Mobile Menu

Creating Life’s Building Blocks: Models for De Novo Protein Design

A key area for AI models is in the field of protein design, where new proteins are created with specific structures and functions that are not found in nature. In particular, designed proteins offer the ability to create novel functionality, whilst being engineered so that their activity can be precisely controlled to create the desired output. This is especially helpful in biomedicine, where de novo protein design has already been applied in preventing viral infections, modulating the immune system and sense-and-respond signalling systems1. However, designing a protein from scratch is no easy feat, with de novo design being experimentally challenging due to the complex nature of protein tertiary structures.

Luckily, AI models are uniquely positioned to handle this task. Deep learning methods have recently been shown to create protein structures that have the precise geometry needed for their desired function. In this feature, we explore the major tools used for de novo protein design and commercial options for AI-based novel protein creation.

The long road ahead: four key hurdles in protein design

Generally speaking, the process of designing novel proteins can be split into four major steps; the protein backbone conformation design, sequence sampling, scoring and functional site design (see Figure 1).

  1.  Backbone sampling: Protein backbones are composed of a sequence of dihedral angles (phi, psi and omega) for each residue. The conformational space for these angles is vast, leading to an extremely high-dimensional search space that makes it difficult to identify optimal structures. Additionally, proteins can adopt a multitude of conformations. Sampling must account for this flexibility to ensure that designed backbones are realistic and can stably fold into the desired structure.
    To conduct backbone sampling, researchers may choose to use a variation of an existing structure (e.g., redesigning native backbones from the PDB), assemble local structures, use fold family design or newer machine learning methods1.
  2. Sequence optimisation: The next step involves the addition of amino acid side-chain types and conformations into the design. Amino acid interactions are essential for creating functional three-dimensional structures and backbone stabilisation. For a protein of average length, the number of possible amino acid sequences for a given protein backbone (with 20 possible amino acids at each position) is astronomical, making it computationally intensive to explore all potential sequences. Therefore, methods such as HBNet and trRosetta are invaluable for sequence optimisation.
  3. Scoring: How do you rank your designs to select one with the desired properties? In the next stage of the design pipeline, scoring functions are used to evaluate and rank protein sequences based on their predicted stability, structure and functionality. These functions estimate the free energy of a protein structure and help identify sequences that are likely to fold into a desired conformation.
  4. Functional site design: Lastly, the designed protein must have a functional site complementary to the desired function. Importantly, researchers must determine which amino acid residues are crucial for the desired activity, which involves identifying the catalytic residues for an enzyme or the binding residues for a ligand or interaction partner. There are many methods to assist in this process, depending on the type of functional site is required – catalytic sites, ligand-binding sites and membrane interaction sites are common examples.

Figure 1: Major aspects of the de novo protein design. The design of a functional de novo protein, for example, a binder (middle, magenta) to a target protein (middle, gray), requires sampling of the backbone structure space to find a backbone compatible with the function, sequence optimization to stabilize the backbone, and designing the functional site interactions. A scoring function is necessary to select designs with desired properties, typically by identifying low-energy sequence–structure combinations. Sourced from Pan et al., 20211.

ProteinMPNN

It’s clear that deep learning models have massively improved protein structure prediction, but until recently, experimentally verified de novo protein design still relied on physical methods such as Rosetta2 (see Figure 2 for an overview of common design approaches). However, this changed with the 2022 release of ProteinMPNN3, a deep learning method for protein sequence design based on a message-passing neural network (MPNN). The model works by autoregressively predicting protein sequences when provided a set of protein backbone coordinates as input. Dauparas et al. illustrated that ProteinMPNN achieves a sequence recovery rate of 52.4% on native protein backbones, significantly higher than Rosetta’s 32.9%. This improvement in sequence prediction is done by using backbone features and interatomic distances, enabling it to accurately design single or multiple chains for a variety of protein design challenges.

Figure 2: Protein design concepts and approaches. (A) Most design methods divide the process into two steps: first, a structure containing only the polypeptide backbone is generated, and then a sequence is designed for that backbone. For each step, design methods that use atomistic modeling (blue) or AI-based approaches (orange) are indicated. (B) Classical design methods use a “blueprint” defining a protein fold topology and then assembles a three-dimensional backbone from ideal helix, strand, and loop peptide fragments. (C) Backbone generation methods can systematically sample geometries within a given fold. These methods generate synthetic fold families that, similar to evolved protein families, can be optimized for diverse functions. (D) A recent AI-based method, protein diffusion, generates protein backbones through a denoising process from random backbone starting coordinates. This method generates diverse protein folds without having to pre-specify a topology as input. Sourced from Kortemme, 20244.

Frame2seq

ProteinMPNN is not the only deep learning tool that can generate de novo protein designs using protein backbone inputs. Recently, a new model called Frame2seq was published as a preprint, detailing a novel structure-conditioned masked language model for protein sequence design5. The authors demonstrate that the model accurately designs soluble, monomeric, and stable proteins, establishing it as a state-of-the-art fixed-backbone design method to rival ProteinMPNN. To illustrate this, Akpinaroglu et al. benchmarked their model against a ProteinMPNN model trained on the CATH 4.2 dataset, showing that Frame2Seq outperforms ProteinMPNN by 2% in native sequencer recovery, with over six times faster inference.

RFdiffusion

From the pioneers that developed RoseTTAFold, last year saw the publication of a novel protein design tool that has already had a significant impact on the proteomics community4. The tool, RFdiffusion6, uses a protein diffusion model – a newer AI approach that uses a machine learning algorithm similar to those used for image generation. The diffusion model is distinct from other design strategies in that its input is amino acid residue noise, which then undergoes a “denoising” process (see Figure 1). The RFdiffusion model was developed by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks. The authors report that this approach typically results in high quality structures, outperforming other preexisting design methods in areas including protein binder design, enzyme active site scaffolding and metal-binding protein design7.

As previously mentioned, an advanced version of RFdiffusion, called RFdiffusion All-Atom, was released this year. Designed specifically to handle small-molecule binder design, the model generates folded proteins surrounding a small molecule of interest, with unique binding pockets being generated for each ligand. As a proof of concept, the authors also used the model to generate small-molecule binding designs for the heart disease drug digoxigenin and enzyme cofactor heme, which were then experimentally validated to show that some designed had the desired binding activity.

Overall, RFDiffusion All-Atom is a versatile generative model capable of creating novel proteins, without the need for native or preexisting designed scaffolds. Importantly, the code is available for free on Github or you can access it through Tamarind Bio.

Protein design without the hassle?

For those less computationally inclined, or eager for faster results, there’s now a multitude of companies using AI-based models for de novo protein design and protein optimisation (see Table 1). Some companies, such as Arzeda and AI Proteins, provide a good option for outsourcing the entire production workflow – in-house design, synthesis and optimisation included. Alternatively, companies providing access to proprietary AI design software are also a popular choice for many research, industry and pharmaceutical organisations, allowing users to freedom to generate protein sequences themselves.

Table 1: Selected protein design and optimisation companies.
CompanyDesign focusProduct
ArzedaAntibodies, enzymes, proteinsIntelligent Protein Design Technology: protein design, DNA synthesis and experimental validation
Cradle BioEnzymes (vaccines, peptides and antibodies currently in beta)Proprietary design software
ProteinQurePeptidesProteinStudio: an integrated platform for drug peptide design
AI ProteinsMiniproteinsProtein design, characterisation and optimisation pipeline
ExazymeEnzymes, antibodies, peptidesProprietary design web app
Scala BiodesignEnzymes, antibodies, vaccine immunogens, proteinsProprietary design software
Cyrus BiotechnologyAntibodies, vaccine immunogens, proteinsProprietary design software
DeepChainProteinsProprietary design web app

Want to learn more? Keep an eye out for our Proteomics Playbook, releasing late 2024.

References

  1. Pan, X. & Kortemme, T. Recent advances in de novo protein design: Principles, methods, and applications. J. Biol. Chem. 296, 100558 (2021).

2.    Rohl, C. A., Strauss, C. E. M., Misura, K. M. S. & Baker, D. Protein structure prediction using Rosetta. Methods in Enzymology 383, 66–93 (2004).

3.    Dauparas, J. et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022).

4.    Kortemme, T. De novo protein design-From new structures to programmable functions. Cell 187, 526–544 (2024).

5.    Akpinaroglu, D. et al. Structure-conditioned masked language models for protein sequence design generalize beyond the native sequence space. bioRxiv 2023.12.15.571823 (2023) doi:10.1101/2023.12.15.571823.

6.    Watson, J. L. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023).

7. Baker Lab – RFdiffusion: A generative model for protein design. Baker Lab – UW Institute for Protein Design https://www.bakerlab.org/2023/07/11/diffusion-model-for-protein-design/ (2023).


More on these topics

AI / Machine Learning / Proteomics