💫 Binder Design Pipeline — Modular Singularity Containers for Protein Binder Design with RFdifusion

This repository contains a quality-of-life mod for the modern protein engineer.

The tools used in this pipeline — RFdiffusion, ProteinMPNN, and AlphaFold2 — were originally integrated in the dl_binder_design workflow using Conda environments and monolithic setups. While effective, that structure can be difficult to reproduce or adapt in scientific computing environments, where software restrictions and lack of root access are common.

Here, each tool is packaged in a standalone Singularity container, making the pipeline easier to run, modify, and deploy on shared HPC systems without relying on Docker or Conda.

This restructuring makes the workflow:

• Easier to deploy in real-world cluster environments
• More modular and transparent for debugging or retooling
• Better suited for labs working between computation and experiment

RFdiffusion Container (Singularity)

This container runs RFdiffusion using Singularity and is intended for use on HPC clusters with NVIDIA GPUs. It supports protein structure generation using diffusion-based modeling.

Features

• CUDA 11.6.2 and cuDNN 8 support
• Python 3.9 with PyTorch 1.12.1 and DGL (CUDA 11.6)
• Includes SE3Transformer and RFdiffusion source installation
• Modular setup: works with a locally cloned RFdiffusion repo

---

Instalation

Clone the official RFdiffusion repository into the same directory as your Singularity definition file:

git clone https://github.com/RosettaCommons/RFdiffusion.git
singularity build RFdiffusion.sif RFdiffusion.def

RFdiffusion Container (Docker)

While designed for Singularity (ideal for shared HPC environments without root access), the same build is also available as a Docker image for local development or cloud deployment.

docker pull cheems154/my_rfdifusion:9.0

ProteinMPNN Container (Singularity)

This container builds and runs ProteinMPNN using Singularity for use on HPC clusters with NVIDIA GPUs. The container includes a pre-configured Conda environment and can optionally support PyRosetta integration (if licensed credentials are available).

Features

• CUDA 11.6.1 and cuDNN 8 development environment
• Pre-installed Conda environment based on proteinmpnn_fastrelax.yml
• PyRosetta configuration support (optional)
• Designed to work with the dl_binder_design fork or standalone MPNN modules

Installation

Place the cloned dl_binder_design repo in the same directory as your proteinmpnn.def file, then build the container:

The bootstrap.sh script is used to unpack and inject Rosetta binaries (brians_score_jd2 and extract_pdbs) into the correct location within the dl_binder_design repository. This is necessary because, these Rosetta binaries are not included in the public dl_binder_design repo.

git clone https://github.com/nrbennet/dl_binder_design.git
./bootstrap.sh
singularity build proteinmpnn.sif proteinmpnn.def

AlphaFold Initial Guess Container (Singularity)

This container packages AlphaFold2.3.2 Initial Guess, a modified AlphaFold2 protocol designed to improve complex prediction in de novo binder design.

In this context, most failures fall into:

• Type I errors – the designed binder does not fold into the intended monomeric structure.
• Type II errors – the binder folds correctly but fails to bind the target.

Initial Guess specifically addresses Type II errors by seeding AlphaFold’s pair representation with the designed backbone and target structure, biasing the model toward the intended interface. This significantly improves multimer prediction success rates and enables early filtering of designs that are unlikely to bind.

Predictions are scored using pAE_interaction, the average predicted alignment error for interchain residue pairs, with values < 10 strongly correlated to experimentally confirmed binders.

Note: There is an important trade-off. Too little bias increases false negatives, while too much bias inflates false positives. The Institute for Protein Design demonstrated that initializing with an encoding of the binder–target complex achieves a balanced bias, recovering more accurate complexes without overfitting.

This container makes the Initial Guess method fully deployable on Singularity-based HPC systems enabling reproducible large-scale complex prediction and binder prioritization.

Features

• Based on base-2.3.2.sif by @jsgro
• Custom conda environment for binder design (af2_binder_design.yml)
• All required .npz parameter weights are bundled
• Includes silent_tools for Rosetta-based helper scripts

---

Building the container

Download base image

wget -O base-2.3.2.sif https://github.com/jsgro/alphafold_singularity/releases/download/v2.3.2/base-2.3.2.sif

singularity build alphafold_initial_guess.sif alphafold_initial_guess.def