DockingDiffDock

What is Chai-1? AI Protein-Ligand Complex Prediction Explained

Learn how Chai-1 predicts protein-ligand complexes with state-of-the-art accuracy. Compare to AlphaFold3 and DiffDock, explore use cases, and try it via SciRouter API.

Ryan Bethencourt
May 4, 2026
10 min read

What Is Chai-1?

Chai-1 is an open-source AI model for predicting the 3D structures of biomolecular complexes. Developed by Chai Discovery, it can predict how proteins bind to small molecule drugs, nucleic acids, and other proteins with accuracy comparable to AlphaFold3 while being fully open-source and commercially available.

Unlike traditional molecular docking tools that treat the protein as rigid and search for ligand poses, Chai-1 jointly predicts the structure of both the protein and its binding partner. This means it can model induced-fit effects where the protein changes shape upon binding — a critical phenomenon in drug discovery that classical methods handle poorly.

Why Chai-1 Matters for Drug Discovery

Predicting how a drug molecule binds to its protein target is one of the central challenges in medicinal chemistry. Traditional docking tools like AutoDock Vina assume a rigid protein structure and search for favorable ligand orientations. This works for some targets but fails when binding causes conformational changes.

Chai-1 changes this paradigm. By predicting the full complex structure from sequence and SMILES inputs, it captures protein flexibility, allosteric effects, and binding-induced conformational changes. This makes it particularly valuable for:

  • Drug-target binding prediction: Predict how a candidate drug sits in the binding pocket, including induced-fit effects
  • Antibody-antigen complexes: Model how antibodies recognize and bind their targets, critical for biologics development
  • Covalent drug design: Predict binding modes for covalent inhibitors that form chemical bonds with the target
  • Protein-nucleic acid interactions: Model how proteins bind DNA or RNA, useful for gene therapy and antisense drug design

Chai-1 vs AlphaFold3 vs DiffDock

Each tool occupies a different niche in the structural prediction landscape. Understanding their strengths helps you choose the right tool for your use case.

Chai-1

Joint structure prediction for arbitrary biomolecular complexes. Open-source weights, commercially licensable. Handles proteins, ligands, nucleic acids, covalent modifications, and multi-chain assemblies. Best when you need the full complex structure predicted from scratch.

AlphaFold3

Similar capabilities to Chai-1 but with restricted access. The model weights are not publicly available for commercial use. Available through a limited API from Google DeepMind. Slightly higher accuracy on some benchmarks, but the access restrictions limit practical use in drug discovery pipelines.

DiffDock

A diffusion-based docking model that predicts ligand binding poses given a fixed protein structure. Faster than Chai-1 for simple docking tasks. Does not predict protein conformational changes. Best when you already have a reliable protein structure and need rapid pose prediction.

Note
For most drug discovery workflows, start with DiffDock for rapid screening of many compounds, then use Chai-1 for detailed analysis of top candidates where induced-fit effects matter.

Using Chai-1 via the SciRouter API

SciRouter provides Chai-1 as an API endpoint, so you do not need to download model weights, configure GPU infrastructure, or manage dependencies. Submit a protein sequence and a ligand SMILES string, and receive the predicted complex structure.

Predict a protein-ligand complex with Chai-1
import os, requests, time

API_KEY = os.environ["SCIROUTER_API_KEY"]
BASE = "https://api.scirouter.ai/v1"
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

# Submit a protein-ligand complex prediction
job = requests.post(f"{BASE}/complexes/predict", headers=HEADERS, json={
    "model": "chai-1",
    "protein_sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKAV",
    "ligand_smiles": "CC(=O)Oc1ccccc1C(=O)O",  # aspirin
}).json()

print(f"Job ID: {job['job_id']}")

# Poll for results
while True:
    result = requests.get(
        f"{BASE}/complexes/predict/{job['job_id']}", headers=HEADERS
    ).json()
    if result["status"] == "completed":
        break
    if result["status"] == "failed":
        raise RuntimeError(result.get("error", "Unknown error"))
    time.sleep(5)

print(f"Confidence: {result['confidence']:.2f}")
print(f"Binding energy estimate: {result['binding_energy']:.1f} kcal/mol")

# Save the complex structure
with open("complex.pdb", "w") as f:
    f.write(result["complex_pdb"])
print("Complex structure saved to complex.pdb")
Tip
The confidence score ranges from 0 to 1. Predictions above 0.7 are generally reliable. Always validate high-stakes predictions against experimental data when available.

Practical Use Cases

Virtual Screening with Structural Insight

Use Chai-1 to predict binding poses for your top drug candidates after initial filtering with faster tools. The full complex structure reveals specific interactions — hydrogen bonds, hydrophobic contacts, and water-mediated bridges — that explain why a compound binds.

Antibody-Antigen Modeling

Predict how an antibody variable domain recognizes its antigen. This is essential for understanding binding epitopes, guiding CDR engineering, and assessing cross-reactivity. Chai-1 handles the flexibility of CDR loops better than rigid docking approaches.

Covalent Drug Design

Covalent inhibitors form permanent bonds with target residues (typically cysteine). Chai-1 can model these interactions, predicting the pre-reactive complex geometry that traditional docking tools cannot handle without extensive manual setup.

Next Steps

Chai-1 is one of several structure prediction tools available on SciRouter. For single-chain protein folding, use ESMFold for speed or Boltz-2 for multi-chain complexes. For rapid ligand docking against known structures, DiffDock is faster and still highly accurate.

Sign up for a free SciRouter API key and start predicting protein-ligand complexes in minutes. With 500 free credits per month and no infrastructure to manage, it is the fastest way to go from hypothesis to structural insight.

Frequently Asked Questions

What is the difference between Chai-1 and AlphaFold3?

Chai-1 and AlphaFold3 both predict biomolecular complex structures, but Chai-1 is fully open-source and commercially licensable. AlphaFold3 has restricted weights and a limited API. Chai-1 also supports covalent ligands and modified residues out of the box, whereas AlphaFold3 focuses primarily on non-covalent interactions.

How long does a Chai-1 prediction take?

A typical protein-ligand complex prediction with Chai-1 takes 30 to 120 seconds depending on the size of the protein and complexity of the ligand. The SciRouter API handles this asynchronously — you submit a job and poll for results.

Can Chai-1 predict protein-protein interactions?

Yes, Chai-1 can predict protein-protein complexes, protein-ligand complexes, protein-nucleic acid complexes, and multi-chain assemblies. It handles a wide range of biomolecular interaction types in a single unified model.

What input formats does Chai-1 accept?

Through SciRouter, Chai-1 accepts protein sequences in FASTA format and ligands as SMILES strings. For protein-protein docking, provide multiple sequences. The API also accepts PDB files for constrained prediction.

How accurate is Chai-1 compared to traditional docking?

Chai-1 achieves state-of-the-art accuracy on the PoseBusters benchmark, outperforming traditional docking methods like AutoDock Vina on success rate for drug-like ligands. It is particularly strong for flexible binding sites where traditional methods struggle.

Try It Free

No Login Required

Try this yourself

500 free credits. No credit card required.