Zero-shot design of drug-binding proteins via neural selection-expansion 1. This study introduces a powerful method for the zero-shot design of small-molecule binding proteins using neural networks. The approach combines two reciprocal neural networks, LASErMPNN and RoseTTAFold-All Atom (RFAA), in an iterative design process called NISE, achieving high-affinity small-molecule binders from scratch. 2. The core of the method involves a self-consistency optimization algorithm that iterates between sequence design and co-structure prediction. LASErMPNN designs protein sequences based on protein-ligand co-structure, while RFAA predicts protein-ligand co-structures from the designed sequences, optimizing both protein and ligand coordinates. 3. The method was tested by designing a protein binder for the anticancer drug exatecan, a topoisomerase I inhibitor. The designs showed promising results, with the best-performing design binding exatecan with a dissociation constant (Kd) as low as 120 nM, surpassing state-of-the-art methods by up to 70-fold in terms of affinity. 4. A major innovation of this work is the introduction of "neural proofreading," where LASErMPNN suggests mutations to improve binding affinity. By adding just two predicted mutations, the affinity was improved by over 100-fold, achieving a Kd of 1.2 nM for the exatecan-binding protein. 5. The designed proteins were also tested for their ability to protect exatecan from hydrolysis, a common issue for the drug in the bloodstream. The best design, EPIC, protected exatecan from hydrolysis for days, a dramatic improvement compared to the drug’s typical half-life of a few hours. 6. This approach provides a general framework for designing small-molecule binding proteins, with potential applications in drug delivery, antidotes, and sensors. It demonstrates that deep learning models can be used to automate and accelerate the design of highly specific protein binders for a wide range of molecules. 7. The study also highlights the challenges in protein-ligand co-structure prediction, particularly in accurately modeling ligand functional groups, but the results suggest that future improvements in structure prediction will further enhance the method's performance. @nickpolizzi_ @benf549 @PolizziLab 📜Paper: biorxiv.org/content/10.1101/… #ProteinDesign #DeepLearning #DrugDiscovery #Bioinformatics #MachineLearning #ComputationalBiology #AIinBiology #NeuralNetworks #SmallMoleculeBinders