Changelog

We're excited to announce the integration of AlphaFold 3, DeepMind's latest protein structure prediction model, into the Curie platform.

Key Features

• Higher accuracy for protein-ligand complexes
• Improved pLDDT scores across all protein types
• PAE matrix output for confidence analysis
• Competitive pricing at $0.015/call

Usage Example

result = client.run(
    model="deepmind/alphafold3",
    sequence="MKTIIALSYIFCLVFA"
)

print(f"pLDDT: {result.plddt}")  # → 0.92
print(f"PAE shape: {result.pae.shape}")  # → (16, 16)

Performance

Average latency: ~800ms on standard sequences (up to 500 residues).

This model is now live in production and available through both Python and Node.js SDKs.

Major release of the curieai Python SDK with breaking changes and new features.

Breaking Changes

• Renamed Client.predict() → Client.run()
• Removed deprecated batch_predict() method
• Updated response format to include job_id in all responses

New Features

Async Support

import asyncio
from curieai import AsyncClient

async def main():
    client = AsyncClient(api_key="sk-...")
    result = await client.run("esm/esmfold-v1", sequence="MKTII...")
    print(result.plddt)

asyncio.run(main())

Streaming for Long Jobs

For models with latency >5s, enable streaming to receive progress updates:

for update in client.run_stream("chroma/generate", description="..."):
    print(f"Progress: {update.progress}%")

Type Hints

Full type coverage with py.typed marker for better IDE support.

Migration Guide

Update your imports and method calls:

# Before
from curie import Client
result = client.predict(model="...", input="...")

# After
from curieai import Client
result = client.run(model="...", sequence="...")

Install the latest version: pip install --upgrade curieai

Significant infrastructure upgrades to improve reliability and reduce latency across all endpoints.

Performance Improvements

• 40% reduction in P95 latency for /v1/run endpoint
• 99.97% uptime achieved over the past 30 days
• Cold start optimization reducing initial request time by 2.3s

Rate Limiting

New rate limits to ensure fair usage:

| Tier | Requests/min | Burst | |------|-------------|-------| | Free | 10 | 20 | | Pro | 100 | 200 | | Enterprise | Unlimited | Unlimited |

Rate limit headers now included in all responses:

X-RateLimit-Limit: 100
X-RateLimit-Remaining: 95
X-RateLimit-Reset: 1622548800

Error Handling

Improved error messages with actionable guidance:

{
  "error": {
    "code": "invalid_sequence",
    "message": "Sequence contains invalid amino acid 'X' at position 42",
    "suggestion": "Use standard 20 amino acid codes (A-Z excluding B,J,O,U,X,Z)"
  }
}

Monitoring

New status page available at status.curie.sh with real-time metrics:

• API availability
• Model-specific latency
• Incident history

MIT's DiffDock model is now available on Curie for protein-ligand docking predictions.

What is DiffDock?

DiffDock uses diffusion models to predict how small molecules (ligands) bind to protein targets. This is critical for drug discovery and understanding molecular interactions.

Input Format

Provide a protein structure (PDB format) and a ligand (SMILES string):

result = client.run(
    model="mit/diffdock",
    protein_pdb=open("protein.pdb").read(),
    ligand_smiles="CC(=O)OC1=CC=CC=C1C(=O)O"  # Aspirin
)

Output

Returns multiple binding poses with confidence scores:

for pose in result.poses:
    print(f"Pose {pose.rank}: confidence={pose.confidence:.3f}")
    with open(f"pose_{pose.rank}.pdb", "w") as f:
        f.write(pose.pdb)

Pricing

$0.010/call — competitively priced for high-throughput screening.

Benchmarks

• Top-1 success rate: 37.8% (ligand RMSD < 2Å)
• Top-5 success rate: 52.1%
• Average runtime: ~500ms per docking job

Perfect for virtual screening campaigns and hit identification.