Protein–Peptide Interactions in Structural Genomics

These two modes exist because peptides lack the autonomous folding energy of proteins. They either adopt a conformation templated by the receptor (SLiM) or dock into a pre-formed surface (groove). Each mode maps to a distinct therapeutic strategy: SLiM-class interfaces are the natural target for stapled or cyclised mimetics of disordered motifs because the binding penalty is conformational entropy; shallow grooves underpin Bcl-2-style PPI inhibitors and the broader peptide modulator class because the receptor is already organised. Recent deep-learning work predicting peptide binders directly from sequence and structure leans heavily on PDB-derived complex sets seeded by structural genomics output — corpus quality directly affects prediction accuracy.

Key takeaways

• The JCSG corpus reveals two distinct protein–peptide binding modes: SLiM-based induced-fit recognition and shallow hydrophobic groove binding, each requiring different peptide chemistry strategies.
• Structural genomics at consortium scale captures population-level patterns that single-lab crystallography cannot, providing the training data that computational predictors now depend on.
• Binding mode determines design strategy: SLiM targets require conformational pre-payment through cyclisation or stapling; groove binders reward backbone rigidification and hotspot-spanning side chains.
• The corpus is structurally biased towards readily crystallisable complexes, affinity values lack normalisation, and deposition lags the literature by 12–24 months.
• Computational predictions in 2026 benefit from triangulation between homology templates, AI methods, and biological priors rather than reliance on any single approach.

What protein–peptide interactions are (and our 5–40 residue cutoff)

A protein–peptide interaction, in our JCSG working definition, is the non-covalent association between a folded globular domain and a linear or weakly structured peptide ligand of 5–40 residues, captured at atomic resolution and deposited as a single complex entry. The lower bound excludes dipeptide and tripeptide cofactor fragments that behave as small molecules because they lack the conformational flexibility that defines peptide recognition. The upper bound excludes mini-domains that fold autonomously and are better treated as PPIs because they do not require receptor templating.

The cutoff matters because the literature disagrees. Some 2024 proteome-wide SLiM studies treat anything under 30 residues as a peptide. The Pelay-Gimeno PPI-modulator review implicitly extends the class to ~50 residues when discussing stapled helices and miniproteins. We retain 40 as the JCSG ceiling because beyond that, secondary-structure content typically becomes self-supporting and the induced-fit assumptions underlying our crystallisation and docking protocols break down.

Why this differs from a PPI

A PPI involves two pre-folded surfaces. A protein–peptide complex involves one folded surface and one ligand whose conformation is largely templated on binding. That asymmetry is the structural basis for the two recognition modes discussed above, and it dictates which therapeutic route applies — stapled SLiM mimetic or groove-binding helix — because each mode imposes different entropic costs. For the upstream biology, see how protein–peptide interactions shape cellular function. For the downstream translation, see from protein structures to peptide therapeutics.

Why structural genomics is the right lens

Consortium-scale determination is the right approach because the gap between known sequences and experimental complex structures is too wide for any single lab to close. Taherzadeh and colleagues (2018) showed that only a small fraction of predicted protein–peptide interactions have an experimental complex structure. The ratio has worsened as proteomic and interactomic catalogues have grown faster than the PDB. This divergence means computational methods lack sufficient training data unless they draw from consortium pipelines.

In our JCSG pipeline, parallelised crystallography and cryo-EM were designed around exactly that population problem: standardised constructs, automated crystallisation trials, common refinement protocols, and uniform deposition metadata. The output is a corpus in which complexes can be compared on like-for-like terms rather than a scatter of one-offs with idiosyncratic data quality. That uniformity lets us treat SLiM induced-fit and shallow hydrophobic groove binding as empirically distinguishable classes rather than anecdotes.

Scale changes the question

With a handful of structures, you ask whether a particular peptide binds in a particular groove. With several hundred deposited complexes processed through a common pipeline, you can ask which side-chain registers recur across SH3-, WW- and PDZ-class grooves, where hotspot residues cluster on the folded partner, and how often induced-fit ordering exceeds 8–10 residues. Those statistics convert a structural archive into a design substrate for peptide chemistry.

Inside the JCSG pipeline: expression to PDB deposition

The JCSG pipeline routes every protein–peptide target through a fixed sequence: recombinant expression, modality triage, diffraction or imaging, and curated deposition with annotated contacts. The point is uniformity of metadata, not novelty per structure.

Expression and complex assembly

Expression host selection follows the folded partner, not the peptide, because the receptor's folding requirements determine whether the complex will assemble correctly. E. coli (BL21(DE3) and Rosetta derivatives) handles the bulk of single-domain bait proteins: SH3, WW, PDZ, BIR, BRCT. Insect cell expression (Sf9/High Five with baculovirus) is reserved for multi-domain scaffolds requiring eukaryotic chaperones. HEK293 suspension is held back for glycosylated ectodomains and signalling complexes where the peptide is a phosphorylated motif. The peptide is almost always synthesised separately and added at the complex-assembly step because synthetic peptides are cheaper and faster than recombinant expression.

Co-crystallisation is the default for peptides with Kd below roughly 50 μM, typically at 2–5× molar excess against the purified bait. Soaking is used where the apo crystal form is robust and the binding groove is solvent-accessible — common for shallow hydrophobic grooves but rarely viable for SLiM-binding domains that require induced-fit ordering of loops over the bound motif.

Modality triage: X-ray, cryo-EM, NMR

Diffraction data are collected primarily at SSRL beamlines 12-2 and 9-2, with APS 23-ID used for difficult crystal forms and microcrystals. Single-particle cryo-EM enters the pipeline for transient assemblies and membrane-embedded complexes where crystallisation is unproductive; sub-3 Å maps are now routine for baits above ~80 kDa.

Solution NMR is the chosen route for short, flexible peptides where the bound ensemble matters as much as a single pose. When the peptide samples multiple registers in the groove, a crystal structure picks the lattice-favoured conformer and hides the equilibrium. ¹⁵N-HSQC titration plus NOE-restrained docking recovers the population.

Deposition and annotation

Every JCSG entry is deposited to the PDB with contact-residue tables (4.0 Å cutoff, both directions), measured or estimated affinity where available, and a binding-mode tag. Those annotations make the corpus queryable for the downstream translation work covered in from protein structures to peptide therapeutics.

Pattern 1: SLiM-based induced-fit binding

Short linear motifs (SLiMs) bind their partner domains by folding from a disordered state into a defined backbone geometry inside the receptor pocket. They pay an entropic penalty that caps achievable affinity in the low-micromolar to high-nanomolar range for unmodified peptides. In our JCSG pipeline this is the dominant pattern for signalling and trafficking baits. The apo domain shows ordered secondary structure but disordered specificity loops; only the holo complex resolves the loop register over the bound motif.

The structural signature is consistent across the corpus. The peptide adopts an extended or polyproline-II backbone of 4–8 residues, with two or three anchor side chains buried in well-defined sub-pockets and intervening residues solvent-exposed and tolerant of substitution. The receptor contributes induced-fit loop closure, typically reordering a 6–12 residue lid that is invisible in the apo map. That lid closure converts a shallow recognition site into a buried interface and explains why apo soaking almost never works for this class.

The motif vocabulary is small. Vanhee and colleagues showed in 2009 that the functional SLiM space is heavily reused, with a few hundred motif classes covering most regulatory interactions. Our deposition statistics mirror this: SH3-PxxP, WW-PPxY, PDZ C-terminal hooks, and 14-3-3 phosphopeptide complexes dominate the SLiM-class entries. Representative JCSG and community exemplars are 1SSH (SH3 with proline-rich peptide), 1JMQ (WW domain with PPxY), and 1BE9 (PDZ C-terminal recognition).

The affinity implication is direct. Because induced fit consumes 3–5 kcal mol⁻¹ of binding energy on receptor reorganisation, linear SLiM peptides rarely break 100 nM without conformational pre-payment via cyclisation, stapling, or backbone N-methylation. That biophysical ceiling is the entry point for the design strategy covered in from protein structures to peptide therapeutics.

Pattern 2: shallow hydrophobic groove hotspots

The second dominant pattern in our JCSG corpus is the shallow hydrophobic groove: a pre-formed, solvent-exposed channel 15–25 Å long and 4–6 Å deep. A peptide docks into this channel as an extended strand or short helix without significant receptor reorganisation. Unlike the SLiM induced-fit class, the apo and holo maps are near-superimposable, which makes these targets amenable to fragment soaking and crystallographic screening.

Groove geometry is what disqualifies most small molecules. The interface is broad but shallow, so a 400 Da ligand can occupy at most one or two of the four-to-six hotspot residues lining the channel. Peptides of 8–15 residues match the groove length, present side chains at the correct register, and can be rigidified to recover the entropic cost without a deep anchor pocket. This is the structural rationale behind Bcl-2 family BH3-groove binders and MDM2–p53 helix mimetics.

Surface-area and hotspot thresholds

Buried surface area (BSA) in the JCSG shallow-groove entries clusters between 700 and 1,100 Ų per partner, roughly double the SLiM-class median. Three to five hotspot residues contribute ≥ 1.5 kcal mol⁻¹ each on alanine scanning. Representative exemplars worth pulling from the PDB are 2XA0 (MCL-1 with BH3 peptide), 1YCR (MDM2 with p53 transactivation helix), and 1LB6 (14-3-3 in groove-binding mode without phosphorylation dependence).

Why this maps to a distinct design strategy

Because the receptor is pre-organised, the design problem collapses to peptide pre-payment of conformational entropy: helical stapling, β-hairpin grafting, or backbone constraint. That downstream workflow is treated in from protein structures to peptide therapeutics.

From binding pattern to peptide drug design strategy

The structural class of the interface dictates the chemistry. SLiM-class targets demand peptides that pre-pay conformational entropy and lock the bound rotamer set. Shallow hydrophobic grooves reward backbone rigidification and hotspot-mimetic side chains spanning the full groove length. In our JCSG triage, this maps to two distinct chemistry stacks with different developability profiles.

SLiM-class: lock the induced-fit conformation

For SLiM substrates against SH3, WW, PDZ, and 14-3-3 phosphopeptide pockets, the design objective is to constrain a 4–8 residue motif into its bound geometry before it reaches the receptor. Three tactics dominate: head-to-tail or side-chain cyclisation to remove backbone flexibility, N-methylation of solvent-exposed amides to tune permeability and protease resistance, and selective d-residue substitution at non-contact positions. Hydrocarbon stapling is generally over-engineered here because the bound conformation is rarely a clean α-helix.

Shallow-groove targets: span and rigidify

Groove binders against Bcl-2 family proteins, MDM2, and 14-3-3 in groove mode tolerate, and usually require, longer peptides of 12–25 residues with stabilised secondary structure. Hydrocarbon stapling, lactam bridges, and β-hairpin grafting all earn their keep because the receptor is pre-organised and the peptide simply needs to present three to five hotspot side chains at the correct register. The clinically validated GLP-1 receptor agonists semaglutide and tirzepatide are instructive as agonist-mode groove binders.

Developability trade-offs

Cyclised SLiM mimetics generally win on oral or sublingual potential because their molecular weight stays below ~1.2 kDa. They fail more often at selectivity across paralogous domains because the short motif is shared across many receptors. Stapled groove binders carry better selectivity and affinity but pay in molecular weight, synthetic cost, and almost always parenteral-only delivery.

Computational prediction in 2026: what the JCSG corpus trains

The JCSG corpus contributes to peptide-complex prediction in three concrete ways: as template input for Rosetta FlexPepDock and homology pipelines, as raw material for curated subsets like PepBDB, and as held-out validation for AlphaFold-Multimer and its 2024–2026 successors. None of these uses is captured by a wwPDB-level "protein–peptide complex" count.

What classical methods still earn their keep on

Template-based docking remains competitive when a close structural analogue exists for either the receptor groove or the bound peptide register, because homology can transfer both geometry and energetics. Johansson-Åkhe and colleagues showed in 2019 that templates were available for 81.3% of peptide–protein complexes in their benchmark. In our JCSG pipeline experience, that ceiling holds best for shallow-groove receptors with multiple paralogous structures (Bcl-2 family, 14-3-3, MHC class I) and collapses for SLiM induced-fit cases where the unbound receptor surface is genuinely different.

The AI gap, and why the numbers won't sit still

Deep-learning peptide–complex predictors have moved past template ceilings on the harder induced-fit cases. The field lacks a canonical leaderboard. The Bradley lab's 2023 sequence-and-structure approach to peptide–protein binder prediction trains on PDB-derived complex sets without quoting a single docking-style RMSD benchmark suitable for cross-method comparison. The practical consequence: any "X% improvement over AlphaFold-Multimer" headline you read in 2026 is tied to a private test split and will be stale within twelve months.

For groups wiring JCSG-derived structures into a prediction stack, the defensible workflow is to retain a homology baseline (FlexPepDock with PepBDB templates), run an AI predictor in parallel, and reconcile disagreements against the biological priors set out in how protein–peptide interactions shape cellular function before committing to a chemistry programme.

Limits, caveats, and what the corpus still misses

The JCSG-derived corpus of protein–peptide complexes is structurally biased, affinity-heterogeneous, and lags both the literature and the regulatory pipeline by 12–24 months. Five constraints matter when wiring it into a peptide drug programme in 2026.

First, crystallisation selects against the very biology peptide chemists care about most. Complexes that survive vapour diffusion tend to have ordered receptors and pre-organised peptide conformers. Intrinsically disordered partners and transient SLiM engagements are systematically under-sampled.

Second, affinity values attached to PDB entries come from ITC, SPR, fluorescence polarisation, AlphaScreen, and competition assays with no enforced normalisation. Treating Kd numbers as cross-comparable without re-reading the methods section will mislead any QSAR or scoring-function recalibration built on top of the corpus.

Third, PDB deposition lag is real and uneven. The wwPDB does not publish a dedicated protein–peptide complex count, and curated subsets such as PepBDB and PepX are themselves snapshots.

Fourth, AI benchmarks shift faster than the corpus does. A model that was state-of-the-art in March 2026 may be retrained against a test split that now leaks. Lock your benchmark dates.

Fifth, MHRA and EMA approval documents do not classify peptides by binding mode. The structure-to-drug mapping detailed in from protein structures to peptide therapeutics must be reconstructed case by case from primary structural papers, not regulator summaries.

Where to go next at JCSG

The two sibling explainers extend the pipeline view in opposite directions along the structure-to-function-to-drug axis.

For the upstream biology, how protein–peptide interactions shape cellular function sets out why SLiM-mediated signalling and shallow-groove recognition dominate eukaryotic regulation. Treat it as the prerequisite if you are coming from a computational background and need the cellular rationale before reading PDB metadata.

For the downstream translation, from protein structures to peptide therapeutics walks through the structure-to-lead workflow that this overview only gestures at: stapling, backbone constraint, and shallow-groove mimetics on one track, SLiM-derived induced-fit ligands on the other.

A note on conceptual scaffolding

For readers benchmarking 2026 predictors against PDB-derived peptide sets, pair the cellular-function piece with this article before consulting the therapeutics page. Reading them in that order keeps the binding-mode taxonomy stable across all three.