mims-harvard/tooluniverse-peptide-target-deorphanization

Peptide Target Deorphanization

Deorphanize a peptide: given a peptide sequence plus an observed phenotype (and often a hypothesized target the peptide does NOT actually bind), find its likely real protein target(s) — using ToolUniverse's keyless characterization, homology, target-family, phenotype, cross-species, and (optional, key-gated) co-folding tools.

The target can be anything, not just a GPCR. A target-class router (GPCR ligand / ion-channel toxin / protease target / cytokine or growth-factor receptor / integrin ligand / antimicrobial / unknown) classifies the peptide up front and adapts the enumeration strategy. GPCRs are the best-trodden case (and the validated control), but the pipeline's spine — homology, motif, phenotype, cross-species, co-fold — is target-class-agnostic, and family enumeration uses HGNC gene-family (general) + InterPro (general) + GPCRdb (GPCR-only cross-check).

Core reasoning: LOOK UP, DON'T GUESS

The failure mode this skill defends against is guessing a target from the peptide's name or its assumed mechanism. A peptide can be phenotypically active yet not bind the hypothesized target — because it hits a paralog, a different family member, or the same receptor in a different species whose binding interface has diverged. So:

1. Never assert a target from memory. Every candidate must come from a tool result (homology hit, family enumeration, phenotype association, or structural co-fold), with the tool name and accession recorded.
2. Anchor on PHENOTYPE × STRUCTURE/SEQUENCE plausibility, not on the peptide's reputed mechanism. The real target is the intersection of (a) what the sequence/motif/structure says it could bind and (b) what the phenotype says is biologically relevant. A name-level guess ("it's a GLP-1 analog so it's GLP1R") is exactly what produces off-target errors.
3. Reconcile across species. "Binds in species A but not B" is usually interface sequence divergence, not a different target. Always pull the ortholog set and align the candidate receptor's ligand-binding interface across the assay species before concluding the peptide "doesn't work."
4. A non-binding result against the hypothesized target is a clue, not a dead end. It promotes the paralogs and phenotype-shared receptors to the top of the candidate list.

This skill is built on validated, mostly keyless tools (BLAST, ELM/PROSITE, GPCRdb/HGNC/GtoPdb, OpenTargets, EnsemblCompara/Alliance). The single key-gated step is the optional structural confirmation by co-folding (NVIDIA NIM), used only to rank an already-narrowed shortlist.

---

Automated pipeline (scripts) — the fast path

Two runnable scripts in scripts/ execute the whole pipeline so you don't have to chain the phase calls by hand. Both load ToolUniverse via the SDK and run from the repo root.

deorphanize_peptide.py — keyless candidate generation + ranking (Phases 1–4)

No API key. For each peptide it characterizes it (PepCalc/ProtParam), flags non-canonical/cyclic residues, scans PROSITE + ELM signatures, flags protease/degradation liability, classifies the target class (GPCR / channel / protease / cytokine-receptor / integrin / …) and enumerates the candidate target family accordingly, anchors on phenotype (OpenTargets), and — for the top candidates — resolves the ortholog protein sequences and aligns the binding interface across human / assay-species / source-species (the mechanistic "binds in A, not B" step) and suggests a ClusPro-ready PDB structure. Prints a ranked candidate shortlist with evidence tiers.

python3 scripts/deorphanize_peptide.py \
  --sequence <PEPTIDE_SEQ> \             # OR  --fasta peptides.fasta  for BATCH mode
  --hypothesized-target <GENE> \         # optional; seeds family enumeration (e.g. GLP1R). OMIT for SEEDLESS mode
  --phenotype "<disease name>" \         # optional, REPEATABLE; OpenTargets anchor (use the DISEASE node, not a symptom) — pass several to union plausible phenotypes
  --assay-species mus_musculus \         # species of the NEGATIVE binding assay
  --source-species <organism> \          # optional; species where binding WAS observed -> 3-way interface alignment
  [--no-blast] [--out result.json]

Modes:

• Seeded (--hypothesized-target GLP1R) — enumerate that gene's family as the candidate panel (cleanest). Even seeded, the sequence-derived candidates (below) are always unioned in, so a wrong hypothesized seed cannot blind the search to the real target's family — exactly the deorphanization premise.
• Seedless (omit it) — derive candidate targets from PROSITE and BLAST-homolog keywords × the target-class nouns (e.g. receptor/channel/protease, chosen by the router) via UniProt; degrades to phenotype-only if that resolver is transiently down. No longer receptor-only.
• Multi-phenotype — --phenotype is repeatable; pass every plausible disease and the anchor is the union (max score per target). Best when you don't know the single right phenotype.
• Batch (--fasta) — one record per FASTA entry, sharing --phenotype/--assay-species.

Extra signals it always reports:

• Target class — the router's call (gpcr_ligand / ion_channel_toxin / protease_inhibitor_or_substrate / cytokine_or_growth_factor / integrin_ligand / guanylyl_cyclase_ligand / antimicrobial / unknown) with the evidence that triggered it and the seedless nouns it selected. This is what makes the skill general rather than GPCR-only.
• DPP4 / protease liability — a peptide can be assay-negative because it is cleaved, not because it fails to bind. Native GLP-1 (A@P2) is DPP4-LABILE; exendin-4 (G@P2) is resistant. A labile flag triggers a "re-test with a DPP4 inhibitor or protease-resistant analog" note — a key alternative explanation for "works in vitro, not in the mouse assay."
• ELM LIG motifs (ranked by rarity) + the Pfam binding domain each engages — low-confidence context for peptides without a named PROSITE family.
• Non-canonical / cyclic flag — any residue outside the 20 standard L-amino acids is surfaced, because BLAST/PROSITE/ProtParam silently assume a canonical linear peptide and will mischaracterize a non-ribosomal/cyclic peptide (common for unicellular-organism natural products). Look such peptides up by name with Norine_get_peptide and pass --cyclic to cofold_screen.py.
• Cross-species interface alignment (top ≤3 candidates) — resolves each candidate's human + assay-species (+ optional source-species) ortholog sequence (UniProt) and aligns them (EBI_msa_align), reporting per-pair % identity and substitution count. A low human-vs-assay identity flags the ortholog whose binding interface most plausibly diverged — the mechanistic answer to "binds in A, not B". If the source organism is a protist absent from UniProt, it reports insufficient and tells you to supply the partner sequence by hand.
• ClusPro-ready PDB (top ≤3 candidates) — PDBeSIFTS_get_best_structures resolves a representative solved PDB id you can feed straight to ClusPro_submit_peptide_docking.

Validated on the control (--sequence HGEGTFTSDLSKQMEEEAVRLFIEWLKNGGPSSGAPPPS --hypothesized-target GLP1R --phenotype "type 2 diabetes mellitus"): recovers the class-B panel {GCGR, GHRHR, GIPR, GLP1R, GLP2R, SCTR}, flags GLP1R as hypothesized (tested negative), and promotes GIPR to Tier 1 (family + phenotype, score 0.674) as the leading real-target hypothesis — exactly the deorphanization re-ranking, produced with zero API keys.

cofold_screen.py — structural confirmation (Phase 5, key-gated)

Co-folds the peptide against each shortlisted receptor and ranks by interface confidence (ipTM). Requires NVIDIA_API_KEY; without it, runs a DRY RUN that still resolves every receptor sequence (GPCRdb → UniProt fallback) and prints the co-fold plan, so you can verify inputs before paying for GPU time.

python3 scripts/cofold_screen.py --peptide <SEQ> --candidates GIPR GCGR GLP2R \
  [--backend boltz2|alphafold2_multimer|openfold3] [--assay-species mus_musculus] [--out cofold.json]

Use the scripts for the fast path. When you need to run, debug, or extend a single step by hand, the full per-phase manual reference — every tool call the scripts automate, with exact parameter names, gotchas, fallback chains, runtime notes, and two fully worked examples — lives in references/phases.md. Read it when a script step fails, when you want to drive a phase manually, or when you extend the pipeline to a new tool.

---

The pipeline at a glance

Six phases; the scripts automate 1–4 (and the Phase-5 dry run). Full detail + exact tool calls + gotchas are in references/phases.md — read it before driving any phase by hand.

| Phase | What it does | Key tools | |---|---|---| | 0 Verify | Confirm every tool loads; substitute fallbacks | tooluniverse.cli run … | | 1 Characterize + motif + classify | Properties, non-canonical/cyclic flag, PROSITE/ELM signature → ligand family, target-class router (GPCR/channel/protease/cytokine/integrin/…) | PepCalc/ProtParam, ScanProsite→PROSITE_get_entry, ELM_*, ESMFold | | 2 Candidate generation | 4 independent routes — homology, motif→domain, target-family enumeration (class-aware), phenotype anchor → union | BLAST/EBI_msa_align, HGNC+InterPro+GPCRdb+GtoPdb, OpenTargets | | 3 Cross-species | Resolve "binds in A not B": align the ortholog interface across human / assay / source species | EnsemblCompara, Alliance, UniProt, EBI_msa_align | | 4 Narrow + rank | Score on sequence × phenotype × pharmacology × cross-species → shortlist ≤15 | (intersection logic) | | 5 Structural confirm (optional, key-gated) | Co-fold top candidates, rank by interface ipTM; or academic-free ClusPro docking | NvidiaNIM_boltz2/…, ClusPro_submit_peptide_docking | | 6 Report | Ranked shortlist with evidence tiers + wet-lab plan | (inline report) |

Core intersection rule: the real target is where phenotype-plausible (Route 2D) meets sequence/structure-plausible (Routes 2A–2C). A name-level guess ("it's a GLP-1 analog so it's GLP1R") is exactly what produces off-target errors — never assert a target the union of routes didn't surface.

---

Output format — ranked shortlist report (return inline, no extra files)

1. Peptide characterization — length, MW, pI, GRAVY, instability, PROSITE/ELM signature, non-canonical/cyclic flag, fold confidence (if run).
2. Ranked candidate table — one row per candidate target: gene + accession · target class · evidence tier · routes that surfaced it · OpenTargets phenotype score · known-pharmacology note (GtoPdb peptide ligands?) · cross-species status (ortholog_one2one? interface % identity in the assay species?).
3. Cross-species reconciliation note — for the lead, the human-vs-assay(-vs-source) interface comparison and what it predicts for "binds A not B."
4. Recommended wet-lab validation — binding/competition + a class-appropriate functional assay against the top ≤3 candidates and their assay-species orthologs (cAMP/β-arrestin for GPCRs; electrophysiology for channels; enzymatic/inhibition for proteases; reporter for cytokine receptors), with the GtoPdb-listed family antagonist as control where one exists.

Evidence tiers — always state which evidence is keyless/validated vs key-gated (co-fold not run), and flag any candidate that is a negative against the originally hypothesized target so the reader sees the re-ranking:

• Tier 1 (strong): ≥2 independent sequence/structure routes AND present in the phenotype anchor AND (if Phase 5 run) top interface ipTM.
• Tier 2 (moderate): 1 sequence/structure route + phenotype support, OR ≥2 sequence routes without phenotype.
• Tier 3 (weak/hypothesis): single-route only (deep paralog, or a phenotype-only hit not corroborated by sequence/structure).

---

Validation & test set

Validated on the exendin-4 → GLP1R control: recovers the class-B panel {GCGR, GHRHR, GIPR, GLP1R, GLP2R, SCTR}, flags GLP1R as the (negative) hypothesized target, and promotes GIPR to Tier 1 — the deorphanization re-ranking, produced with zero API keys. The full step-by-step (control and the real "binds in the source organism, not in mouse" case) is in references/phases.md.

Reproducible test prompts + checkable assertions are in evals/evals.json (the exendin-4 control, the source-organism/mouse-negative case, and a non-GPCR seedless case). To check the skill still behaves, run a prompt through it and verify the output against that eval's assertions.