AQAffinity Building Block

Overview

AQAffinity is part of the AQFold suite — SandboxAQ's family of structure and interaction prediction tools built on folding-based AI (internal codename: Origami). AQAffinity specifically focuses on high-throughput binding affinity prediction between protein targets and small molecule ligands. Unlike traditional virtual screening pipelines that require three separate, disconnected tools — structural characterization, docking, and affinity scoring — AQAffinity handles folding, posing, and affinity prediction in a single, structure-free step.

No crystal structure is needed. Provide a protein amino acid sequence and a ligand SMILES string, and AQAffinity returns the predicted binding affinity (pIC₅₀), a 3D co-folded complex structure, and a confidence score — in approximately 1–2 minutes.

AQAffinity is available via the SandboxAQ MCP Server — SandboxAQ's fully managed MCP server, optimized for LLM and agentic integration.

What is AQAffinity?

AQAffinity is a layer on top of OpenFold 3 that integrates structural prediction and affinity scoring into a unified model. It is designed for:

• Medicinal Chemists who want to rapidly evaluate compound libraries against novel targets without crystal structure prerequisites.
• Computational Chemists running high-throughput virtual screening campaigns who need a faster, more affordable alternative to Schrödinger FEP+.
• AI/LLM-driven workflows via Claude and other MCP clients, where natural language queries translate directly to affinity predictions.

AQAffinity was validated by members of the OpenFold Consortium in December 2025. Partner companies including TamarindBio confirmed that AQAffinity's binding predictions directly support streamlined virtual screening, improved hit-to-lead optimization, and accelerated small molecule discovery.

How It Works

A single AQAffinity inference call processes one protein–ligand pair through a unified pipeline:

1. MSA Generation: The system uses a pre-computed or user-supplied Multiple Sequence Alignment (MSA) file, or queries the public ColabFold MSA server, to generate structural alignment features.
2. Structure Co-Folding: AQAffinity co-folds the protein and ligand simultaneously using OpenFold 3 as the structural backbone, producing a predicted 3D complex.
3. Affinity Scoring: The co-folded structure is passed through the AQAffinity scoring head, which returns a predicted pIC₅₀ value and a model confidence metric.
4. Output: A JSON response containing the affinity score, confidence metric, and a path to the 3D complex structure file (PDB format).

Inference time is approximately 1–2 minutes per query on NVIDIA H100 (80GB) hardware, depending on sequence length.

Using AQAffinity

Setup Overview

No infrastructure provisioning required. Compute is managed entirely by SandboxAQ's backend.

1. Request Access: Contact mcp-prod@sandboxaq.com for enterprise onboarding.
2. Configure Claude Desktop: Add the MCP server to your claude_desktop_config.json. The run_affinity_prediction tool is available on the same endpoint as all other SandboxAQ MCP tools.
3. Verify: Ask Claude: "What tools does the SandboxAQ MCP Server provide?" — it should list run_affinity_prediction with parameter descriptions.
4. Run your first query: Use a prompt like: "Predict the binding affinity of [SMILES] against this protein sequence: [sequence]."

// claude_desktop_config.json
"mcpServers": {
  "sandboxaq-mcp": {
    "type": "http",
    "url": "https://mcp.sandboxaq.com/aqfold/v1",
    "headers": {
      "Authorization": "Bearer YOUR_API_KEY",
      "X-Tenant-ID": "YOUR_TENANT_ID"
    }
  }
}

Request & Response Schemas

Request Schema

{
  "request_id": "req-12345",
  "protein": {
    "sequence": "MTEYKLVVVGAGGVGKSALTIQLIQNHFVDEYDPTIEDSYRKQVVIDGETCLLDILDTAGQ",
    "msa_path": "s3://bucket/path/to/optional_precomputed.msa"
  },
  "ligands": [
    { "id": "cmpd_1", "smiles": "CC1=C(C=C(C=C1)NC(=O)C2=CC=C(C=C2)CN3CCN(CC3)C)..." },
    { "id": "cmpd_2", "smiles": "COC1=C(C=C2C(=C1)N=CN=C2NC3=CC(=C(C=C3)F)Cl)OCCCN4CCOCC4" }
  ],
  "parameters": {
    "num_recycles": 3
  }
}

Response Schema

{
  "request_id": "req-12345",
  "status": "completed",
  "results": [
    {
      "ligand_id": "cmpd_1",
      "affinity_score_pIC50": 8.4,
      "complex_structure_path": "s3://bucket/output/req-12345_cmpd_1.pdb",
      "confidence_metric": 0.89
    },
    {
      "ligand_id": "cmpd_2",
      "affinity_score_pIC50": 6.2,
      "complex_structure_path": "s3://bucket/output/req-12345_cmpd_2.pdb",
      "confidence_metric": 0.75
    }
  ],
  "execution_time_seconds": 124
}

MSA Input Options

AQAffinity supports two MSA strategies:

• User-supplied MSA files: Provide the msa_path field pointing to pre-computed colabfold_msa.npz, colabfold_templates.npz, uniref90_hits.a3m, runner.yaml, and L1000.json artifacts. This is the recommended approach for enterprise deployments to avoid public server dependencies.
• ColabFold MSA server: Omit msa_path to have the system query the public ColabFold MSA server automatically. Note: high usage may result in rate limiting by the public server.

Pricing

AQAffinity uses an all-inclusive tiered pricing model. SandboxAQ absorbs all cloud compute costs. You pay one predictable rate per molecule, scaled by protein sequence length. No upfront commitment is required.

Limits & Quotas

Computational requirements scale with protein sequence length. The following constraints are enforced to maintain response times within LLM timeout windows.

Troubleshooting

• 422 Invalid SMILES: Validate your SMILES strings using RDKit or a similar tool before submission. The response body will identify which compound failed and why.
• 413 Sequence too long: The protein sequence exceeds 512 AA and has been automatically rejected. Contact mcp-support@sandboxaq.com for guidance on large targets.
• 504 MSA timeout: The public ColabFold MSA server did not respond within the timeout window. Provide a pre-computed MSA file (msa_path) to bypass the dependency.
• Low confidence score: Predictions below a confidence of 0.5 indicate the query may be outside the model's applicability domain. Treat these predictions as exploratory only.
• 402 Insufficient Balance: Pre-paid credit balance is zero. Contact mcp-support@sandboxaq.com to add funds.

Security & IP Protection

• Zero weight exposure: AQAffinity model weights are never transmitted via the API. All inference runs on SandboxAQ-managed infrastructure.
• Tenant isolation: Usage, billing, and data are isolated by Tenant ID. Cross-tenant access is not possible.
• No training on your data: SandboxAQ does not use your protein sequences, ligand SMILES, or results to train or fine-tune any model. This is an inference-only deployment.
• Audit logs: Usage logs (timestamp, Tenant ID, molecule count, sequence tier, compute duration) are available to account administrators for billing reconciliation.
• Privacy: See the SandboxAQ Privacy Policy for complete details on data handling and retention.