Exclusive compilation, unauthorized use will be pursued.
Protein-protein interactions are at the core of all biological processes. Traditional protein binder design relies on methods such as immunization, antibody library screening, or directed evolution, which are often time-consuming and have a low success rate. Recently, the team led by Bruno E. Correia at the École Polytechnique Fédérale de Lausanne (EPFL) published significant research findings in the journal Nature, developing an open-source automated protein binder design platform called BindCraft, achieving experimental success rates of 10-100% without the need for high-throughput screening to directly obtain nanomolar affinity binders.
How Does AlphaFold2 Become a New Engine for Protein Design?
The core innovation of BindCraft lies in directly utilizing the AlphaFold2 network for the design of protein binders. Traditional physics-based methods like Rosetta allowed for early binder designs but had a success rate of less than 0.1%, requiring thousands of designs to be sampled. The latest deep learning methods like RFdiffusion have significantly improved success rates but still rely on placing rigid target scaffolds without side chains, ultimately depending on AF2 complex predictions to identify potential interactions.
BindCraft adopts a novel design strategy: directly transforming new binders and interfaces through the backpropagation of the AlphaFold2 network. Figure 1a illustrates the schematic of the BindCraft protein binder design pipeline. Given a target protein structure, the AF2 multimer is used to generate the binder scaffold and sequence, followed by optimizing the binder’s surface and core while maintaining the integrity of the interface. Finally, the AF2 monomer model is used to predict and filter the designs.
The design loss function consists of multiple terms, including: (1) binder confidence pLDDT (weight 0.1); (2) interface confidence i_pTM (weight 0.05); (3) normalized predicted alignment error within the binder (weight 0.4); (4) normalized pAE between the binder and the target (weight 0.1); (5) contact loss among residues within the binder (weight 1.0); (6) contact loss between the target and the binder (weight 1.0); (7) radius of gyration of the binder (weight 0.3); (8) helical loss (weight -0.3).
The loss function can be expressed as:
How Does BindCraft Achieve Dynamic Adaptation of Target Proteins?
Unlike methods such as RFdiffusion or RIFdock that keep the target scaffold fixed during the design process, BindCraft re-predicts the binder-target complex at each design iteration. This allows for defined levels of flexibility in the side chains and scaffolds of both the binder and the target, producing scaffolds and interfaces that match the target binding site. Results show that the root mean square deviation (RMSD) of the target scaffold ranges from to (Extended Data Figure 1a shows the **target structures after the initial AF2 binder transformation for different design targets.**
Sequence optimization is conducted in four stages: the first stage optimizes using logits in continuous sequence space, with sequence representations based on a linear combination of where and temperature is 1.0. The second stage normalizes the sequence logits into sequence probabilities using the softmax function, with temperature equal to .
What Breakthroughs Have Been Achieved in Designing Cell Surface Receptor Binding Proteins?
Figure 1b shows an overview of the protein design targets. The green parts in the model are used for design, while the gray areas are excluded. The values in the blue box represent the ratio of the number of successful designs observed in SPR measurements to the total number of designs tested. The values in the yellow box indicate the highest affinity binders measured without experimental sequence optimization, while the values in the orange box represent estimates due to poor fitting.
The team first designed binders targeting the human PD-1 protein, a key immune checkpoint receptor expressed on T cells. They purified and screened 53 designs, detecting binding using biolayer interferometry (BLI) in a bivalent Fc fusion format. Thirteen binders were observed to have binding signals, with the best binder showing an apparent dissociation constant of less than 1 nM. (Figure 2a,b).
Figure 2a shows the design model of binder 2 in complex with PD-1, while Figure 2b displays a representative BLI sensorgram showing the binding kinetics of binder 2 (bivalent Fc fusion) with PD-1. To confirm the binding site, the team performed a competition analysis with the well-characterized anti-PD-1 monoclonal antibody pembrolizumab, which should bind to the same site. Indeed, the designed binder could not compete with the antibody (pM) for binding, indicating it targets an overlapping binding site.
The team also designed binders against PD-L1 and interferon alpha 2 receptor (IFNAR2). Of the nine designs targeting PD-L1, seven showed binding signals, while three of the nine designs targeting IFNAR2 were detected to bind. The top binder for PD-L1 showed an affinity of 615 nM (Figure 2c,d), while the top binder for IFNAR2 showed an affinity of 260 nM (Figure 2e,f).
How to Tackle Protein Targets with Unknown Binding Sites?
The team next tested whether the pipeline could design binders for extracellular receptors lacking clearly defined binding sites. CD45 was chosen as the target due to the structural complexity of its extracellular domain, composed of four immunoglobulin-like domains d1-d4, with heavy N-glycosylation in the minimal isoform. Sixteen binders were experimentally tested, with four showing binding in SPR. The best binder 1 showed an affinity of 14.7 nM, targeting the connecting region between domains d3 and d4 (Figure 2g,h).
Membrane proteins lack clearly defined extracellular domains, which are of significant biological and therapeutic importance, but pose major challenges for binder design due to difficulties in experimental validation and screening. The team proposed that computationally designed soluble analogs could provide a promising solution by preserving natural epitopes, enabling rapid pre-screening of potential binders.
What Application Potential Does BindCraft Show in Allergen Masking?
Allergic rhinitis and seasonal allergies affect up to 50% of the population in some countries. The team designed binders against the dust mite allergens Der f7 and Der f21, as well as the major birch allergen Bet v1, which is responsible for up to 95% of birch-related allergies.
Figure 3a shows the design model of binder 2 targeting the dust mite allergen Der f7 and the SPR binding affinity fit, with binder 2 showing the highest binding affinity of 12.8 nM. To confirm the binding mode of binder 2, the team solved the crystal structure of the complex with Der f7, obtaining two crystal forms with resolutions of 2.2 Å and 3.0 Å (Extended Data Figure 4a,b).
Aligned on the allergen, binder 2 showed a backbone RMSD of 1.7 Å (Figure 3b), confirming the structural accuracy of the design.
Figure 3c shows the design model of binder 10 targeting the dust mite allergen Der f21 and the SPR binding affinity fit, with the best binder 10 showing an apparent affinity of 793 nM. Figure 3d shows the crystal structure of the Der f21-binder 10 complex (colored) superimposed with the design model (gray).
For the birch allergen Bet v1, the team identified two successful designs from seven tested binders. Binder 2 showed a binding affinity of 120 nM (Figure 3e). To assess the specificity of the anti-allergen binders, the team incubated the top binders with three allergens. Even at a concentration of 10 μM binder, no off-target binding to other allergens was observed, indicating that the designed anti-allergen binders have high specificity.
How is the Regulatory Mechanism of Multi-Domain Nucleases Redefined?
Nucleic acid interaction interfaces are widely regarded as undruggable due to their large, charged, and convex surfaces, making them difficult to target with small molecules. The team focused on the multi-domain CRISPR-Cas9 nuclease (SpCas9) from Streptococcus pyogenes.
Figure 4a shows an enlarged view of the REC1 domain of SpCas9 binding to the guide RNA, with the designed binder superimposed in the binding pocket. All six tested binders bound to the full-length apo SpCas9 enzyme. The top binders 3 and 10 showed apparent binding affinities of approximately 300 nM.
To validate the binding mode, the team attempted to solve the cryo-EM structure of binders 3 and 10 in complex with full-length SpCas9 apo enzyme. Despite high data quality and clear observable density for the binders, satisfactory cryo-EM density could not be obtained to build atomic models due to poor resolution in the target region.
Figure 4b,c show the cryo-EM structures of binders 3 and 10 binding to the apo form of SpCas9.
To assess functionality, the team co-transfected CRISPR-SpCas9 with the designed binders or natural Acrs into human embryonic kidney 293T (HEK293T) cells. A significant reduction in gene editing activity was observed in the presence of the designed binders (Figure 4d). They outperformed AcrILC2, which uses a different targeting mechanism to inhibit guide RNA loading.
The team also extended binder design to other large nucleases, designing binders for the multi-domain Argonaute (Ago) nuclease (CbAgo) from Clostridium butyricum. Figure 4e shows the structural architecture of CbAgo in complex with gDNA and tDNA. The team tested twelve binders for their effects on CbAgo-mediated tDNA cleavage, with two binders strongly inhibiting CbAgo activity. (Figure 4f).
Kinetic analysis of cleavage showed: while the cleavage rate of 4 μM CbAgo alone was 0.004 s, in the presence of 2 μM binder 2 and binder 3, it was reduced by 80-fold to s and 40-fold to s (Figure 4g).
How Does AAV Gene Delivery Targeting Technology Achieve Precision Therapy?
Viral vectors, such as those from adeno-associated virus (AAV), expand the possibilities of gene therapy by leveraging the natural ability of viruses to introduce genetic material into cells and tissues. However, AAV has poor specificity for cell types, tissues, and organs.
Figure 5a shows a schematic of AAV-cmv-GFP targeting, replacing the natural primary attachment to cell surface glycans with genetically inserted cell type receptor-specific microprotein binders. The team proposed that BindCraft can effectively design microprotein binders that can re-target AAV to cell type-specific receptors (Figure 5a).
Based on extensive mutagenesis studies of AAV capsid adaptability, the team explored alternative insertion sites between residues 497 and 498 located near the threefold symmetry axis of the AAV capsid (Figure 5b).Figure 5c shows the transduction efficiency of different AAV variants targeting HER2 or PD-L1, determined after transferring the supernatant from packaging cells to HEK293 cells stably overexpressing their respective target receptors.
Figures 5d,e show the design models of binder 1 targeting HER2 and binder 202 targeting PD-L1, respectively. The most effective variants HER2-b1 and PD-L1-b202 exhibited enhanced specificity for cells expressing their target receptors (Figure 5f).
What Are the Limitations of BindCraft Technology and Future Directions?
Despite significant successes, the BindCraft design approach still has limitations. The backpropagation through AF2 is GPU-intensive, and the final design filtering using AF2 monomer in single-sequence mode may exclude potential high-affinity binders. The team evaluated the possibility of filtering using the recently released AlphaFold3 model, but still found a large number of false positive predictions.
Additionally, AF2 is known to be insensitive to point mutations, which may be detrimental at PPI interfaces, but can be mitigated by orthogonal physics-based scoring methods (such as Rosetta). The potential limitations of ranking designs using the AF2 i_pTM metric is that it has become a powerful binary predictor of binding activity but does not correlate with interaction affinity.
Conclusion and Outlook
BindCraft represents a significant advancement towards a “one design, one binder” approach in computational design, with enormous potential in therapeutics, diagnostics, and biotechnology. This study demonstrates the performance of BindCraft on twelve diverse, challenging, and therapeutically relevant protein targets. The affinities of the designed binders are primarily in the nanomolar range, with success rates ranging from 10% to 100%, averaging 46.3%, which is remarkable for a purely computational approach.
These success rates allow for fewer designs to be experimentally screened to identify functional binders, compared to the current state-of-the-art RFdiffusion and the recently described closed-source AlphaProteo binder design pipelines. Notably, the binders designed using this pipeline recently ranked first in a community-wide binder design competition, demonstrating an affinity of 82 nM against the challenging epidermal growth factor receptor (EGFR) target.
Through iterative improvements to the design pipeline, the team envisions ultimately reaching the “one design, one binder” stage, eliminating the need for screening. This will enable a wide range of research groups without protein design expertise to rapidly generate binders for research, biotechnology, and therapeutic applications.
Paper Citation: Pacesa, M., Nickel, L., Schellhaas, C., Schmidt, J., Pyatova, E., Kissling, L., … & Correia, B. E. (2025). One-shot design of functional protein binders with BindCraft. Nature, 625, 598-608. https://doi.org/10.1038/s41586-025-09429-6
