{
  "version": 3,
  "created_at": "2026-07-07T20:07:40.577Z",
  "task_summary": "Phase 2: mechanistic validation, novelty/genetics grounding, patient stratification, and protein-therapeutic design for prioritized EoE targets",
  "agents": [],
  "phases": [
    {
      "name": "Plan",
      "delegations": [
        {
          "steps": [
            {
              "title": "Epithelial differentiation trajectory analysis",
              "description": "On the annotated single-cell object (gse218607_annotated / slim h5ad), subset the epithelial compartment (basal + suprabasal, ~143k cells) and run pseudotime + RNA-velocity (scanpy/scVelo or diffusion pseudotime if spliced counts unavailable) to test whether EoE arrests basal→differentiated maturation. Place CAPN14, CDH26, SPINK7, DSG1, FLG and lead targets along the trajectory; compare EoE vs Healthy trajectory density. Deliverables: epithelial_trajectory.png (pseudotime UMAP + gene-along-trajectory curves, EoE vs Healthy), trajectory_gene_dynamics.csv. Quality bar: publication-grade, EoE/Healthy split explicit, barrier genes shown declining along maturation axis."
            },
            {
              "title": "Interferon signature deconvolution",
              "description": "Resolve the type-I vs type-II IFN program flagged in Phase 1. Score type-I (ISG15, MX1, OAS1/2, IFIT1) vs type-II/IFN-γ (GBP1/5, IRF1, STAT1, CXCL9/10/11) modules per cell type in the single-cell data, identify source vs responder cells, and test whether IFN tracks the severe endotype (link to Phase-1 heterogeneity preview). Cross-check against bulk meta-signature. Deliverables: ifn_deconvolution.png (per-cell-type type-I/II scores, source-vs-responder), ifn_module_scores.csv. Quality bar: clear statement of whether IFN-γ is upstream of epithelial HLA induction."
            },
            {
              "title": "Differential cell-cell communication (EoE vs GERD vs healthy)",
              "description": "Re-load the held-out GERD samples (91k cells) with the EoE+Healthy set; run liana rank_aggregate separately per condition and compute differential ligand-receptor interactions so EoE-specific signaling (S100A8/A9→mast/myeloid, IL33→IL1RL1, mast VIM-CD44→epithelium) is separated from generic esophageal inflammation. Deliverables: differential_cellcomm.png (EoE-specific vs shared LR axes), cellcomm_eoe_vs_gerd.csv. Quality bar: each highlighted axis annotated as EoE-specific or inflammation-generic."
            },
            {
              "title": "Mast-cell state characterization",
              "description": "Subcluster the mast compartment (2,107 cells) to resolve disease-associated states; test whether SIGLEC6 marks a depletable EoE-enriched subset and characterize protease (CPA3, TPSAB1, CTSG) and activation programs. Quantify state proportions EoE vs Healthy. Deliverables: mastcell_states.png (subcluster UMAP + marker dotplot + composition), mastcell_state_markers.csv. Quality bar: explicit call on whether a discrete SIGLEC6+ disease state exists."
            },
            {
              "title": "Patient endotype stratification (cross-cohort)",
              "description": "Extend the Phase-1 GSE250595 endotype preview to all powered Tier-1 cohorts: cluster patients on the 567-gene signature per cohort, characterize by Th2/mast/IFN/barrier modules, and test endotype reproducibility across cohorts (consensus clustering / module-score concordance). Assign each patient a mild/intermediate/severe/fibrostenotic label. Deliverables: endotype_crosscohort.png (per-cohort endotype heatmaps + module profiles), patient_endotypes_all.csv. Quality bar: reproducibility explicitly quantified, not assumed."
            },
            {
              "title": "PPI-response target triage",
              "description": "Re-fetch original GEO sample metadata for GSE303169 (PPI-responder study) and any other cohorts with treatment/response coding to recover responder/non-responder labels. Classify each Tier-A target as PPI-reversible (normalizes in responders) vs PPI-refractory (persists) — refractory-persistent targets are highest-value novel leads. Deliverables: ppi_response_triage.png (target FC in responders vs non-responders), ppi_target_triage.csv. Quality bar: state which targets remain elevated in the treatment-refractory population."
            },
            {
              "title": "GWAS risk-locus overlap",
              "description": "Cross-reference prioritized targets against EoE/EGID GWAS risk loci using the human-genetics connector (GWAS Catalog, eQTL Catalogue) — known EoE hits include CAPN14, TSLP, STAT6, LRRC32, ANKRD27. Report which shortlist targets carry genetic support and whether risk alleles are eQTLs for them; note endotype-specific genetic associations where available. Deliverables: gwas_overlap.png (targets × genetic-evidence), gwas_target_overlap.csv. Quality bar: every reported association traced to a retrieved GWAS/eQTL record, not memory."
            },
            {
              "title": "Literature grounding and novelty ledger",
              "description": "Using the literature and preprint connectors, ground each major Phase-1 finding against published EoE literature and classify it confirmatory / novel / contradictory with real retrieved citations (load literature-review skill for citation discipline). Focus on the novel/contradictory calls: IFN dominance over pure-Th2, epithelial antigen presentation, SIGLEC6 mast targeting, S100A8/A9 alarmin axis, and the FLG direction reversal in GSE278888. Deliverables: novelty_ledger.csv (finding, class, citations, confidence), novelty_summary.md. Quality bar: no fabricated citations; contradictory findings flagged for independent confirmation."
            },
            {
              "title": "Independent-dataset confirmation of novel/contradictory findings",
              "description": "For each novel or contradictory item in the ledger, re-test in independent datasets not used to derive it (unused single-cell series GSE126250/GSE249276/GSE201153, additional bulk, or CELLxGENE oesophagus atlas). Confirm or retract the FLG reversal, the IFN dominance, and epithelial antigen presentation. Deliverables: confirmation_tests.png (per-finding replication), confirmation_results.csv. Quality bar: each finding marked replicated / not-replicated with the independent dataset named."
            },
            {
              "title": "Target–biomarker pairing dossier",
              "description": "For each Tier-A target, assemble a one-page dossier: cascade node, endotype it marks, addressable patient fraction (from positivity analysis), PPI-reversibility, genetic support, mechanistic role, candidate predictive/companion biomarker, and explicit differentiation vs broad IL-4/13 blockade (dupilumab) including the targeted+stratified advantage. Deliverables: target_biomarker_dossier.md, target_biomarker_matrix.csv + summary figure. Quality bar: each target paired with a concrete patient-selection biomarker hypothesis."
            },
            {
              "title": "Protein-therapeutic design specifications",
              "description": "For the top secreted/surface leads (CCL26, IL1RL1/ST2, POSTN, SIGLEC6), pull sequences/structures (UniProt, AlphaFold/PDB), generate ESM-2 embeddings and per-residue analysis, map receptor-ligand interaction interfaces, and write a design brief per target (modality, epitope/interface, developability considerations). Confirm compute/tooling availability for any de novo binder design at this step. Deliverables: design_specs/ per-target briefs, interface_analysis.png, target_sequences.fasta. Quality bar: each brief names the interface to engage and the modality rationale."
            },
            {
              "title": "Consolidate revised target dossier and Phase-2 report",
              "description": "Integrate all Phase-2 evidence into a revised prioritized dossier and report: mechanistic node + trajectory placement, endotype/biomarker package, genetic support, novelty class with citations, PPI-refractoriness, and design readiness per target. Produce a final ranked shortlist for Phase 3 with go/no-go rationale. Deliverables: eoe_phase2_report.md, revised_target_dossier.csv, phase2_summary.png (integrated dashboard). Quality bar: every ranking claim traceable to a Phase-2 artifact; limitations stated."
            }
          ]
        }
      ],
      "id": "phase-0"
    }
  ],
  "desired_outputs": [
    "Mechanistic validation report (epithelial trajectory, IFN deconvolution, differential EoE/GERD/healthy cell-cell communication, mast-cell state)",
    "Novelty ledger: each Phase-1 finding classified confirmatory / novel / contradictory with retrieved citations",
    "GWAS risk-locus overlap table for prioritized targets",
    "Independent-dataset confirmation of novel and contradictory findings",
    "Patient endotype stratification + per-target addressable-fraction biomarker tables",
    "PPI-response target triage (reversible vs refractory)",
    "Target–biomarker pairing dossier per Tier-A target with differentiation vs dupilumab",
    "Structure/sequence design specifications for lead targets (ESM embeddings, interface analysis, binder design brief)",
    "Revised prioritized target dossier consolidating all Phase-2 evidence"
  ],
  "feasibility": {
    "rationale": "Phase 1 delivered a reproducible 567-gene meta-signature, single-cell validation (166k cells), cell-cell communication, druggability annotation, and a Tier-A shortlist — all as artifacts. Phase 2 builds directly on these with established methods (trajectory/velocity, IFN deconvolution, differential ligand-receptor, patient endotyping, GWAS overlap, literature grounding) and connector-backed evidence (human-genetics, literature, clinical-trials, drug-regulatory). The protein-design track depends on ESM-family models and structure prediction; sequence-based analysis and embeddings are in-scope now, while de novo binder design may need compute/tooling confirmed at that step. Main risks: some metadata re-fetches (PPI-responder coding), cross-cohort endotype reproducibility, and eosinophil dropout in droplet scRNA-seq — all flagged and worked around, none blocking.",
    "confidence": "high"
  }
}