Top-down design of protein nanomaterials with reinforcement learning

Abstract

The multisubunit protein assemblies that play critical roles in biology are the result of evolutionary selection for function of the entire assembly, and hence the subunits in structures such as icosahedral viral capsids often fit together with remarkable shape complementarity1,2. In contrast, the large multisubunit assemblies that have been created by de novo protein design, notably the icosahedral nanocages used in a new generation of potent vaccines3–7, have been built by first designing symmetric oligomers with cyclic symmetry and then assembling these into nanocages while keeping the internal structure fixed8–14, which results in more porous structures with less extensive shape matching between the components. Such hierarchical “bottom-up” design approaches have the advantage that one interface can be designed and validated in the context of the cyclic oligomer building block15,16, but the disadvantage that the structural and functional features of the assemblies are limited by the properties of the predesigned building blocks. To overcome this limitation, we set out to develop a “top-down” reinforcement learning based approach to protein nanomaterial design in which both the structures of the subunits and the interactions between them are built up coordinately in the context of the entire assembly. We developed a Monte Carlo tree search (MCTS) method17,18 which assembles protein monomer structures in the context of an overall architecture guided by a loss function which enables specification of any desired overall structural properties such as shape and porosity. We demonstrate the power of the approach by designing hyperstable icosahedral assemblies more compact than any previously observed protein icosahedral structure (designed or naturally occurring), that have very low porosity and are robust to fusion and display of proteins as complex as influenza hemagglutinin. CryoEM structures of two designs are very close to the computational design models. Our top-down reinforcement learning approach should enable the design of a wide variety of complex protein nanomaterials by direct optimization of overall system properties.

Fast and versatile sequence-independent protein docking for nanomaterials design using RPXDock

“…Docking monomeric and oligomeric building blocks into higher-order symmetric complexes followed by protein-protein interface design is an established and successful paradigm for accurately creating novel self-assembling protein nanomaterials (King et al 2014(King et al , 2012Fallas et al 2017;Shen et al 2018;Ben-Sasson et al 2021;Bale et al 2016;Gonen et al 2015). While deep learning-based generative models have recently proven successful in designing de novo oligomers (Wicky et al 2022) and small nanocages (Lutz et al 2022), the ability of RPXdock to use experimentally verified or previously designed scaffolds in a stepwise manner enables the use of specific building blocks that have optimal features for specific applications (Marcandalli et al 2019;Boyoglu-Barnum et al 2021;Ueda et al 2020;Walls et al 2020;Brouwer et al 2019). The RPXdock code can accommodate specific user requirements for complex docking problems, and the efficiency at which 34 high-quality docks are found has been greatly improved compared to its predecessors (tcdock; sicdock; sicaxel (King et al 2014;Fallas et al 2017;Brouwer et al 2019;Marcandalli et al 2019)) due to the hierarchical search and scoring algorithms.…”

De Novo Design of Polyhedral Protein Assemblies: Before and After the AI Revolution

“…of the nanomaterial, and the model systematically builds up subunits which pack together to optimise these parameters. [67] c) RFdiffusion, a diffusion generative model based on RoseTTAFold, generates polyhedral protein assemblies by conditioning the neural network. [68] d) Chroma, a correlated diffusion process with chain and radius of gyration constraints, gradually generates symmetric protein backbones from random collapsed polymers, while protein sequences and side-chain conformations suitable to fold into the generated backbone are generated by a graph-based design network.…”

“…Recently, a "topdown" Monte Carlo Tree Search (MCTS)-based approach to de novo protein design (Figure 3b) has been described. [67] This method involves the coordinated construction of both the ASU 3D structure and the protein-protein interfaces between adjacent copies within the context of the entire assembly. Integrating the principles of decision-making and reinforcement learning, MCTS uses an algorithm for searching combinatorial spaces represented by trees.…”

“…[106][107][108] Such an approach has been used, for example, to target the angiopoietin pathway, through clustering of the cellular surface receptors, demonstrating the potential of these materials as multivalent agonists. [67] Alternatively, protein assemblies can be enzymatically functionalized by conjugation of proteins carrying polyglycine tags to scaffolds containing sort-tags using sortase. [123] This approach has been used to study the effect of spatially organising multiple cellulosedegrading enzymes on their activity, showing activity enhancement upon co-localization at the assembly surface, a promising route towards higher efficiency enzyme materials.…”

“…While structure prediction algorithms can be used as a validation step prior to experimental testing (Figure 5a and 5b), typical experimental success rates for the design of polyhedral protein assemblies, where success is defined as the production of proteins that selfassemble into the designed architecture, lay around 10 %, [136] with even lower success rates observed for fully AI-based approaches. [67,68] One way to improve individual pipelines is to combine different methods and subsequent design rounds, which can diversify and optimise generated backbones and sequences towards user defined parameters (e. g. Figure 5c). Another way is by using high-throughput methods to screen and validate a large number of computationally designed polyhedral assemblies to generate rich data sets.…”