{
  "version": 3,
  "created_at": "2026-07-09T04:22:13.880Z",
  "task_summary": "Design neutralizing binders/ligand-traps for CCL26 and POSTN, validate, and render 3D videos",
  "agents": [],
  "phases": [
    {
      "name": "Plan",
      "delegations": [
        {
          "steps": [
            {
              "title": "Mechanistic rationale + POSTN barrier-restoration literature",
              "description": "Before design, settle the biology. Run a focused literature synthesis (host.llm over targeted PubMed abstracts, grounded in real retrieved sources) on two questions: (1) CCL26/eotaxin-3 as the CCR3-axis eosinophil chemoattractant in EoE — confirm the neutralization rationale and the ligand-trap vs anti-CCR3 tradeoff; (2) whether POSTN blockade could restore esophageal barrier function, including evidence that barrier dysfunction persists in histologic remission and POSTN's role in remodeling/integrin signaling. Deliverable: eoe_design_rationale.md with cited evidence and an explicit design goal per target (what to neutralize, which interface to block). Save as artifact."
            },
            {
              "title": "Fetch and prepare target structures",
              "description": "Pull experimental structures for CCL26 (RCSB via structures-interactions MCP; fall back to AlphaFold) and POSTN (AlphaFold model; isolate the relevant FAS1 domain(s)). Clean chains, select the mature secreted sequence (strip signal peptide), and identify the functional interface to target: CCL26's CCR3-binding / GAG-binding surface, and POSTN's integrin-interaction / fibrillogenesis surface. Deliverable: cleaned target PDBs + a hotspot-residue definition file per target, saved as artifacts, plus a figure showing each target with the targeted epitope highlighted."
            },
            {
              "title": "Build the Modal RFdiffusion+ProteinMPNN GPU env",
              "description": "In the compute_provider kernel, build the bundled proteomics_rfd_diffdock_gpu image (RFdiffusion + ProteinMPNN), hydrate any needed weights into a Modal Volume, and record the resulting im- image ref and volume in the compute_details ledger. Validate with a short GPU smoke probe (load RFdiffusion weights, confirm CUDA). This is a prerequisite for design and is treated as its own validated step."
            },
            {
              "title": "RFdiffusion binder backbone generation",
              "description": "For each target, run RFdiffusion in binder mode against the defined hotspot residues to generate ~50-100 candidate mini-binder backbones (typically 60-100 aa) docked to the target epitope. Dispatch as GPU jobs on Modal; harvest backbones. Deliverable: backbone PDB sets per target saved as a checkpoint artifact, plus a summary of how many backbones passed initial radius-of-gyration/contact filters."
            },
            {
              "title": "ProteinMPNN sequence design on backbones",
              "description": "Inverse-fold each surviving backbone with ProteinMPNN (SolubleMPNN variant for expression-friendly sequences), generating several sequences per backbone across a small temperature sweep, holding the target fixed. Filter on ProteinMPNN score and sequence-recovery. Deliverable: designed-sequence FASTA sets per target + a scored table, saved as artifacts."
            },
            {
              "title": "Fold-back validation of binder-target complexes",
              "description": "Co-fold each designed binder with its target using Boltz-2 (primary) and/or Chai-1/ESMFold2-Fast as a cross-check, on Modal GPU. Compute interface metrics: ipTM, complex pLDDT, interface pAE, and predicted-aligned-error at the binding site. Rank designs; filter to those with ipTM above threshold and a well-packed interface at the intended epitope. Deliverable: eoe_binder_validation.csv ranking all designs with metrics, plus a metrics-distribution figure, saved as artifacts."
            },
            {
              "title": "Select and characterize the lead binder per target",
              "description": "Pick the top-scoring, interface-correct design for CCL26 and for POSTN. Characterize each: report sequence, length, predicted complex structure, interface residues and buried surface area, and confirm the binder engages the intended functional epitope (blocks CCR3 site / integrin site). Save the final complex structure PDB/CIF for each as an artifact (renders in the Mol* viewer), plus a one-page spec sheet per binder."
            },
            {
              "title": "Render 3D rotation videos for the judges",
              "description": "For each lead binder-target complex, render a publication-quality 3D rotation movie with PyMOL (open-source, local CPU): cartoon representation with target and binder in distinct colors, the interface epitope highlighted, a 360-degree turntable plus an optional zoom to the interface, ray-traced frames stitched to mp4 via ffmpeg. Deliverable: ccl26_binder.mp4 and postn_binder.mp4, saved as artifacts."
            },
            {
              "title": "Compile the design deliverable",
              "description": "Assemble a hackathon-ready summary (eoe_binder_design_report.md) tying together: the mechanistic rationale, the two lead designs with sequences and interface metrics, embedded links to the 3D videos and complex structures, honest caveats (in silico only, computational success rates, wet-lab validation needed), and suggested next experimental steps. Save all final artifacts."
            }
          ]
        }
      ],
      "id": "phase-0"
    }
  ],
  "desired_outputs": [
    "Neutralizing binder / ligand-trap design (sequence + structure) for CCL26",
    "Neutralizing binder / ligand-trap design for POSTN",
    "Fold-validated complex models with interface quality metrics for each",
    "3D rotation video (mp4) of each binder-target complex for the hackathon judges",
    "Mechanistic design-rationale brief covering CCL26 (chemokine/CCR3 axis) and POSTN (barrier-restoration evidence)"
  ],
  "feasibility": {
    "rationale": "Both targets are secreted and structurally characterized (CCL26 has experimental structures; POSTN via AlphaFold), the full RFdiffusion→ProteinMPNN→fold-validation toolchain is available on the now-connected Modal GPU, and 3D video rendering is straightforward local work. The honest caveat: de novo binder design has modest in silico success rates — we will generate many candidates and filter hard on interface metrics (ipTM, pLDDT, interface pAE, shape complementarity) to surface a strong exemplar for each target, but these are computational designs requiring wet-lab validation, not finished therapeutics. The POSTN barrier-restoration question is a literature-evidence synthesis, answerable but with a mechanistic-inference caveat.",
    "confidence": "medium"
  }
}