Boltz-1: Protein Complex Prediction

• Boltz-1 is an open-source deep learning framework re-implemented from AlphaFold 3 with targeted modifications for protein–protein complex prediction.
• It utilizes an Evoformer stack, cross-chain attention, and dedicated interface metrics like ipTM and complex_pLDDT to accurately assess binding interactions.
• Integrated into binder design pipelines, Boltz-1 statistically ranks candidate complexes to prioritize high-fidelity interfaces for further evaluation.

Boltz-1 is an open-source deep learning framework specifically re-implemented for the accurate prediction and evaluation of protein–protein complexes, with a primary objective of supporting de novo protein binder design. Developed as a derivative of the AlphaFold 3 architecture, Boltz-1 is distinguished by its fine-tuning and algorithmic modifications targeting the assessment of interface quality, per-chain and per-interface confidence in highly multi-chain biological assemblies. Its adoption enables robust, structure-based screening in protein engineering pipelines, where the accurate discrimination and prioritization of binder–target pairs is essential for downstream computational and experimental validation (Ding et al., 21 Jan 2026).

1. Theoretical Foundation and Architecture

Boltz-1 builds upon the transformer-based AlphaFold 3 core but incorporates targeted modifications for protein–protein complex prediction. The underlying network includes:

• An “Evoformer” stack that processes paired multiple-sequence alignments (MSAs) and complex-specific representation features.
• A structure refinement module that generates all-atom 3D coordinates iteratively for the composite system.
• Cross-chain attention mechanisms that allow for dynamic inter-chain information flow, essential for multi-chain docking and conformation prediction.
• The inclusion of a dedicated interface-predicted TM-score (ipTM) head, designed to quantify binding interface quality, and fine-tuning of score aggregation across chain boundaries.

Boltz-1 accepts as input primary sequences of both binder and receptor, with optional MSAs. In the referenced TAS1R2 sweet-taste receptor study, MSAs were leveraged for both chains using the --use_msa_server flag, enabling full exploitation of available evolutionary data (Ding et al., 21 Jan 2026).

2. Output Metrics and Quantitative Scoring

Boltz-1 provides a suite of quantitative confidence and quality metrics, integral for complex evaluation and design selection:

Metric	Range	Biological Interpretation
pLDDT	[0,100]	Per-residue confidence in atomic coordinates
complex_pLDDT	[0,1]	Average per-residue pLDDT normalized over all chains
ipTM	[0,1]	Predicted interface TM-score, i.e., interface correctness
complex_pde	≥0	Mean predicted aligned error between chains (in Å)
confidence_score	[0,1]	Weighted average of ipTM and complex_pLDDT

The confidence score is defined (implementation-specific) as:

$$\mathrm{confidence\_score} \approx 0.5 \times \mathrm{ipTM} + 0.5 \times \mathrm{complex\_pLDDT}.$$

A high confidence_score correlates with structurally plausible and tight-binding interfaces.

3. Integration into De Novo Protein Binder Design Pipelines

Boltz-1 has been structurally integrated into iterative binder discovery workflows. The canonical protocol as demonstrated involves:

1. Backbone Generation: RFdiffusion produces thousands of candidate binder backbones, engineered to exploit receptor “hotspot” residues.
2. Sequence Optimization: ProteinMPNN proposes amino-acid sequences for each backbone, followed by structural relaxation.
3. Structure Evaluation: For each candidate complex, Boltz-1 generates an ensemble of five independent structure predictions to statistically sample conformational and scoring variability.
4. Screening and Ranking: Designs are subjected to two hard filtering criteria: (i) confidence_score > 0.6 and (ii) complex_pLDDT > 0.7. Designs passing these criteria are retained for downstream energetic and experimental prioritization (Ding et al., 21 Jan 2026).

This modular integration enables Boltz-1 to serve as a high-throughput screening filter, acting upstream of energetics calculations such as MM/GBSA.

4. Algorithmic and Implementation Details

The operational workflow in Boltz-1 for complex prediction consists of:

• Running the model in complex mode with evolutionary augmentation (i.e., MSA server enabled) to maximize information capture.
• Generating multiple “recycles” (model seeds), each corresponding to an independent sampling of network noise and initial conditions, resulting in a statistically robust distribution of metrics per candidate.
• Producing chain-specific, interface-specific, and global confidence assessments, explicitly reporting ipTM and complex_pLDDT for quantitative comparison.
• Calculating the complex predicted distance error (complex_pde) as the mean predicted aligned error across all inter-chain residue pairs, providing a geometric measure of interface reliability.

The selection of candidate designs is thus based on ensemble statistics, ensuring that retained complexes are not artifacts of a single anomalous model prediction.

5. Quantitative Case Study: Sweet Protein Binders

In the application to TAS1R2 binder design, Boltz-1 enabled precise quantitative comparison among natural (brazzein) and de novo designed binders (Binder1–Binder5):

• ipTM and confidence_score robustly distinguished high-quality complexes (brazzein, Binder2) from inferior candidates, with the highest values (ipTM ≈ 0.55–0.60, confidence_score > 0.75) observed for the most native-like designs.
• complex_pde provided a stringent geometric filter; top candidates displayed values <0.8 Å, reflecting stable interface predictions.
• complex_pLDDT was uniformly high (all >0.78), but subtle distinctions enabled further ranking.
• The workflow illustrated that Boltz-1 metrics identify both overall and interfacial design fidelity, and, when coupled to MM/GBSA, facilitate the selection of diverse final pools including binders with exceptional predicted binding free energy, even when ipTM is not maximized (Ding et al., 21 Jan 2026).

6. Protocol Summary and Researcher Guidance

For effective use of Boltz-1 in protein complex modeling or binder design:

1. Construct receptor–binder models using compatible backbone and sequence generators (e.g., RFdiffusion, ProteinMPNN).
2. Run Boltz-1 with multiple, independent model seeds and MSAs enabled.
3. Extract ipTM, complex_pLDDT, complex_pde, and overall confidence_score for each run.
4. Filter for confidence_score > 0.6 and complex_pLDDT > 0.7; optionally, further rank by ipTM.
5. Advance promising complexes to energetic evaluation or experimental validation.

This protocol ensures that high-fidelity, confidently modeled interfaces—distinguished by Boltz-1 scoring—are prioritized for further study, optimizing the likelihood of successful de novo binder discovery (Ding et al., 21 Jan 2026).

7. Comparative Context and Impact

Boltz-1’s use of dedicated complex-centric metrics (ipTM, complex_pLDDT, complex_pde) and its ensemble-based scoring distinguish it from models relying on isolated chain confidence or undifferentiated global structure metrics. By directly evaluating atomic-level features relevant to binding interfaces and providing robust, interpretable filtering criteria, Boltz-1 addresses longstanding bottlenecks in computationally guided protein–protein interaction design and validation.

Its integration into open-source, experimental-ready workflows, as exemplified in the rational engineering of sweet-taste protein binders, positions Boltz-1 as a key enabling technology for next-generation protein engineering research (Ding et al., 21 Jan 2026).