Generalized Biomolecular Modeling and Design with RoseTTAFold All-Atom

Figure 1: Copied from Figure 4A of the preprint. 

Background 

Proteins constitute the building blocks of biological processes. Consequently, knowledge of their native structures is a critical pre-requisite to any form of basic and therapeutic research (1). Experimental approaches to solving protein structures are infamous for being laborious, non-economic, and often beyond the scope of in vitro conditions (I can personally relate to such misfortune, as my PhD project primarily revolved around the challenges of protein isolation and structural characterisation). Conversely, primary amino acid sequences are readily available through widely used databanks, and can be used to estimate a protein’s folded conformation (1, 2). The advent of artificial intelligence-based tools for structurally predicting biomolecules has revolutionised the life sciences over the past decade (2, 3). 

The evolution of the protein modelling field is truly remarkable. In the 1950s, pioneers such as Linus Pauling, G. N. Ramachandran, Alexander Rich, and F. R. Crick used a combination of stereochemical and geometric considerations, model-building, and mathematical equations to assemble secondary structures of alpha-keratin (4) and collagen (5, 6) into 3D-models. However, the analysis of the first protein crystal by Kendrew and colleagues in 1961 (7) unveiled that protein structure prediction can be more complicated than previously thought (8). After half a century of scientists trying to solve the ‘protein folding’ problem, the emergence of platforms like AlphaFold (AF; 9) and RoseTTAfold (RF; 10) radically transformed the protein structural prediction scene. Both approaches are highly accurate in predicting domain structures but exhibit low confidence when it comes to modelling loops, coiled-coil proteins, and disordered regions. The recently modified versions of AF and RF – AlphaFoldMultimer and RoseTTAFold – can predict complex protein assemblies for some canonical amino acids. 

However, they lack the training to accurately predict the coordinates of small molecules, ligands, and cofactors usually associated with a protein complex. 

What has this study achieved? 

In this new multidisciplinary work coordinated by David Baker, a generalised structure prediction algorithm was used to gather a holistic understanding for all atoms within a biological unit, including proteins, nucleic acids, small molecules, cofactors, and chemical modifications. Using a novel prediction tool – RFdiffusion All Atom (RFdiffusionAA) – the authors designed and experimentally validated binding partners of a therapeutic small molecule, an enzymatic cofactor, and optically active photosynthetic molecules. 

Summary of the paper 

Part 1: Development of a RoseTTAFold-based model for prediction of universal biomolecules 

• Generalised Biomolecular Prediction with RoseTTAFold All-Atom (RFAA) 

Using the architectural framework of RoseTTAFold2 (RF2), the authors developed RoseTTAFold All-Atom (RFAA) to include input information about a biomolecular assembly, i.e., amino acids and nucleic acid base sequences, metal ions, small molecule bonded structure, and covalent bonds between proteins and their interactors. 

The three-track system in RF2 was improved as follows: 

1) the 1D track includes 46 new elements representing most common element types found in the Protein Data Bank (PDB), in addition to the 20 residue and 8 nucleic acid base representation in RoseTTAFold Nucleic Acid (RFNA; 10), 

2) the 2D track comprises types of bonds between elements (single, double, triple, etc), 

3) the 3D tracker encodes stereochemistry information (chirality of the molecules). 

Figure 2. Processing of molecular input information on RFAA for generalised biomolecular prediction. Copied from the preprint (Fig 1B). 

• Training RFAA to predict protein-small molecule complexes. 

In AlphaFold (AF), the Frame Aligned Point Error (FAPE) loss function helps minimise the distance between the predicted structure and the actual PDB (protein data bank) structure, thereby enhancing accuracy of a model. The authors incorporated an all-atom version of FAPE in RFAA, where every atom was assigned a local coordinate frame based on neighbouring 

atoms. Similar to AF, the RFAA network also predicts atom and residue-wise confidence (pLDDT) and pairwise confidence (PAE) to help identify high quality predictions. 

Of note, the authors demonstrated three successful examples where they used RFAA to model higher order biomolecular complexes. Comparison to other deep learning-based docking methods showed that RFAA predicts 42% of the complexes correctly, which is significantly higher than what was achieved by other tested methods. 

• Prediction of protein covalent modifications using RFAA 

Posttranslational modifications (PTMs) involve the covalent addition of modifying groups to some proteins after their biosynthesis. The ability to model these covalent modifications can have direct implications for therapeutics and diagnostic designs. RFAA uses a process called ‘atomisation’ to treat proteins as atoms, rather than residues. Consequently, it models modifications in proteins by assigning atomic identity to a particular residue and the attached chemical moiety. 

In this work, 931 recent entries in the PDB were used for the prediction of covalently modified proteins. In 46% of the cases, the network made accurate predictions with modification RMSD <2.5Å. High confidence structures (PAE interaction<10) were obtained for 60% of the predictions, akin to that reported for protein-small molecule complexes. 

• Small molecule binding protein design 

Although there have been several efforts aiming to ‘dock’ molecules into native protein scaffold structures, designing proteins that bind small molecules remains a challenge in the field. Building on recent work (11), the authors in this study refined RFAA by developing a diffusion model, RFdiffusion All-Atom (RFdiffusionAA). This model was conditionally trained on the distribution of proteins pertaining to a biomolecular structure. Information about protein sequences or ‘motifs’ were included in the training language, since motifs play a big role in determining ligand conformation. The authors then evaluated RFdiffusionAA in silico for four independent small molecules, utilising a combination of platforms like LigandMPNN and RosettaGALigandDock. The results show that RFdiffusionAA predictions score better binding energy evaluations than the existing RF diffusion platform that utilises an attractive/repulsive potential. 

Part 2: Experimental validation of designed binders 

Next, the authors designed proteins that bind three diverse small molecules using RFAA, and experimentally validated their binding characteristics to further strengthen the prediction tool. 

1. Previously, binders for Digoxigenin – a small molecule therapeutic steroid for treating cardiovascular diseases – were designed based on co-crystal structures, binding fitness landscapes, and thermodynamic binding parameters (12). Such approaches may not be ideal for small molecules of diverse origins due to challenges in determining experimental characteristics. In this work, the authors employed RFAA to design Digoxigenin-binding backbones without utilising prior knowledge about protein-ligand interface or backbone structure. After fitting sequences into the backbones with the assistance of LigandMPNN and Rosetta FastRelax, over 4000 designs were selected based on AF2 predictions and Rosetta metrices and screened for binding via fluorescence-activated cells sorting. Finally, 3 of these designed proteins, displaying high binding signals, were purified in vitro and characterised. The most potent Digoxigenin binder demonstrated a dissociation constant (Kd) of 10 nM and exhibited remarkable thermostability at temperatures as high as 98 °C. This clearly proves that RFAA can successfully generate novel binders for small molecules in a robust and resourceful manner. 

1. Heme is a porphyrin Iron-binding compound and a critical enzymatic cofactor in the transport and storage of oxygen in vertebrates. Hemoglobin, myoglobin, cytochrome, etc. all comprise the heme auxillary group. Therefore, designing novel high affinity heme binders holds promise for potential therapeutic and diagnostic avenues in the future. Using RFAA, the authors designed heme binders and selected 168 of the designs based on AF predictions and confidence of the backbone design (indicated by RMSD). The critical part in these designs was that, unlike the small molecule Digoxigenin, the catalytic function of heme depends on its binding to a central Iron molecule, which is coordinated by a Cysteine residue above the porphyrin ring. Finally, after purification from the bacterial heterologous system, 38 of the designs were obtained as monomeric, thermostable, heme-binding structures, indicating therapeutic and industrial applications. 
2. Employing RFAA-based characterisation, the researchers could also build optically enhanced binders for bilins, essential light-harvesting molecules in photosynthetic organisms. Of particular interest, 3 designs reported satisfactory fluorescence quantum yields, demonstrating potential in building novel complexes with increased light capturing capacity and subsequently higher photosynthetic output. 

What I loved about this work 

• What truly stood out for me in this study was the utilisation of in vitro tools to validate in silico predictions performed with RFAA and RFdiffusionAA. It satisfied my scientific curiosity to see orthogonal evidence that further strengthened the presented model and offered tangible proof to appreciate the credibility of the predictions. 
• The idea of “atomisation” stood out particularly because of the elegance of the concept where protein residues are simply treated as atoms. This model assigns atomic identity not only to the constituent residues of a protein, but also extends it to diverse small molecules closely associated with an individual protein or a protein complex. By reducing macromolecules to their elemental units, this approach encapsulates the essence of all matter with a remarkable level of simplicity. 
• Another interesting attribute of the study is that the authors tested RFAA’s efficiency and “trainability” based on its ability to predict protein-small molecule interactions outside of their training dataset. By extending the tool’s memory to encompass a diverse array of proteins and small molecules beyond those included in the training set, this attribute not only enhances the tool’s adaptability but also underscores the robust and comprehensive approach taken in its development. 

Questions to the authors. 

• Is there a maximum limit to the number of atoms or residues that can be accurately predicted using this training network? 
• Given the promising outcomes observed in three experimental cases in this study, can RFAA-mediated modelling potentially contribute to understanding dynamic transient interactions crucial for accurate cellular functioning? For example, in cell cycle pathways, small molecules external to critical protein complexes dynamically interact with enzymes and functional domains within short time frames. Often, it is challenging to capture such interactive complexes in vitro. I was wondering whether RFAA has been tested to understand such transiently/ reversibly stable complexes. 

References: 

1. Lupas, A. N., Pereira, J., Alva, V., Merino, F., Coles, M., & Hartmann, M. D. (2021). The breakthrough in protein structure prediction. Biochemical Journal, 478(10), 1885–1890. https://doi.org/10.1042/BCJ20200963 
2. Kuhlman, B., & Bradley, P. (2019). Advances in protein structure prediction and design. Nature Reviews Molecular Cell Biology, 20(11), 681–697. https://doi.org/10.1038/s41580-019-0163-x 
3. Hekkelman, M. L., De Vries, I., Joosten, R. P., & Perrakis, A. (2023). AlphaFill: Enriching AlphaFold models with ligands and cofactors. Nature Methods, 20(2), 205–213. https://doi.org/10.1038/s41592-022-01685-y 
4. Pauling, L., & Corey, R. B. (1953). Compound Helical Configurations of Polypeptide Chains: Structure of Proteins of the α-Keratin Type. Nature, 171(4341), 59–61. https://doi.org/10.1038/171059a0 
5. Ramachandran, G. N., & Kartha, G. (1955). Structure of Collagen. Nature, 176(4482), 593–595. https://doi.org/10.1038/176593a0 
6. Rich, A., & Crick, F. H. C. (1955). The Structure of Collagen. Nature, 176(4489), 915–916. https://doi.org/10.1038/176915a0 
7. Kendrew, J. C., Bodo, G., Dintzis, H. M., Parrish, R. G., Wyckoff, H., & Phillips, D. C. (1958). A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis. Nature, 181(4610), 662–666. https://doi.org/10.1038/181662a0 
8. Lupas, A. N., Pereira, J., Alva, V., Merino, F., Coles, M., & Hartmann, M. D. (2021). The breakthrough in protein structure prediction. Biochemical Journal, 478(10), 1885–1890. https://doi.org/10.1042/BCJ20200963 
9. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S. A. A., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., … Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2 
10. Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. N., Schaeffer, R. D., Millán, C., Park, H., Adams, C., Glassman, C. R., DeGiovanni, A., Pereira, J. H., Rodrigues, A. V., Van Dijk, A. A., Ebrecht, A. C., … Baker, D. (2021). Accurate prediction of protein structures and interactions using a three-track neural network. Science, 373(6557), 871–876. https://doi.org/10.1126/science.abj8754 
11. Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., Wicky, B. I. M., Hanikel, N., Pellock, S. J., Courbet, A., Sheffler, W., Wang, J., Venkatesh, P., Sappington, I., Torres, S. V., … Baker, D. (2022). Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models [Preprint]. Biochemistry. https://doi.org/10.1101/2022.12.09.519842 
12. Tinberg, C. E., Khare, S. D., Dou, J., Doyle, L., Nelson, J. W., Schena, A., Jankowski, W., Kalodimos, C. G., Johnsson, K., Stoddard, B. L., & Baker, D. (2013). Computational design of ligand-binding proteins with high affinity and selectivity. Nature, 501(7466), 212–216. https://doi.org/10.1038/nature12443 

doi: https://doi.org/10.1242/prelights.36373

(No Ratings Yet)