bring your own complex to a research project. Everything in steps 3–6 works on any two-chain structure; the only thing special about 1YCR is that it is small and has a published answer key.