BindCraft Enables Direct AlphaFold2‑Driven Intelligent Protein Binder Design (46% Success on 12 Targets)

Protein function in living systems largely depends on protein‑protein interactions (PPIs), making the design of specific binders a promising therapeutic and biotechnological strategy.

Traditional binder generation methods such as immunization, antibody‑library screening, or directed evolution are labor‑intensive and provide limited control over binding sites. Early computational approaches like Rosetta attempted physics‑based modeling and side‑chain optimization but reported success rates below 0.1%.

The advent of deep‑learning structure predictors, especially AlphaFold2, dramatically improved the accuracy of single‑protein folding and complex modeling, yet true "intelligent" design remained elusive. Existing generative tools (e.g., RFdiffusion, ProteinMPNN) still require manually defined scaffolds and docking interfaces before AlphaFold validation.

To address this gap, researchers from EPFL and MIT introduced BindCraft, an open‑source automated workflow that back‑propagates error gradients through AlphaFold2’s multi‑body weights to directly generate binder sequences. By integrating a neural network with AlphaFold2, BindCraft jointly optimizes protein structure, sequence, and interface, eliminating the need for high‑throughput screening, experimental iteration, or prior knowledge of binding sites, and can produce nanomolar‑affinity de novo binders in silico.

Experimental evaluation on twelve structurally complex, pharmacologically relevant targets showed success rates ranging from 10% to 100%, with an average success rate of 46.3%. This indicates that designs previously requiring hundreds or thousands of experimental screens can now be obtained with a single computational run.

BindCraft has been released as a tutorial on the HyperAI website, allowing users to launch the workflow with a single click. The tutorial guides users through the following steps:

Navigate to the HyperAI homepage, select the "BindCraft: Protein Binder Design" tutorial, and click "Run this tutorial online".

Clone the tutorial repository to a personal container.

Choose the "NVIDIA GeForce RTX 5090" GPU and a PyTorch image, then start the execution (options include pay‑as‑you‑go or subscription billing).

Wait for resource allocation (approximately two minutes), then open the workspace once the status changes to "Running".

Open the README within the workspace, configure parameters as described, and view the generated binder designs and visualizations.

Images illustrating the demo interface and result visualizations are included in the original tutorial.