The rapid expansion of the AlphaFold Database has provided unprecedented structural coverage of plant proteomes, thereby creating new opportunities for the computational design of functional proteins tailored for agricultural and biotechnology applications. However, existing deep learning-based antibody design methods encounter substantial computational bottlenecks when scaled for high-throughput screening of plant targets. We present PhytoNB, an automated, parallel-accelerated framework for the de novo design of plant nanobodies. This pipeline integrates domain-level segmentation via Chainsaw, multi-view binding site prediction using GPSite and MVGNN, generative nanobody design with IgGM, and large-scale energy-based filtering through Rosetta. To address the considerable computational demands of large-scale generative models, we developed a parallel acceleration engine incorporating dynamic GPU scheduling and multi-threading optimization. This engine enables efficient task allocation across multiple GPUs and CPU cores. Starting from structural inputs, PhytoNB autonomously performs structure prediction, epitope localization, sequence-structure co-generation, and biophysical validation. The pipeline thereby identifies high-stability nanobody candidates that target key functional regions of plant proteins. Benchmarks demonstrate that the parallelized workflow achieves orders-of-magnitude acceleration in design throughput while maintaining structural fidelity and binding specificity. PhytoNB provides an efficient and scalable platform for plant nanobody discovery, with an interface that supports streamlined application of the design workflow. Extensible applications of this platform include multi-epitope targeting, multi-specific binder design, and cross-species applications.
Wu et al. (Thu,) studied this question.