Leveraging artificial intelligence and deep learning to generate proteins de novo has unlocked new frontiers of protein design. By training deep learning models on protein sequences and experimental structures, we can sample new structural landscapes unexplored by evolution. This approach can be used to design bespoke binders that target specific proteins and domains. However, generating successful binders has a low in silico success rate, often requiring thousands of designs and hundreds of GPU hours to obtain enough hits for experimental testing. While commercial web-apps and workstation implementations are available for binder design, these are limited in both scalability and throughput. There is a lack of efficient open-source pipelines designed for high-performance computing (HPC) systems that can maximise hardware resources and parallelise the workflow efficiently.
Here, we present 'ProteinDJ'1 - an implementation of a protein binder design workflow that is deployable on HPC systems using the Nextflow pipeline language and Singularity containerisation. It automatically batches and parallelises the workload across both GPUs and CPUs, facilitating the generation and testing of hundreds of designs per hour, dramatically accelerating the discovery process. Importantly, it allows central deployment of the workflow on shared HPC resources, rather than workstations that are restricted to individual research groups.
ProteinDJ is modular by design and currently includes RFdiffusion or BindCraft for fold generation, ProteinMPNN or Full-Atom MPNN for sequence design, and AlphaFold2 or Boltz-2 for prediction and validation of binder-target interfaces, with supporting packages for structural evaluation of designs. We have included structure-based filters to reject problematic designs mid-stream to minimise manual inspection of designs by the user. As the most effective parameters for binder design can vary between target proteins, we also developed 'BindSweeper' – a Python-based command-line tool that can automatically coordinate multiple ProteinDJ runs to systematically explore design parameters. BindSweeper can be used to 'sweep' multiple parameters in ProteinDJ, such as target hotspot residues or structure prediction methods. Our approach democratises protein binder design in an easy-to-use and robust implementation freely available on GitHub (https://github.com/PapenfussLab/proteindj).