Boltz-2 Molecular Predictor — Structure & Affinity Explorer

Beyond Structure: Predicting How Tightly Drugs Bind

Boltz-2 is the first deep learning model to approach the accuracy of physics-based free-energy perturbation (FEP) methods for binding affinity prediction — while running 1,000× faster. Open-sourced by MIT Jameel Clinic and Recursion Pharmaceuticals.

1,000×

Faster than FEP

Approaches FEP accuracy at a fraction of the computational cost

0.62

Pearson r (FEP+ benchmark)

Comparable to OpenFE on unseen protein targets

CASP16 Affinity

Outperforms all submitted methods across 140 complexes

3.5M+

Training Data Points

Curated from ChEMBL, PubChem, BindingDB, CeMM, MIDAS

Why This Matters: Small molecule binding affinity — how tightly a drug sticks to its protein target — is the single most measured property in early-stage drug R&D. Until now, only slow physics simulations (FEP) could predict it accurately. Boltz-2 changes that.

🧬 What Boltz-2 Predicts

3D Complex Structures — protein-ligand, protein-protein, protein-DNA/RNA, antibody-antigen
Binding Affinity (continuous) — log₁₀(IC₅₀) for hit-to-lead optimization
Binder Probability (binary) — active vs. decoy classification for hit discovery
Molecular Dynamics — RMSF prediction competitive with AlphaFlow/BioEmu
Confidence Scores — iPTM, pLDDT for reliability assessment

🔑 Key Innovations

Unified Structure + Affinity — first co-folding model to jointly predict both
Method Conditioning — specify X-ray, NMR, or MD emulation mode
Template Steering — input reference structures as prior knowledge
Contact Constraints — enforce specific distance constraints
MD Ensemble Training — trained on MISATO, ATLAS, mdCATH dynamics

Discovery Use Cases

🎯

Hit Discovery

Screen large chemical libraries. Boltz-2 discriminates binders from decoys on MF-PCBA benchmark, doubling average precision over docking/ML methods.

⚗️

Hit-to-Lead

Rank-order compounds by binding affinity. Approaches FEP accuracy (Pearson 0.62) on the protein-ligand-benchmark at 1,000× the speed.

🧪

Lead Optimization

Predict how small chemical modifications affect binding. Pairwise intra-assay difference training captures subtle SAR relationships.

🔬

De Novo Generation

Paired with SynFlowNet generative model: top-10 TYK2 compounds all predicted to bind via ABFE simulation. Validated generative design workflow.

📅 Boltz Timeline

Nov 2024

Boltz-1 Released — First fully open-source model to approach AlphaFold3 accuracy. MIT license.

Mar 2025

Boltz-1x — Added inference-time steering with physics-based potentials for improved physical quality.

Jun 2025

Boltz-2 Released — Unified structure + affinity prediction. First DL model to approach FEP accuracy. 1,000× faster.

Jun 2025

NVIDIA NIM Integration — Boltz-2 available as NVIDIA NIM microservice for cloud inference.

Jul 2025

PMC Publication — Full manuscript published in bioRxiv. Community adoption across top 20 pharma companies.

Model Architecture

Boltz-2 extends the AlphaFold3/Boltz-1 architecture with an affinity module, controllability features, and GPU optimizations. Four main components: Trunk → Denoising → Confidence → Affinity.

Architecture Components

Component	Role	Key Details	New in Boltz-2?
Trunk	Core representation learning	PairFormer stack with triangle attention, bf16, trifast kernel, 768-token crops	Optimized ✓
MSA Module	Multiple sequence alignment processing	Column attention over evolutionary sequences	—
Diffusion Module	Structure generation via denoising	Atom transformer, AF3 σ hyperparameters	Updated ✓
Steering (Boltz-2x)	Physical quality enforcement	Inference-time physics potentials, clash removal	✓
Method Conditioning	Experimental mode specification	X-ray / NMR / MD emulation modes	✓ New
Template Steering	Prior structure integration	Multi-chain templates, soft/strict modes	✓ New
Confidence Module	Prediction reliability	iPTM, pLDDT, PAE, pDE scores	—
Affinity Module	Binding strength prediction	Pocket PairFormer → binary head + affinity value head	✓ New
B-factor Head	Local dynamics prediction	Supervised on experimental + MD B-factors	✓ New

🔢 Training Data Sources

⚡ Compute Optimizations

Mixed precision (bf16) — Reduced memory, faster matmuls
trifast kernel — Optimized triangle attention
768-token crops — Matching AF3 training scale
cuEquivariance — NVIDIA GPU acceleration
Pre-computed pockets — Efficient affinity training

Drug Discovery Pipeline

From protein sequence to binding affinity prediction — the complete Boltz-2 workflow for computational drug discovery.

Input Preparation

Define the biomolecular system in YAML format: protein sequences, ligand SMILES, DNA/RNA chains. Optionally provide MSA, templates, method conditioning, and contact constraints.

sequences: - protein: id: A sequence: "MSKGEELFTG..." - ligand: id: B smiles: "CC(=O)Nc1ccc(O)cc1" predict_affinity: true method: xray

MSA Generation

ColabFold MSA server generates multiple sequence alignments from protein databases. Evolutionary co-variation patterns inform structural contact predictions.

Trunk Processing

The PairFormer stack processes pair and single representations through triangle attention layers (bf16, trifast kernel). B-factor supervision captures local dynamics. Crop size: 768 tokens.

Structure Prediction (Denoising)

Diffusion module generates 3D coordinates through iterative denoising. Atom transformer refines all-atom positions. Boltz-2x applies physics-based steering to remove clashes and fix stereochemistry.

Confidence Scoring

Confidence module outputs iPTM (interface predicted TM-score), pLDDT (predicted local distance difference test), PAE (predicted aligned error), and pDE (predicted distance error) for each prediction.

Binding Affinity Prediction

The affinity module's pocket PairFormer processes protein-ligand interactions. Two output heads: binary P(binder) for hit discovery screening, and log₁₀(IC₅₀) continuous affinity for lead optimization.

Generative Design (Optional)

Coupled with SynFlowNet for de novo molecule generation. Iteratively generates and scores synthesizable compounds. Validated on TYK2 target: all top-10 compounds predicted as binders by ABFE simulations.

Speed Comparison: Boltz-2 vs Traditional Methods

📥 Input Modalities

Input	Format	Required?
Protein	Amino acid sequence	✓
Ligand	SMILES / SDF	For affinity
DNA	Nucleotide sequence	Optional
RNA	Nucleotide sequence	Optional
MSA	Auto-generated or custom	Recommended
Templates	CIF files	Optional
Constraints	Distance / contact pairs	Optional

📤 Output Predictions

Output	Unit	Use Case
3D Structure	CIF coordinates	Visualization, docking
P(binder)	0 → 1 probability	Hit discovery
Affinity value	log₁₀(IC₅₀) in μM	Lead optimization
iPTM	0 → 1	Interface quality
pLDDT	0 → 100	Local confidence
B-factors	Å²	Flexibility

Benchmarks & Performance

Boltz-2 sets new standards across structure prediction, binding affinity, and virtual screening benchmarks.

Headline Result: On the CASP16 affinity track, Boltz-2 outperforms all submitted competition entries across 140 complexes — out of the box, with no fine-tuning.

Affinity Prediction: FEP+ Benchmark

Pearson correlation on 4-target FEP+ subset (CDK2, TYK2, JNK1, P38). Higher = better. Unseen proteins held out of training.

Structure Prediction: Cross-Modality Accuracy

Structural accuracy across biomolecular modalities. Boltz-2 matches or exceeds Boltz-1 across all categories, with notable gains on antibody-antigen and DNA-protein complexes.

Virtual Screening: MF-PCBA Hit Discovery

Average precision for binder/decoy discrimination in high-throughput screens. Boltz-2 doubles average precision over docking and ML baselines.

Benchmark Summary

Benchmark	Task	Metric	Boltz-2	Best Competitor	Note
FEP+ (4-target)	Affinity (lead opt.)	Pearson r	0.62	0.65 (OpenFE)	1,000× faster than FEP
CASP16 Affinity	Affinity (140 complexes)	Ranking	#1	All submitted methods	Retrospective, no fine-tuning
MF-PCBA	Hit discovery	Avg. Precision	2× baseline	Docking / ML methods	Binder vs decoy screen
Protein-Ligand	Structure	Success rate	≥ Boltz-1	AlphaFold3	Improved over Boltz-1
Antibody-Antigen	Structure	DockQ	Notable gains	Boltz-1	Challenging modality
DNA-Protein	Structure	LDDT	Improved	Boltz-1	Better with distillation
RNA Structure	Structure	TM-score	Improved	Boltz-1	Expanded training data
Dynamics (RMSF)	Flexibility	Correlation	Competitive	AlphaFlow / BioEmu	MD-conditioned mode
TYK2 Generative	De novo design	Top-10 validated	10/10 binders	—	Via ABFE simulation

Interactive Affinity Predictor

Simulate Boltz-2 binding affinity predictions. Select a protein target and ligand, then explore how structural features influence predicted binding strength.

🎯 Select Target & Ligand

Protein Target

Ligand Compound

Prediction Parameters

Method Conditioning

Template Steering

Num. Recycling 3

Num. Samples 5

📊 Prediction Results

−6.82

log₁₀(IC₅₀) μM

IC₅₀ ≈ 151 nM — Moderate Binder

P(binder) 0.87

iPTM (interface quality) 0.78

pLDDT (local confidence) 82.4

Binding Pocket Interactions

🧪 Compound Series Comparison

Simulated SAR (Structure-Activity Relationship) analysis across the compound series for the selected target.

Model Arena

Compare Boltz-2 against leading biomolecular structure prediction and affinity models across key dimensions.

Model	Organization	Structure	Affinity	Open Source	License	Modalities	Speed
Boltz-2	MIT + Recursion	★★★★☆	★★★★★	✅ Full	MIT	Protein, Ligand, DNA, RNA, Ab-Ag	~30s / complex
AlphaFold3	DeepMind	★★★★★	★★☆☆☆	⚠️ Server only	Restricted	Protein, Ligand, DNA, RNA, ions	Server queue
Boltz-1	MIT + Recursion	★★★★☆	—	✅ Full	MIT	Protein, Ligand, DNA, RNA	~25s / complex
Chai-1	Chai Discovery	★★★★☆	★★☆☆☆	✅ Weights	Non-commercial	Protein, Ligand, DNA, RNA	~30s / complex
OpenFold	Columbia	★★★☆☆	—	✅ Full	Apache 2.0	Protein (monomer/multimer)	~45s / chain
FEP+ (Schrödinger)	Schrödinger	—	★★★★★	❌ Commercial	Commercial	Protein-Ligand only	~8h / compound
OpenFE	Open Source	—	★★★★☆	✅ Full	MIT	Protein-Ligand only	~4h / compound
RoseTTAFold2	UW Baker Lab	★★★★☆	—	✅ Full	BSD	Protein, NA, Ligand	~40s / complex
ESMFold	Meta FAIR	★★★☆☆	—	✅ Full	MIT	Protein (single-sequence)	~5s / chain
DiffDock	MIT	★★★☆☆	★★★☆☆	✅ Full	MIT	Protein-Ligand docking	~10s / pose
NeuralPLexer	Iambic	★★★★☆	★★☆☆☆	✅ Weights	Research	Protein-Ligand complexes	~20s / complex
BoltzGen	MIT + Recursion	★★★☆☆	—	✅ Full	MIT	Generative protein design	~15s / sample

Capability Radar

Why Boltz-2 Stands Out

🏆 Only model combining structure + affinity
AlphaFold3 does structure; FEP does affinity. Boltz-2 does both.
🔓 Fully open-source (MIT)
Weights, code, and training pipeline all under MIT license. AF3 is server-only.
⚡ 1,000× faster than FEP
Approaches FEP accuracy in seconds, not hours. Enables large-scale virtual screening.
🎛️ Controllable predictions
Method conditioning, template steering, contact constraints — no retraining needed.
📊 CASP16 #1 in affinity
Outperforms all submitted methods on the 140-complex CASP16 affinity challenge.

References & Resources

Primary literature, code repositories, and community resources for Boltz-2.

Primary Papers

Passaro, S., Corso, G., Wohlwend, J., et al. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv. doi:10.1101/2025.06.14.659707
Wohlwend, J., Corso, G., Passaro, S., et al. (2024). Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv. doi:10.1101/2024.11.19.624167
Abramson, J., Adler, J., Dunger, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500.
Ross, G.A., et al. (2023). Large-scale protein-ligand binding free energy benchmark. J. Chem. Inf. Model.
Buterez, D., et al. (2023). MF-PCBA: Multi-fidelity high-throughput screening benchmarks. NeurIPS Datasets and Benchmarks.
Cretu, A., et al. (2024). SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints. arXiv.
Jing, B., et al. (2024). AlphaFlow: autonomous molecular dynamics with diffusion models. ICML.
Lewis, M., et al. (2025). BioEmu: scalable and accurate biomolecular dynamics with ML. bioRxiv.
Mirdita, M., et al. (2022). ColabFold: making protein folding accessible to all. Nature Methods.
Kim, S., et al. (2023). PubChem 2023 update. Nucleic Acids Research.
Zdrazil, B., et al. (2024). The ChEMBL Database in 2023. Nucleic Acids Research.
Lin, T.Y., et al. (2017). Focal Loss for Dense Object Detection. ICCV.

Resources

💻

GitHub Repository

github.com/jwohlwend/boltz
MIT-licensed code, weights, training pipeline.

📄

Full Manuscript

PDF (jeremywohlwend.com)
Complete technical details and supplementary.

🖥️

NVIDIA NIM

NVIDIA NIM API
Cloud inference via NIM microservice.

🧪

Tamarind Bio

Tamarind web UI
Run Boltz-2 in browser, upload templates.

💬

Slack Community

Join Slack
Discuss with developers and users.

📊

Recursion (RXRX)

rxrx.ai/boltz-2
Industry partner, NASDAQ-listed TechBio.

Citation

@article{passaro2025boltz2, author = {Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and Reveiz, Mateo and Thaler, Stephan and Somnath, Vignesh Ram and Getz, Noah and Portnoi, Tally and Roy, Julien and Stark, Hannes and Kwabi-Addo, David and Beaini, Dominique and Jaakkola, Tommi and Barzilay, Regina}, title = {Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction}, year = {2025}, doi = {10.1101/2025.06.14.659707}, journal = {bioRxiv} }

Built by Q ⚡ · Data sourced from Passaro et al. (2025) bioRxiv · Dashboard for educational exploration