Structural Genomics in Peptide Drug Discovery

Key takeaways

• Experimental structures remain essential even after AlphaFold3 because predictions degrade sharply on flexible peptides, induced-fit receptors, and membrane-embedded targets (40–50% of test cases fail).
• Structural genomics outputs feed directly into each stage of peptide drug discovery: target validation, binding-site definition, scaffold selection, modification placement, and lead optimization.
• SG-built synchrotron infrastructure (Diamond I04-1/I24, APS-U, SSRL) now processes 1,000 to 20,000 datasets per day, compressing lead optimization cycles from months to weeks.
• Open PDB deposits under CC0 licensing eliminate IP friction and capital barriers for small biotechs competing against large pharma.
• Crystallization remains the rate limiter in 2026; data collection is no longer the bottleneck.

Why experimental structures still matter after AlphaFold3

AlphaFold3 reports interface backbone RMSD ≤ 2 Å on roughly 50–60% of protein–peptide test cases. The other 40–50% (flexible peptides, induced-fit receptors, membrane-embedded targets) still require experimental structures because predictions degrade sharply when backbone flexibility or large conformational rearrangements dominate.

Synchrotron co-crystallography at Diamond's I04-1 and I24 clears 1,000 to 3,000 crystals per 24-hour shift in soaking mode. JCSG alone deposited roughly 1,200 to 1,300 unique entries before PSI wrap-up. That public corpus makes a 2026 peptide pipeline tractable — without it, each program would need to solve its own target structures from scratch.

Some teams worry that AlphaFold3 has made experimental structures redundant. The accuracy gap on flexible and membrane-embedded targets proves otherwise: one in three predicted complexes is wrong enough to mislead a stapling or N-methylation decision.

Why structural genomics matters for peptide drugs

Structural genomics produces open, target-agnostic experimental structures (PDB deposits, ligand-soaked co-crystals, apo-form scaffolds) that peptide chemists need to design 10–50 residue ligands against flat, extended, or shallow interfaces. Small molecules cannot productively engage these interfaces.

Peptides occupy a modality gap: too large for classical Lipinski-space pockets, too small and metabolically fragile to behave like antibodies.

Peptide leads typically engage 600–1,500 Ų of interface. Every productive design cycle requires a high-resolution view of side-chain rotamers, backbone hydrogen bonds, and ordered waters across the binding groove. Predicted models alone rarely resolve those details on induced-fit or membrane-embedded targets because AlphaFold3 confidence metrics (pLDDT and PAE) were trained on experimental structures and cannot extrapolate reliably beyond the training set's accuracy ceiling.

The 2026 SG ecosystem

The PSI (Protein Structure Initiative) centers (JCSG, NESG, MCSG, and others) formally ended in 2015. Their roughly 6,000 to 7,000 deposited structures remain queryable through RCSB's Structural Genomics center filter and continue to seed homology templates for AlphaFold3. Fold-family saturation reduces prediction error on new targets in the same family.

The Structural Genomics Consortium (SGC) operates as the active successor, running open-access crystallography and cryo-EM on under-characterized human targets. For deeper background on how these deposits feed peptide campaigns, see our hub on structural genomics and peptide drug discovery.

The SG infrastructure that peptide programs actually use

Peptide programs draw on three concrete SG outputs: open PDB deposits under permissive licenses, robotic crystallization plus automated synchrotron data collection, and standardized cloning and expression protocols that made both possible at scale. None of these are abstractions in 2026 — they are the daily plumbing of any structure-enabled peptide campaign.

Open PDB deposits as permissive scaffolds

JCSG alone deposited on the order of 1,200 to 1,300 unique structures between 2000 and 2015. The four major PSI centers (JCSG, NESG, MCSG, plus SGC operating in parallel) together account for roughly 6,000 to 7,000 entries. Every coordinate set is CC0 or equivalent.

This matters for peptide chemists building proprietary macrocycle libraries. A homology template lifted from a JCSG apo structure carries no downstream IP friction — the chemist can incorporate it into a proprietary docking pipeline without a license agreement. The SGC extends this with an explicit open-access mandate covering reagents, constructs, and pre-publication structures on under-characterized human targets including kinases, bromodomains, and solute carriers.

Robotic crystallization and automated beamlines

The methodology Grabowski and colleagues codified for PSI-era pipelines (nanoliter sitting-drop screens, automated imaging, standardized cryo-loop handling) is now baseline equipment at peptide-focused biotechs running co-crystallography in-house. The payoff shows up at the synchrotron:

• Diamond's I04-1 and I24 process 1,000 to 3,000 crystals per 24-hour shift in fragment or ligand-soaking mode.
• APS-U beamlines advertise 10,000 to 20,000 data collections per week post-upgrade.
• SSRL's automated MX lines support several hundred to over 1,000 ligand-bound datasets per day when campaigns are tuned.

For a peptide team soaking a cyclic library against a target crystal form, that throughput compresses a six-month SAR cycle to six weeks.

Standardized cloning and expression

The construct-design heuristics SG centers published (truncation boundaries, fusion-tag matrices, parallel expression screens) remain the default starting point for producing receptor extracellular domains, kinase catalytic cores, and protease constructs that peptide leads engage.

Stage 1: Target validation at scale

Structural genomics earns its keep before a single peptide is synthesized by answering two questions chemistry cannot: Is this target expressible in a form a peptide can engage? Does its surface present a ligandable groove rather than a featureless interface?

A peptide program engaging a class B GPCR like GLP-1R or GIPR requires the full extracellular domain (ECD) plus a stable transmembrane bundle. The binding mode of agonist peptides spans both regions, and a truncated construct will miss critical contacts. The SG-era construct libraries remain the starting deck teams reach for when building a tirzepatide-style co-agonist construct in 2026.

Family-wide coverage as validation prior

GPCR-targeted peptide programs move faster today than they did pre-PSI because the family has been mapped, not just sampled. A target-validation team can pull dozens of related structures from the PDB to assess ligandability before committing to expression work.

Stage 2: Defining the peptide binding site

Defining a peptide binding site means mapping the full interaction epitope, not just identifying a small druggable cleft. Peptides typically engage 600–1,200 Ų of extended, often flat surface across multiple secondary-structure elements.

For class B GPCRs, the design-stage value is concrete. GLP-1R and GIPR cryo-EM structures captured between 2020 and 2025 resolved the active-state ECD plus transmembrane bundle bound to peptide agonists at 2.5–3.5 Å, exposing the deep orthosteric pocket geometry that tirzepatide-style co-agonists exploit at residues lining TM1, TM5, TM6, and TM7.

Hot-spot residues and conformational flexibility

The practical workflow in 2026 pulls every deposited apo and holo structure of the target from the PDB, aligns them, and quantifies side-chain variance at putative hot-spot residues. For BCL-2 family targets driving stapled-peptide programs, the spread of MCL-1 and BCL-xL structures reveals which BH3-groove residues (the conserved hydrophobic h1–h4 sub-pockets) are rigid anchors versus which shift on binding.

Cryo-EM filling the membrane-target gap

SG-era X-ray crystallography stalled on full-length membrane receptors, transporters, and large complexes. Cryo-EM after 2020 closed the gap, and the structures land in the same PDB that SG infrastructure normalized as open public deposit.

Stage 3: Scaffold and modification selection

Scaffold choice falls out of three structural readings: the bound-peptide secondary structure in the holo PDB ensemble, the depth and continuity of the binding groove, and the side-chain vectors of anchor residues. Helical grooves with 3–4 turns of buried hydrophobic contact push you toward stapled alpha-helices. Shallow, discontinuous interfaces push you toward macrocyclic peptides or linear N-methylated scaffolds.

Stapled helices versus macrocycles

For BCL-2 family targets, the BH3 groove is a continuous amphipathic channel with four hydrophobic sub-pockets (h1–h4) accepting helical i, i+3, i+4, i+7 side chains. That geometry is why MCL-1 and BCL-xL programs converge on i,i+7 hydrocarbon staples. KRAS surfaces tell the opposite story — the shallow switch-II pocket explains why clinical-stage KRAS peptide programs trend toward macrocyclic and bicyclic scaffolds.

N-methylation placement

N-methylation sites are chosen at backbone amides that the structure shows pointing into solvent, not into the target. If the carbonyl or NH participates in a hydrogen bond visible in the holo structure, leave it alone. If it projects away from the interface, it is a candidate for methylation to block protease recognition without paying an affinity tax.

Lipidation and half-life extension

Semaglutide's C18 diacid lipidation at Lys26 is the canonical example of structure-guided half-life engineering. Lys26 sits on the solvent-facing face of the GLP-1 helix in GLP-1R complexes, well away from the ECD and TMD contact surfaces. The same logic now drives lipidation site selection across GLP-1R, GIPR, and glucagon receptor co-agonists.

Stage 4: Lead optimization via iterative co-crystallography

By 2026, weekly analog-to-co-crystal-to-SAR cycles are table stakes for peptide biotechs. Ship a tray of 96 analogs on Monday, get processed datasets and refined difference maps back by Friday. The JCSG-developed automation stack is now the default on these beamlines rather than a specialist offering.

Where the cycle bottlenecks in 2026

Crystallization itself, not data collection, is the rate limiter for most peptide programs. Macrocycles and N-methylated analogs that diverge significantly from the parent often require re-screening of crystallization conditions, adding two to four weeks. Programs that maintain a soakable apo or parent-complex crystal form compress the loop to five to seven days per analog batch.

Cost framing

A de novo membrane-target structure solved entirely in-house at a mid-size biotech runs roughly $1M to $3M all-in, with median budgets near $1.5M per novel high-resolution structure. Iterative co-crystallography on an established crystal form, run through an SG-derived beamline pipeline, costs a small fraction of that per analog.

The AlphaFold2/3 hybrid workflow (2024–2026)

AlphaFold2 (2021) and AlphaFold3 (2024) accelerated hypothesis generation for peptide–target interfaces but did not retire experimental structural genomics. Instead, they shifted SG output from primary evidence to validation set and training data. A predicted complex is treated as a starting pose that must be checked against an SG-deposited experimental structure of the same fold family before any medicinal chemistry decision is made.

The 2024 AlphaFold3 release reports interface backbone RMSD ≤ 2 Å on roughly 50% to 60% of protein–peptide test complexes. Independent 2025 benchmarks recover DockQ ≥ 0.5 in 55% to 70% of blind cases. For a peptide program, roughly one in three predicted complexes is wrong enough to mislead a stapling or N-methylation decision.

The loop most programs run in 2026

1. AF3 prediction of the peptide–target complex, with multiple seeds and MSA-depth sweeps to gauge confidence.
2. Cross-check against SG-deposited PDB entries for the target's fold family. JCSG, NESG, MCSG, and SGC deposits serve as the de facto validation set.
3. Co-crystal or cryo-EM confirmation of the lead peptide series, typically routed through Diamond I04-1/I24, APS-U MX beamlines, or SSRL 12-2.

Named peptide programs tracing back to SG structures

Tirzepatide (GIP/GLP-1R co-agonist, Eli Lilly, approved 2022)

The class B GPCR fold underpinning both GLP-1R and GIPR was characterized across multiple SGC and academic deposits during the PSI era. SG's contribution is upstream: it normalized class B ECD crystallography and provided the calibration set against which AlphaFold2/3 models of GIPR were validated.

MCL-1 stapled peptides (AMG 176/tapotoclax-class, Phase 1/2)

BCL-2 family structures, including several MCL-1 BH3-binding-groove complexes deposited during PSI-2, anchor essentially every published MCL-1 stapled peptide design paper. This is the strongest case for SG involvement in a clinical asset.

PCSK9 peptide antagonists (Merck MK-0616, Phase 3)

MK-0616's structural design traces to PCSK9–LDLR interface crystallography. SG-era PCSK9 deposits provide the catalytic/prodomain coordinates used in early macrocycle screening models.

KRAS-binding macrocycles (Revolution Medicines RMC-6236, Phase 2/3)

RAS GTPase and SOS1/RAF complex structures from SGC and MCSG deposits form the public scaffold used in most KRAS macrocycle design pipelines.

Open data as competitive accelerant for small biotechs

The SG open-access model lets a 10-person peptide biotech compete on chemistry rather than on capital infrastructure. A 2025 biopharma benchmarking survey put the median internal budget for a novel high-resolution membrane protein structure at roughly $1.5M at mid-scale firms, with all-in ranges of $1M to $3M. When the apo structure and several ligand complexes already exist as SG deposits, that entire budget line collapses to the cost of a computational chemist running AlphaFold3 refinement against the experimental scaffold.

Limitations and what SG still can't deliver

Structural genomics outputs degrade sharply for three target classes that dominate current peptide pipelines: intrinsically disordered proteins (IDPs), transient signaling complexes, and conformationally heterogeneous membrane proteins. The PDB does not contain ground-truth coordinates for ensembles, and neither AlphaFold2 nor AlphaFold3 produces calibrated conformational populations for disordered regions as of 2026.

For IDP targets such as c-Myc, p53 transactivation domain, and tau, SG deposits give you fragments of bound states at best. A peptide chemist optimizing against a fuzzy interface still requires NMR chemical-shift perturbations or HDX-MS as the primary structural readout.

FAQ

Next steps

Pull every JCSG, MCSG, NESG, and SGC deposit for your fold family from RCSB this week. Run an AlphaFold3 sweep anchored to those templates. Reserve Diamond I04-1 or SSRL 12-2 remote time for the first soaking campaign. Then walk that pipeline against the stage map in our structural genomics and peptide drug discovery hub to find the cycle where your program is losing weeks. The bottleneck is usually crystallization, not data collection.