IsoDDE vs Chai-1 vs Boltz-2: Which Complex Prediction Tool Should You Use for Drug Discovery?

The Race to Predict How Drugs Bind Their Targets

In February 2026, Isomorphic Labs unveiled IsoDDE, a unified AI drug design engine that doubled AlphaFold 3's accuracy on protein-ligand complex prediction. The AI and drug discovery communities immediately took notice. Some researchers began calling it "AlphaFold 4," recognizing it as the next leap in structure-based drug design. IsoDDE represents the most accurate biomolecular complex predictor ever built.

There is one problem: nobody outside of Isomorphic Labs can use it. IsoDDE is fully proprietary with no public weights, no API, and no academic server. For the vast majority of researchers, biotech teams, and AI agents building drug discovery pipelines, IsoDDE is impressive but inaccessible. This raises a critical question: what are the best tools you can actually use today?

This comparison examines IsoDDE alongside two open-source alternatives that are available right now: Chai-1 from Chai Discovery and Boltz-2 from MIT. We cover accuracy, licensing, speed, API access, and why the open-source ecosystem matters more than ever for computational drug discovery.

What Is IsoDDE?

IsoDDE (Isomorphic Drug Design Engine) was announced by Isomorphic Labs in February 2026. It is a unified AI system that handles multiple aspects of drug design within a single model, including protein-ligand binding prediction, antibody-antigen complex modeling, and molecular property optimization. Unlike AlphaFold 3, which was primarily a structure prediction tool, IsoDDE integrates structure prediction with drug design scoring into one architecture.

• Developer: Isomorphic Labs (an Alphabet company, spun out of DeepMind)
• Announced: February 2026
• Architecture: unified transformer-diffusion model for drug design
• Key claim: 2x improvement over AlphaFold 3 on protein-ligand accuracy; 2.3x on antibody-antigen complexes
• Weights: not publicly available (fully proprietary)
• Access: internal use only at Isomorphic Labs and select pharma partners
• License: proprietary; no academic or commercial access
• API: none

Isomorphic Labs reported that IsoDDE achieves state-of-the-art results across multiple drug discovery benchmarks. On protein-small molecule binding pose prediction, the model reportedly places ligands within 1 angstrom RMSD of the crystal pose for over 60 percent of targets, compared to approximately 30 to 40 percent for AlphaFold 3. On antibody-antigen interface prediction, IsoDDE achieved DockQ scores 2.3 times higher than AlphaFold 3 on challenging targets with limited evolutionary information.

What Is Chai-1?

Chai-1 was released by Chai Discovery in September 2024 as an open-source multi-modal foundation model for molecular structure prediction. It focuses on protein-ligand and protein-protein complex prediction, with particularly strong performance on drug-like small molecule binding poses. Chai-1 is commercially licensable and available through multiple access methods including local GPU deployment and cloud APIs.

• Developer: Chai Discovery
• Released: September 2024
• Architecture: modified diffusion model with trunk transformer and pair representation
• Weights: publicly available (open source)
• Access: local GPU, Chai Discovery web server, or SciRouter API
• License: open source, commercial use permitted
• GPU requirement: A100 80GB minimum
• Inputs: proteins, small molecules (SMILES), protein-protein, protein-nucleic acid

Chai-1 was specifically optimized during training for protein-small molecule interactions. Its trunk transformer architecture processes sequence and pair-level features simultaneously, capturing long-range interactions between protein and ligand during the denoising process. On the PoseBusters benchmark, Chai-1 achieves approximately 40 to 50 percent of ligand poses within 2 angstroms RMSD of the crystal pose, which is competitive with AlphaFold 3 and significantly above traditional docking methods.

What Is Boltz-2?

Boltz-2 was developed by researchers at MIT in collaboration with Genesis Therapeutics and released as a fully open-source biomolecular complex prediction model. Its distinguishing feature is confidence-guided diffusion, where predicted per-residue and per-atom confidence scores are fed back into subsequent denoising steps to steer the model toward higher-quality structures.

• Developer: MIT / Genesis Therapeutics
• Released: 2024-2025
• Architecture: confidence-guided diffusion with pairwise attention
• Weights: publicly available (open source)
• Access: local GPU deployment or SciRouter API
• License: open source, commercial use permitted
• GPU requirement: A100 (40GB or 80GB)
• Inputs: proteins, DNA, RNA, small molecules

Boltz-2 achieves comparable accuracy to AlphaFold 3 on standard protein-ligand benchmarks while being fully open and deployable on standard GPU hardware. The confidence feedback loop helps the model avoid low-confidence regions of structure space, producing more physically plausible poses for flexible binding sites. It also supports broader molecular types including DNA and RNA complexes.

Head-to-Head Comparison

Accuracy

IsoDDE claims the highest accuracy by a significant margin, with reported 2x improvements over AlphaFold 3 on protein-ligand predictions. However, these benchmarks are self-reported and not independently verified. Chai-1 and Boltz-2 both achieve state-of-the-art results that have been validated by the research community on standard benchmarks like PoseBusters and CASP15 targets. On protein-ligand binding poses, both open-source models place ligands within 2 angstroms RMSD for 40 to 50 percent of targets. For most drug discovery applications, this level of accuracy is sufficient for lead identification and prioritization.

Licensing and Access

This is where the three tools diverge most dramatically:

• IsoDDE: fully proprietary. No public weights, no API, no academic access. Only available to Isomorphic Labs and its pharmaceutical partners. Researchers cannot use, benchmark against, or build upon this model.
• Chai-1: fully open source with downloadable weights. Commercial use is permitted. Available through local deployment, the Chai Discovery web server, and the SciRouter API.
• Boltz-2: fully open source with a permissive license. Commercial use permitted. Deployable on any compatible GPU or accessible through the SciRouter API.

Speed and Infrastructure

IsoDDE inference speed is undisclosed, as it runs exclusively on Isomorphic Labs' internal infrastructure. Chai-1 takes approximately 2 to 8 minutes per complex on an A100 80GB GPU, with larger systems taking longer due to quadratic attention scaling. Boltz-2 typically runs in 1 to 5 minutes on an A100 and can sometimes converge faster for high-confidence targets thanks to its adaptive confidence feedback loop. Through SciRouter, both Chai-1 and Boltz-2 predictions complete in 30 seconds to 8 minutes depending on complexity.

Input Types

• IsoDDE: reportedly handles proteins, small molecules, antibody-antigen systems, and molecular property optimization (details undisclosed)
• Chai-1: proteins, small molecules (SMILES), protein-protein complexes, protein-nucleic acid complexes
• Boltz-2: proteins, DNA, RNA, small molecules with the broadest molecular type coverage among available tools

Drug Discovery Integration

For building automated drug discovery pipelines, API access is essential. IsoDDE offers no programmatic interface. Chai-1 and Boltz-2 are both available through REST APIs, making them suitable for integration with LangChain agents, MCP-connected LLMs, virtual screening workflows, and automated lead optimization pipelines. Through SciRouter, both models share a consistent API format, enabling consensus scoring across models with minimal code changes.

Why Open-Source Alternatives Matter

IsoDDE's accuracy improvements are impressive, but accessibility determines real-world impact. The history of AI in science shows that open models drive more innovation than closed ones. AlphaFold 2 transformed structural biology not because it was the most accurate model ever built, but because DeepMind released the weights and code, enabling hundreds of thousands of researchers to use it. The AlphaFold Protein Structure Database made over 200 million predicted structures freely available.

By contrast, AlphaFold 3 took a step backward in openness, restricting access to a web server with daily limits and a non-commercial license. IsoDDE goes even further, offering no public access at all. For researchers at academic institutions, biotech startups, and AI agent developers, the practical choice is between the tools they can actually use.

Open-source models like Chai-1 and Boltz-2 offer several advantages beyond accessibility:

• Reproducibility: published results can be independently verified and reproduced on the same model versions
• Fine-tuning: weights can be adapted to proprietary datasets for specific therapeutic areas or target classes
• Transparency: model behavior can be studied, failure modes identified, and confidence calibration validated
• Pipeline integration: programmatic API access enables automated workflows, batch processing, and agent-driven discovery
• Cost control: self-hosted deployment gives full control over compute costs without per-prediction pricing from a proprietary provider

Accessing Chai-1 and Boltz-2 via SciRouter

SciRouter provides both Chai-1 and Boltz-2 through a unified API. No GPU setup, no model weight downloads, no container orchestration. One API key gives you access to both models plus 40 additional scientific computing tools.

Predict a Protein-Ligand Complex with Chai-1

import requests
import time

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"
headers = {"Authorization": f"Bearer {API_KEY}"}

# Predict a kinase-inhibitor complex with Chai-1
response = requests.post(
    f"{BASE}/complexes/chai1",
    headers=headers,
    json={
        "protein_sequence": "MTEYKLVVVGAGGVGKSALTIQLIQ...",  # KRAS sequence
        "ligand_smiles": "CC1=C(C=C(C=C1)NC(=O)C2=CC=C(C=C2)CN3CCN(CC3)C)NC4=NC=CC(=N4)C5=CN=CC=C5",
        "num_samples": 5
    }
)
job_id = response.json()["job_id"]

# Poll for results
while True:
    result = requests.get(
        f"{BASE}/complexes/chai1/{job_id}",
        headers=headers
    ).json()
    if result["status"] == "completed":
        print(f"Top pose confidence: {result['confidence']:.3f}")
        print(f"DockQ score: {result['dockq_score']:.3f}")
        with open("chai1_complex.pdb", "w") as f:
            f.write(result["pdb"])
        break
    elif result["status"] == "failed":
        print(f"Error: {result['error']}")
        break
    time.sleep(15)

Predict the Same Complex with Boltz-2

import requests
import time

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"
headers = {"Authorization": f"Bearer {API_KEY}"}

# Predict the same complex with Boltz-2
response = requests.post(
    f"{BASE}/proteins/complex",
    headers=headers,
    json={
        "model": "boltz2",
        "chains": [
            {
                "type": "protein",
                "sequence": "MTEYKLVVVGAGGVGKSALTIQLIQ..."  # KRAS sequence
            }
        ],
        "ligands": [
            {"smiles": "CC1=C(C=C(C=C1)NC(=O)C2=CC=C(C=C2)CN3CCN(CC3)C)NC4=NC=CC(=N4)C5=CN=CC=C5"}
        ]
    }
)
job_id = response.json()["job_id"]

# Poll for results
while True:
    result = requests.get(
        f"{BASE}/proteins/complex/{job_id}",
        headers=headers
    ).json()
    if result["status"] == "completed":
        print(f"Complex confidence: {result['confidence']:.3f}")
        print(f"Interface pTM: {result['interface_ptm']:.3f}")
        with open("boltz2_complex.pdb", "w") as f:
            f.write(result["pdb"])
        break
    elif result["status"] == "failed":
        print(f"Error: {result['error']}")
        break
    time.sleep(10)

Consensus Scoring Across Both Models

A powerful workflow is to run both Chai-1 and Boltz-2 on the same target and compare predictions. When both models agree on a binding pose, confidence increases substantially. When they disagree, it flags targets that need additional investigation or experimental validation.

# After collecting results from both models:
chai1_confidence = chai1_result["confidence"]
boltz2_confidence = boltz2_result["confidence"]

# Consensus classification
if chai1_confidence > 0.7 and boltz2_confidence > 0.7:
    print("HIGH CONFIDENCE: both models agree on binding pose")
    print("Proceed to lead optimization")
elif chai1_confidence > 0.7 or boltz2_confidence > 0.7:
    print("MIXED: one model confident, review poses manually")
    print("Consider running DiffDock as a third opinion")
else:
    print("LOW CONFIDENCE: target needs experimental validation")
    print("Consider alternative binding hypotheses")

When to Use Which Tool

The right choice depends on your situation, your licensing constraints, and whether you need programmatic access:

• Use Chai-1 when predicting protein-ligand complexes for drug discovery, when you need strong small molecule binding pose accuracy, for commercial projects, or when GPU memory is not a constraint (requires A100 80GB)
• Use Boltz-2 when you need broader molecular type support (DNA, RNA complexes), when running on a 40GB GPU, for large-scale screening campaigns where speed matters, or when confidence-guided predictions help prioritize targets
• Use both Chai-1 and Boltz-2 for high-value targets where consensus scoring increases confidence, or when building robust automated pipelines that benefit from multi-model agreement
• Use DiffDock when you have a known binding site and need fast molecular docking rather than full complex prediction
• IsoDDE is not currently an option for external researchers, developers, or commercial teams

The Bottom Line

IsoDDE represents a genuine advance in AI-driven drug design, with reported accuracy improvements that could accelerate drug discovery at Isomorphic Labs and its partners. But for the broader scientific community, the tools that matter are the ones you can access, integrate, and build upon.

Chai-1 and Boltz-2 deliver state-of-the-art complex prediction accuracy with full transparency, commercial licensing, and API access. Through SciRouter, you can run both models with a single API key, compare results via consensus scoring, and integrate predictions into automated pipelines alongside DiffDock for docking, Chai-1 for complex prediction, and 40 other scientific computing tools.

Get started with a free API key at scirouter.ai/pricing (5,000 calls per month, no credit card required). For more background, read our Chai-1 vs AlphaFold 3 vs Boltz-2 comparison and introduction to Boltz-2.