arazzo: 1.0.1
info:
  title: NVIDIA NIM BioNeMo Drug Discovery
  summary: Fold a protein from sequence, dock a ligand into the predicted structure, then generate optimized analog molecules.
  description: >-
    A structure-based drug-discovery pipeline chaining three NVIDIA BioNeMo
    NIMs. AlphaFold2 first predicts the 3D structure of a target protein from
    its amino acid sequence; the resulting PDB structure is docked against a
    seed ligand with DiffDock to find candidate binding poses; and MolMIM then
    generates optimized analogs around the seed ligand SMILES to explore the
    chemical neighborhood. Outputs of each stage feed the next, turning a bare
    sequence and seed molecule into a ranked set of docked poses and generated
    candidates. Every step spells out its request inline so the flow can be read
    and executed without opening the underlying OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: biologyApi
  url: ../openapi/nvidia-nim-biology-api-openapi.yml
  type: openapi
workflows:
- workflowId: bionemo-drug-discovery
  summary: Fold a protein, dock a ligand, then generate analog molecules.
  description: >-
    Predicts a protein structure from sequence, docks a seed ligand into it, and
    generates optimized analogs of the seed ligand.
  inputs:
    type: object
    required:
    - apiKey
    - sequence
    - ligandSmiles
    properties:
      apiKey:
        type: string
        description: NVIDIA developer API key (nvapi-...) sent as a Bearer token.
      sequence:
        type: string
        description: Single-letter amino acid sequence of the target protein.
      ligandSmiles:
        type: string
        description: Seed ligand as a SMILES string, used for docking and analog generation.
      numPoses:
        type: integer
        description: Number of docking poses to predict.
        default: 10
      numMolecules:
        type: integer
        description: Number of analog molecules to generate.
        default: 10
  steps:
  - stepId: predictStructure
    description: >-
      Predict the 3D structure of the target protein from its amino acid
      sequence with the AlphaFold2 NIM, returning PDB-format structures.
    operationId: predictProteinStructure
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        sequence: $inputs.sequence
        databases:
        - uniref90
        - mgnify
        - small_bfd
        relax_prediction: true
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      predictedPdb: $response.body#/pdbs/0
  - stepId: dockLigand
    description: >-
      Dock the seed ligand against the predicted protein structure with
      DiffDock to predict candidate binding poses.
    operationId: dockLigand
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        protein: $steps.predictStructure.outputs.predictedPdb
        ligand: $inputs.ligandSmiles
        num_poses: $inputs.numPoses
        time_divisions: 20
        steps: 18
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      ligandPositions: $response.body#/ligand_positions
      topPoseConfidence: $response.body#/position_confidence/0
  - stepId: generateAnalogs
    description: >-
      Generate optimized analog molecules around the seed ligand SMILES with
      MolMIM, optimizing for drug-likeness (QED).
    operationId: generateMolecules
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        smi: $inputs.ligandSmiles
        algorithm: CMA-ES
        num_molecules: $inputs.numMolecules
        property_name: QED
        minimize: false
        min_similarity: 0.3
        particles: 30
        iterations: 10
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      generated: $response.body#/generated
      topAnalogSmiles: $response.body#/generated/0/sample
      topAnalogScore: $response.body#/generated/0/score
  outputs:
    predictedPdb: $steps.predictStructure.outputs.predictedPdb
    topPoseConfidence: $steps.dockLigand.outputs.topPoseConfidence
    topAnalogSmiles: $steps.generateAnalogs.outputs.topAnalogSmiles