Structural Genomics & the Peptide Drug Pipeline

The commercial stakes are concrete. Eli Lilly's tirzepatide franchise (Mounjaro plus Zepbound) booked roughly $8.2 billion in 2024. What follows traces each stage, from JCSG-era expression protocols through stapling and macrocyclization, anchored to named tools, PDB entries, and clinical-stage assets.

What structural genomics actually delivers for drug discovery

Structural genomics enables peptide drug discovery by industrializing the production of target structures. Genome-scale, parallel expression, purification, and structure determination generate pre-validated peptide-binding domains, interface maps, and apo/holo conformer libraries that feed scaffold design directly. No bespoke crystallography campaign per target required. The structural census already exists, saving months of experimental work per program.

Single-target structure-aided drug discovery solves one protein to enable one program. Structural genomics, as Lundstrom framed it in 2007, runs hundreds of targets in parallel under standardized expression, purification, and crystallization protocols. This produces a population of structures whose value is combinatorial — interface families, druggable pocket inventories, and conformer ensembles usable across programs. For peptide leads, that population is the raw material.

The two anchor consortia play different roles in 2026. The Joint Center for Structural Genomics (JCSG), NIH NIGMS-funded under the Protein Structure Initiative from 2000 through PSI termination in 2015, is now an informational and protocol resource. Its Acta Crystallographica F pipeline description (Lesley et al., 2008) remains the reference implementation for high-throughput E. coli expression, IMAC purification, and sparse-matrix screening that most academic peptide-target labs still adapt. The Structural Genomics Consortium (SGC), a public-private partnership across Oxford, Toronto, Frankfurt, Karolinska, UNC, and Campinas, is the active producer.

For a working peptide chemist, the practical handoff looks like this: SGC fragment hotspot maps and probe co-crystals nominate the interface. JCSG-derived protocols produce the construct. Downstream cryo-EM or AF3 modeling resolves the bound state.

Why peptides — and why structural genomics picks them

Peptides are the modality structural genomics is best positioned to support. The interfaces revealed at genome scale are large, shallow protein–protein contacts — exactly the chemical space peptides occupy and small molecules cannot. A peptide in the roughly 500–5,000 Da range presents 10–40 backbone hydrogen-bond donors and acceptors across a contact patch that a Rule-of-Five small molecule simply cannot tile.

SGC fragment screens and JCSG-style high-throughput crystallography map PPI hotspots, hub residues, and conformer ensembles at interfaces where buried surface area routinely exceeds 1,500 Ų. Small-molecule fragment hits at those interfaces typically saturate at low micromolar affinity because the pocket is the wrong shape. A constrained peptide built against the same hotspot can recover the nanomolar binding the native interaction evolved for.

Commercial validation

Eli Lilly's tirzepatide franchise generated roughly $5.16B in Mounjaro and $3.03B in Zepbound revenue in 2024, totaling about $8.2B for a single peptide scaffold against GLP-1R/GIPR. Semaglutide sits in the same commercial tier, confirming that structure-guided peptide design against incretin receptors can produce blockbuster drugs.

Stage 1 — Genome-scale target identification and prioritization

Target prioritization for a peptide program starts with a tractable construct, not a hypothesis. The structural genomics pipeline inherits its front end almost verbatim from the JCSG and SGC playbooks: parallel cloning of every isoform and domain boundary variant of interest into a panel of expression vectors, small-scale induction screens across E. coli, insect, and HEK293 hosts, and a triage step that keeps only constructs producing >1 mg/L of monodisperse, thermally stable protein.

Why this matters for peptides specifically: the constructs that crystallize cleanly are the same constructs that behave in SPR, BLI, and HDX-MS binding assays against candidate peptides. A target that refuses to express as a soluble domain will not yield a co-crystal with a stapled lead three years later.

Modality-specific prioritization criteria

• Buried surface area at the native PPI interface (target: >1,200 Ų), because peptides need real estate to compete.
• Presence of a continuous groove or extended cleft rather than a deep enclosed pocket, which accommodates the extended backbone geometry peptides present.
• Conformational flexibility evidenced by multiple PDB conformers or high B-factors at hotspot residues, which a constrained peptide can exploit through induced-fit binding.
• Open-access structural coverage in the SGC chemical probes portal or wwPDB, which now exceeds 220,000 deposited structures with X-ray crystallography still supplying roughly 75–80% and cryo-EM 15–20%.

AlphaFold2/AF3 as a pre-experimental filter

Since 2024 the workflow runs AlphaFold2 and AlphaFold3 predictions on every prioritized target before committing wet-lab resources: predict the apo structure and any available peptide-bound complex, then score pLDDT and PAE at the putative interface. Targets where interface confidence collapses get deprioritized. The caveat is real — the AlphaFold 3 publication reports strong performance across mixed biomolecular complexes but does not break out a dedicated peptide–protein docking benchmark on a held-out set like PepBDB. Treat AF3 peptide poses as hypothesis-generating rather than confirmatory.

Stage 2 — High-resolution structure determination: cryo-EM and crystallography

Peptide drug design demands structures at 2.5 Å or better. Side-chain rotamers, ordered waters in the binding cleft, and alternate backbone conformations all resolve clearly at that threshold. In 2026 that resolution comes from two complementary methods: X-ray crystallography for soluble domains and well-behaved complexes, and single-particle cryo-EM for membrane proteins, large assemblies, and conformationally heterogeneous targets.

Crystallography and the JCSG inheritance

The JCSG-era pipeline industrialized X-ray crystallography between roughly 2000 and 2015, coupling parallel E. coli expression, IMAC purification, sparse-matrix screening, and automated synchrotron data collection at beamlines including the Advanced Photon Source at Argonne National Laboratory and SSRL at SLAC. For soluble proteases, peptide-binding domains (PDZ, SH3, WW, MHC), or E3 ligase substrate receptors, crystallography remains the default. Anisotropic diffraction below 1.8 Å typically resolves the backbone–side-chain hydrogen-bond network needed to optimize a stapled or cyclized scaffold.

Cryo-EM and the sub-2 Å question

Cryo-EM now supplies roughly 15–20% of wwPDB entries and is the default for GPCRs, ion channels, and signaling assemblies. The GLP-1 receptor–Gs complex (PDB 6X18, 3.3 Å) established the structural template that informed incretin peptide engineering, although no semaglutide- or tirzepatide-bound GLP-1R complex has been identified in PDB searches as of 2024–2025. Methods reviews from 2023 describe overall map resolutions in the 1.8–2.0 Å range for favorable specimens, but a target-matched sub-2 Å peptide-binding pocket should be treated as method-level rather than peptide-specific until verified. For peptide design, local resolution at the interface matters, not just global FSC.

Stage 3 — Fragment screening and binding hotspot mapping

Fragment-based lead discovery (FBLD) coupled to co-crystallography is how energetically privileged subsites on a PPI interface get mapped before committing to a peptide scaffold. Soak or co-crystallize the target with a library of 500–2,000 rule-of-three fragments (typically <300 Da, ≤3 H-bond donors/acceptors, cLogP ≤3). Solve each complex by X-ray diffraction. Overlay the bound poses to identify subpockets that recur across chemically distinct fragments. Recurrence is the signal.

For PPI targets relevant to peptide programs, the SGC's open-access fragment and probe portal is the first resource to check. SGC has run structure-guided fragment campaigns against bromodomains, WD-repeat scaffolds, and E3 ligase substrate receptors, with crystallographic coordinates deposited in the PDB and probe characterization data released without IP encumbrance. Fragment hits are predominantly small-molecule, but the hotspot geometries translate directly into constraints for peptide design.

In US labs, the dominant fragment screening platforms in active use are crystallographic soaking pipelines modeled on Diamond's XChem (replicated at several academic synchrotron beamlines), ¹⁹F NMR screening for soluble targets, and SPR-based primary screens triaged into co-crystallography for hits with K_D <1 mM. The hotspot map is the geometric specification document that fragment chemistry hands to peptide medicinal chemistry. For the binding-mode taxonomy these maps feed into, see protein–peptide interactions in structural genomics.

Stage 4 — Peptide scaffold design and computational optimization

Peptide scaffold design starts by translating the hotspot map into three concrete specifications: backbone topology (helix, hairpin, or constrained loop), sequence length (typically 8–20 residues for stapled helices, 6–14 for cyclized macrocycles), and the pharmacophore residue positions fixed by the fragment vectors from Stage 3. Treat the hotspot geometry as immutable and the surrounding scaffold as the degrees of freedom Rosetta is allowed to optimize.

From hotspot to backbone

For helical interfaces, pull the cognate helix from the experimental complex (or, for GPCR peptides like GLP-1R agonists, from PDB 6X18) and use it as the design template. Fix i and i+4 hotspot residues and let Rosetta's FastDesign with interface-aware scoring sample the remaining positions. The Baker Lab's Rosetta protocols for de novo peptide binder design — particularly the hallucination-then-design pipelines published 2022–2023 — are the production tools when no cognate helix exists.

For hairpins and cyclized scaffolds, the resolution of the underlying target structure matters most. A 1.9 Å crystal structure with ordered waters allows explicit water-mediated H-bond design. A 3.2 Å cryo-EM map does not.

Complex prediction and dynamics

AlphaFold-Multimer (2021) and AlphaFold3 (2024) are now standard for predicting peptide–protein complex geometries before synthesis. Treat their outputs as hypotheses, not endpoints. The AlphaFold3 paper reports accurate modeling across mixed biomolecular complex types, but it does not publish a peptide–protein-only benchmark broken out from the held-out set.

After AF3 triage, run 200–500 ns MD simulations (Amber or GROMACS, explicit solvent) on the top three predicted poses. Check helix persistence, staple-induced strain, and whether the hotspot residues maintain contact through thermal fluctuation. Poses that lose >40% of native contacts in the first 100 ns get cut before synthesis. This is also the stage where candidate staple positions (i, i+4 vs. i, i+7) get scored against backbone RMSF.

Stage 5 — Stapling, cyclization, and stability engineering

Stapling and cyclization solve the two failure modes that kill most linear peptide leads: proteolytic clearance within minutes in plasma, and membrane impermeability that locks the molecule out of intracellular targets. Both modifications constrain backbone conformation, which reduces the entropic penalty on binding and shields scissile bonds from serum and gut proteases.

Hydrocarbon stapling

For α-helical leads, all-hydrocarbon i,i+4 or i,i+7 crosslinks are installed via ring-closing metathesis between two α-methyl,α-alkenyl amino acids. The placement decision is direct structural payoff from the prior stages: overlay the predicted staple geometry onto the bound-state crystal or cryo-EM coordinates and reject any position where the olefin tether clashes with the target surface or sits on the hotspot face. Aileron Therapeutics' ALRN-6924, a stapled MDM2/MDMX dual inhibitor, was the most-cited clinical example of this chemistry; Aileron deprioritized the asset and no active interventional trials were listed by 2024–2025.

Cyclization and tethered lipidation

Head-to-tail amide cyclization, disulfide bridges, and side-chain lactam bridges all reduce the conformational ensemble of the unbound peptide and improve protease resistance at the cyclization site. Semaglutide illustrates a hybrid approach: a C18 fatty diacid is tethered via a γGlu-2xOEG linker to Lys26 of the GLP-1 backbone, driving albumin binding and extending half-life to once-weekly dosing. The lipidation site has to sit on the solvent-exposed face of the receptor-bound helix — exactly the information a GLP-1R complex structure such as PDB 6X18 provides.

Regulatory classification caveat

Stapled and cyclized peptides under ~100 residues that are chemically synthesized are generally regulated as drugs by CDER under the 2021 "Nonclinical Evaluation of Peptide-Related Products" guidance. As of early 2026 no stapled- or macrocyclic-peptide-specific guidance has superseded that framework.

From pipeline to clinic: approved and late-stage peptide drugs

Structure-guided peptide design has produced two of the highest-grossing drugs of the decade: semaglutide and tirzepatide, both GLP-1R-class peptides whose receptor-bound conformations were defined by cryo-EM structures such as PDB 6X18. Eli Lilly's 2024 Form 10-K reports combined tirzepatide franchise revenue of roughly $8.2 billion across Mounjaro (type 2 diabetes, FDA-approved 2022) and Zepbound (obesity, FDA-approved 2023). Tirzepatide is a dual GIP/GLP-1R agonist whose design exploits the shared class B GPCR architecture resolved across multiple incretin receptor complexes.

Older PPI-targeting peptides

Enfuvirtide (Fuzeon, Roche/Trimeris), approved by the FDA in 2003 for HIV-1, remains the cleanest historical example of a structure-guided peptide that disrupts a protein–protein interaction. The target was the gp41 six-helix bundle required for viral fusion. Its 36-residue sequence was derived directly from the gp41 HR2 region, a target geometry that crystallography established before the drug entered development. Enfuvirtide validated the structural logic that semaglutide and tirzepatide later monetized at scale.

Pipeline snapshot caveats

ClinicalTrials.gov does not maintain a curated peptide tally and aggregate counts swing by tens of percent depending on whether peptide vaccines, radioligands, and peptide–drug conjugates are included. A 2024 landscape review estimated 150–200 active peptide therapeutic trials globally, but that range is not anchored to a reproducible search string.

The post-AlphaFold shift in structural genomics workflows (2024–2026)

The pipeline above looks structurally similar to what JCSG ran in 2008, but the decision tree at every stage now branches through a predicted model first. Before AlphaFold2's 2021 release, a new peptide target meant a multi-month experimental campaign just to get a starting model. After AF2 — and now AF3 (Google DeepMind, May 2024) — predicted structures are the entry point for target triage, hotspot mapping, and initial peptide scaffold docking, reserving beamtime or microscope time for structures that actually need experimental ground truth.

What AF3 adds for peptide programs specifically

AF3's headline capability over AF2 is joint prediction of protein complexes with peptides, nucleic acids, ions, and small-molecule ligands in a single forward pass, including some covalent modifications. For peptide discovery, that means a first-pass model of a 10–30 residue candidate bound to its target receptor can be generated without running a separate docking protocol. The DeepMind paper reports median interface backbone accuracies in the 1–2 Å range across mixed biomolecular complexes, but a dedicated peptide–protein subset benchmark comparable to legacy PepBDB evaluations has not been broken out.

What AI still cannot replace

Three things still send the program to the beamline or the Titan Krios. First, ordered water networks in binding pockets — the ones that explain why a single methyl group shifts affinity tenfold — are not reliably predicted. Second, allosteric and apo-vs-bound conformational ensembles, including the inactive-state Class B GPCR geometries relevant to GLP-1R analogs, require experimental capture; the 6X18 GLP-1R–Gs complex at 3.3 Å is still the reference geometry for that family. Third, the sub-angstrom ligand placement needed for lead optimization remains an experimental claim.

Prediction proposes, experiment disposes.

Open infrastructure: JCSG, SGC, and the PDB as shared resources

Three open repositories carry most of the structural data a peptide program needs before it commissions a single expression construct: the worldwide Protein Data Bank, the SGC's probe and structure portal, and the legacy archive of the JCSG. None of them require a pharma-scale budget to query.

The PDB is the canonical archive of experimentally determined macromolecular structures. wwPDB reports more than 220,000 released entries by 2024, the majority still solved by X-ray crystallography and a growing 15–20% fraction by cryo-EM. Receptor reference geometries such as the 6X18 GLP-1R–Gs complex at 3.3 Å are immediately available as docking templates, AF3 input scaffolds, or validation sets.

SGC: open probes and PPI structures

The SGC publishes its chemical probes, crystal structures, and assay data under open-access terms. 2023–2024 campaigns are concentrated on PPI-rich classes including bromodomains, WD-repeat proteins, and E3 ligases. The probes are small molecules, not peptides, but the co-crystal structures are directly useful as hotspot maps when designing a macrocycle or stapled peptide against the same interface.

JCSG: legacy, not active

JCSG was decommissioned with the broader NIH Protein Structure Initiative in 2015 and now functions as an informational resource rather than an active structure-production center. The 2008 Lesley et al. description of the JCSG high-throughput pipeline — standardized E. coli expression, IMAC purification, and sparse-matrix screening — is still a reasonable starting protocol for producing peptide-binding domains or peptide–target complexes.

Key takeaways: the structural genomics-to-peptide pipeline at a glance

The pipeline collapses into five sequential stages, each anchored to a primary tool and a named output or clinical asset.

1. Genome-scale target identification and prioritization — UniProt, Open Targets, and PDB cross-referencing to triage druggable PPI interfaces and GPCRs; named output: GLP-1R prioritization that ultimately seeded the incretin franchise.
2. High-resolution structure determination — cryo-EM and X-ray crystallography deposited to the wwPDB, which surpassed 220,000 entries by 2024 with cryo-EM at roughly 15–20% of the archive; named output: PDB 6X18, the 3.3 Å GLP-1R–Gs complex.
3. Fragment screening and hotspot mapping — FBLD and crystallographic fragment screens via SGC open-access campaigns against bromodomains, WD-repeat proteins, and E3 ligases.
4. Peptide scaffold design and computational optimization — Rosetta, AlphaFold-Multimer, and AlphaFold 3 for peptide–protein complex modeling; named output: de novo and analog designs feeding GLP-1/GIP dual agonists.
5. Stapling/cyclization and stability engineering — hydrocarbon stapling, lactam bridges, and head-to-tail cyclization; named drugs: enfuvirtide, semaglutide, and tirzepatide, the latter generating roughly $8.2 billion in combined Mounjaro and Zepbound revenue in 2024.

Structural genomics infrastructure built for genome-scale enzymology now underwrites the most commercially consequential modality in metabolic disease. The gap between PDB deposition and peptide lead is no longer translational — it is a defined workflow.

Next steps: scoping your own peptide program

Start by querying the wwPDB and SGC probe portal for structures of your target class. If you find a reference complex at 2.5 Å or better with a cognate peptide or protein ligand, you have a template for scaffold design. If not, run AlphaFold3 predictions on your target and any known binders to generate a first-pass model, then use that model to identify hotspot residues and interface geometry. Commission a fragment screen or pull existing SGC fragment data for the same target family. Once you have a hotspot map, you are ready to design and synthesize your first stapled or cyclized candidates.

For the upstream biology of why these interfaces dominate eukaryotic regulation, see how protein–peptide interactions shape cellular function. For the binding-mode taxonomy that informs scaffold choice, see protein–peptide interactions in structural genomics. For the downstream medicinal chemistry, see from protein structures to peptide therapeutics.