Coupling Protein Side-Chain and Backbone Flexibility Improves the Re-design of Protein-Ligand Specificity

Interactions between small molecules and proteins play critical roles in regulating and facilitating diverse biological functions, yet our ability to accurately re-engineer the specificity of these interactions using computational approaches has been limited. One main difficulty, in addition to inaccuracies in energy functions, is the exquisite sensitivity of protein–ligand interactions to subtle conformational changes, coupled with the computational problem of sampling the large conformational search space of degrees of freedom of ligands, amino acid side chains, and the protein backbone. Here, we describe two benchmarks for evaluating the accuracy of computational approaches for re-engineering protein-ligand interactions: (i) prediction of enzyme specificity altering mutations and (ii) prediction of sequence tolerance in ligand binding sites. After finding that current state-of-the-art “fixed backbone” design methods perform poorly on these tests, we develop a new “coupled moves” design method in the program Rosetta that couples changes to protein sequence with alterations in both protein side-chain and protein backbone conformations, and allows for changes in ligand rigid-body and torsion degrees of freedom. We show significantly increased accuracy in both predicting ligand specificity altering mutations and binding site sequences. These methodological improvements should be useful for many applications of protein – ligand design. The approach also provides insights into the role of subtle conformational adjustments that enable functional changes not only in engineering applications but also in natural protein evolution.     

Predicting drug polypharmacology from cell morphology readouts using variational autoencoder latent space arithmetic

A variational autoencoder (VAE) is a machine learning algorithm, useful for generating a compressed and interpretable latent space. These representations have been generated from various biomedical data types and can be used to produce realistic-looking simulated data. However, standard vanilla VAEs suffer from entangled and uninformative latent spaces, which can be mitigated using other types of VAEs such as β-VAE and MMD-VAE. In this project, we evaluated the ability of VAEs to learn cell morphology characteristics derived from cell images. We trained and evaluated these three VAE variants—Vanilla VAE, β-VAE, and MMD-VAE—on cell morphology readouts and explored the generative capacity of each model to predict compound polypharmacology (the interactions of a drug with more than one target) using an approach called latent space arithmetic (LSA). To test the generalizability of the strategy, we also trained these VAEs using gene expression data of the same compound perturbations and found that gene expression provides complementary information. We found that the β-VAE and MMD-VAE disentangle morphology signals and reveal a more interpretable latent space. We reliably simulated morphology and gene expression readouts from certain compounds thereby predicting cell states perturbed with compounds of known polypharmacology. Inferring cell state for specific drug mechanisms could aid researchers in developing and identifying targeted therapeutics and categorizing off-target effects in the future.     

Side-chain and backbone flexibility in protein core design.

We have developed a computational approach for the design and prediction of hydrophobic cores that includes explicit backbone flexibility. The program consists of a two-stage combination of a genetic algorithm and monte carlo sampling using a torsional model of the protein. Backbone structures are evaluated either by a canonical force-field or a constraining potential that emphasizes the preservation of local geometry. The utility of the method for protein design and engineering is explored by designing three novel hydrophobic core variants of the protein 434 cro. We use the new method to evaluate these and previously designed 434 cro variants, as well as a series of phage T4 lysozyme variants. In order to properly evaluate the influence of backbone flexibility, we have also analyzed the effects of varying amounts of side-chain flexibility on the performance of fixed backbone methods. Comparison of results using a fixed versus flexible backbone reveals that, surprisingly, the two methods are almost equivalent in their abilities to predict relative experimental stabilities, but only when full side-chain flexibility is allowed. The prediction of core side-chain structure can vary dramatically between methods. In some, but not all, cases the flexible backbone method is a better predictor of structure. The development of a flexible backbone approach to core design is particularly important for attempts at de novo protein design, where there is no prior knowledge of a precise backbone structure.     

Point-estimating observer models for latent cause detection

The spatial distribution of visual items allows us to infer the presence of latent causes in the world. For instance, a spatial cluster of ants allows us to infer the presence of a common food source. However, optimal inference requires the integration of a computationally intractable number of world states in real world situations. For example, optimal inference about whether a common cause exists based on N spatially distributed visual items requires marginalizing over both the location of the latent cause and 2^N possible affiliation patterns (where each item may be affiliated or non-affiliated with the latent cause). How might the brain approximate this inference? We show that subject behaviour deviates qualitatively from Bayes-optimal, in particular showing an unexpected positive effect of N (the number of visual items) on the false-alarm rate. We propose several “point-estimating” observer models that fit subject behaviour better than the Bayesian model. They each avoid a costly computational marginalization over at least one of the variables of the generative model by “committing” to a point estimate of at least one of the two generative model variables. These findings suggest that the brain may implement partially committal variants of Bayesian models when detecting latent causes based on complex real world data.     

Neural network‐derived Potts models for structure‐based protein design using backbone atomic coordinates and tertiary motifs

Designing novel proteins to perform desired functions, such as binding or catalysis, is a major goal in synthetic biology. A variety of computational approaches can aid in this task. An energy‐based framework rooted in the sequence‐structure statistics of tertiary motifs (TERMs) can be used for sequence design on predefined backbones. Neural network models that use backbone coordinate‐derived features provide another way to design new proteins. In this work, we combine the two methods to make neural structure‐based models more suitable for protein design. Specifically, we supplement backbone‐coordinate features with TERM‐derived data, as inputs, and we generate energy functions as outputs. We present two architectures that generate Potts models over the sequence space: TERMinator, which uses both TERM‐based and coordinate‐based information, and COORDinator, which uses only coordinate‐based information. Using these two models, we demonstrate that TERMs can be utilized to improve native sequence recovery performance of neural models. Furthermore, we demonstrate that sequences designed by TERMinator are predicted to fold to their target structures by AlphaFold. Finally, we show that both TERMinator and COORDinator learn notions of energetics, and these methods can be fine‐tuned on experimental data to improve predictions. Our results suggest that using TERM‐based and coordinate‐based features together may be beneficial for protein design and that structure‐based neural models that produce Potts energy tables have utility for flexible applications in protein science.     

High-resolution structural and thermodynamic analysis of extreme stabilization of human procarboxypeptidase by computational protein design

Recent efforts to design de novo or redesign the sequence and structure of proteins using computational techniques have met with significant success. Most, if not all, of these computational methodologies attempt to model atomic-level interactions, and hence high-resolution structural characterization of the designed proteins is critical for evaluating the atomic-level accuracy of the underlying design force-fields. We previously used our computational protein design protocol RosettaDesign to completely redesign the sequence of the activation domain of human procarboxypeptidase A2. With 68% of the wild-type sequence changed, the designed protein, AYEdesign, is over 10 kcal/mol more stable than the wild-type protein. Here, we describe the high-resolution crystal structure and solution NMR structure of AYEdesign, which show that the experimentally determined backbone and side-chains conformations are effectively superimposable with the computational model at atomic resolution. To isolate the origins of the remarkable stabilization, we have designed and characterized a new series of procarboxypeptidase mutants that gain significant thermodynamic stability with a minimal number of mutations; one mutant gains more than 5 kcal/mol of stability over the wild-type protein with only four amino acid changes. We explore the relationship between force-field smoothing and conformational sampling by comparing the experimentally determined free energies of the overall design and these focused subsets of mutations to those predicted using modified force-fields, and both fixed and flexible backbone sampling protocols.     

Backrub-like backbone simulation recapitulates natural protein conformational variability and improves mutant side-chain prediction

Incorporation of effective backbone sampling into protein simulation and design is an important step in increasing the accuracy of computational protein modeling. Recent analysis of high-resolution crystal structures has suggested a new model, termed backrub, to describe localized, hinge-like alternative backbone and side-chain conformations observed in the crystal lattice. The model involves internal backbone rotations about axes between C-alpha atoms. Based on this observation, we have implemented a backrub-inspired sampling method in the Rosetta structure prediction and design program. We evaluate this model of backbone flexibility using three different tests. First, we show that Rosetta backrub simulations recapitulate the correlation between backbone and side-chain conformations in the high-resolution crystal structures upon which the model was based. As a second test of backrub sampling, we show that backbone flexibility improves the accuracy of predicting point-mutant side-chain conformations over fixed backbone rotameric sampling alone. Finally, we show that backrub sampling of triosephosphate isomerase loop 6 can capture the millisecond/microsecond oscillation between the open and closed states observed in solution. Our results suggest that backrub sampling captures a sizable fraction of localized conformational changes that occur in natural proteins. Application of this simple model of backbone motions may significantly improve both protein design and atomistic simulations of localized protein flexibility.     

Motif‐directed redesign of enzyme specificity

Computational protein design relies on several approximations, including the use of fixed backbones and rotamers, to reduce protein design to a computationally tractable problem. However, allowing backbone and off‐rotamer flexibility leads to more accurate designs and greater conformational diversity. Exhaustive sampling of this additional conformational space is challenging, and often impossible. Here, we report a computational method that utilizes a preselected library of native interactions to direct backbone flexibility to accommodate placement of these functional contacts. Using these native interaction modules, termed motifs, improves the likelihood that the interaction can be realized, provided that suitable backbone perturbations can be identified. Furthermore, it allows a directed search of the conformational space, reducing the sampling needed to find low energy conformations. We implemented the motif‐based design algorithm in Rosetta, and tested the efficacy of this method by redesigning the substrate specificity of methionine aminopeptidase. In summary, native enzymes have evolved to catalyze a wide range of chemical reactions with extraordinary specificity. Computational enzyme design seeks to generate novel chemical activities by altering the target substrates of these existing enzymes. We have implemented a novel approach to redesign the specificity of an enzyme and demonstrated its effectiveness on a model system.     

Improved energy bound accuracy enhances the efficiency of continuous protein design

Flexibility and dynamics are important for protein function and a protein's ability to accommodate amino acid substitutions. However, when computational protein design algorithms search over protein structures, the allowed flexibility is often reduced to a relatively small set of discrete side‐chain and backbone conformations. While simplifications in scoring functions and protein flexibility are currently necessary to computationally search the vast protein sequence and conformational space, a rigid representation of a protein causes the search to become brittle and miss low‐energy structures. Continuous rotamers more closely represent the allowed movement of a side chain within its torsional well and have been successfully incorporated into the protein design framework to design biomedically relevant protein systems. The use of continuous rotamers in protein design enables algorithms to search a larger conformational space than previously possible, but adds additional complexity to the design search. To design large, complex systems with continuous rotamers, new algorithms are needed to increase the efficiency of the search. We present two methods, PartCR and HOT, that greatly increase the speed and efficiency of protein design with continuous rotamers. These methods specifically target the large errors in energetic terms that are used to bound pairwise energies during the design search. By tightening the energy bounds, additional pruning of the conformation space can be achieved, and the number of conformations that must be enumerated to find the global minimum energy conformation is greatly reduced. Proteins 2015; 83:1151–1164. © 2015 Wiley Periodicals, Inc.     

Thoroughly sampling sequence space: large-scale protein design of structural ensembles

Modeling the inherent flexibility of the protein backbone as part of computational protein design is necessary to capture the behavior of real proteins and is a prerequisite for the accurate exploration of protein sequence space. We present the results of a broad exploration of sequence space, with backbone flexibility, through a novel approach: large-scale protein design to structural ensembles. A distributed computing architecture has allowed us to generate hundreds of thousands of diverse sequences for a set of 253 naturally occurring proteins, allowing exciting insights into the nature of protein sequence space. Designing to a structural ensemble produces a much greater diversity of sequences than previous studies have reported, and homology searches using profiles derived from the designed sequences against the Protein Data Bank show that the relevance and quality of the sequences is not diminished. The designed sequences have greater overall diversity than corresponding natural sequence alignments, and no direct correlations are seen between the diversity of natural sequence alignments and the diversity of the corresponding designed sequences. For structures in the same fold, the sequence entropies of the designed sequences cluster together tightly. This tight clustering of sequence entropies within a fold and the separation of sequence entropy distributions for different folds suggest that the diversity of designed sequences is primarily determined by a structure's overall fold, and that the designability principle postulated from studies of simple models holds in real proteins. This has important implications for experimental protein design and engineering, as well as providing insight into protein evolution.     

Deep generative models for 3D molecular structure

Deep generative models have gained recent popularity for chemical design. Many of these models have historically operated in 2D space; however, more recently explicit 3D molecular generative models have become of interest, which are the topic of this article. Dozens of published models have been developed in the last few years to generate molecules directly in 3D, outputting both the atom types and coordinates, either in one-shot or adding atoms or fragments step-by-step. These 3D generative models can also be guided by structural information such as a binding pocket representation to successfully generate molecules with docking score ranges similar to known actives, but still showing lower computational efficiency and generation throughput than 1D/2D generative models and sometimes producing unrealistic conformations. We advocate for a unified benchmark of metrics to evaluate generation and propose perspectives to be addressed in next implementations.     

Recapitulation of protein family divergence using flexible backbone protein design

We use flexible backbone protein design to explore the sequence and structure neighborhoods of naturally occurring proteins. The method samples sequence and structure space in the vicinity of a known sequence and structure by alternately optimizing the sequence for a fixed protein backbone using rotamer based sequence search, and optimizing the backbone for a fixed amino acid sequence using atomic-resolution structure prediction. We find that such a flexible backbone design method better recapitulates protein family sequence variation than sequence optimization on fixed backbones or randomly perturbed backbone ensembles for ten diverse protein structures. For the SH3 domain, the backbone structure variation in the family is also better recapitulated than in randomly perturbed backbones. The potential application of this method as a model of protein family evolution is highlighted by a concerted transition to the amino acid sequence in the structural core of one SH3 domain starting from the backbone coordinates of an homologous structure.     

Accurate computer-based design of a new backbone conformation in the second turn of protein L.

The rational design of loops and turns is a key step towards creating proteins with new functions. We used a computational design procedure to create new backbone conformations in the second turn of protein L. The Protein Data Bank was searched for alternative turn conformations, and sequences optimal for these turns in the context of protein L were identified using a Monte Carlo search procedure and an energy function that favors close packing. Two variants containing 12 and 14 mutations were found to be as stable as wild-type protein L. The crystal structure of one of the variants has been solved at a resolution of 1.9 A, and the backbone conformation in the second turn is remarkably close to that of the in silico model (1.1 A RMSD) while it differs significantly from that of wild-type protein L (the turn residues are displaced by an average of 7.2 A). The folding rates of the redesigned proteins are greater than that of the wild-type protein and in contrast to wild-type protein L the second beta-turn appears to be formed at the rate limiting step in folding.     

A simple model of backbone flexibility improves modeling of side-chain conformational variability

The considerable flexibility of side-chains in folded proteins is important for protein stability and function, and may have a role in mediating allosteric interactions. While sampling side-chain degrees of freedom has been an integral part of several successful computational protein design methods, the predictions of these approaches have not been directly compared to experimental measurements of side-chain motional amplitudes. In addition, protein design methods frequently keep the backbone fixed, an approximation that may substantially limit the ability to accurately model side-chain flexibility. Here, we describe a Monte Carlo approach to modeling side-chain conformational variability and validate our method against a large dataset of methyl relaxation order parameters derived from nuclear magnetic resonance (NMR) experiments (17 proteins and a total of 530 data points). We also evaluate a model of backbone flexibility based on Backrub motions, a type of conformational change frequently observed in ultra-high-resolution X-ray structures that accounts for correlated side-chain backbone movements. The fixed-backbone model performs reasonably well with an overall rmsd between computed and predicted side-chain order parameters of 0.26. Notably, including backbone flexibility leads to significant improvements in modeling side-chain order parameters for ten of the 17 proteins in the set. Greater accuracy of the flexible backbone model results from both increases and decreases in side-chain flexibility relative to the fixed-backbone model. This simple flexible-backbone model should be useful for a variety of protein design applications, including improved modeling of protein-protein interactions, design of proteins with desired flexibility or rigidity, and prediction of correlated motions within proteins.     

Modeling backbone flexibility to achieve sequence diversity: the design of novel alpha-helical ligands for Bcl-xL

Computational protein design can be used to select sequences that are compatible with a fixed-backbone template. This strategy has been used in numerous instances to engineer novel proteins. However, the fixed-backbone assumption severely restricts the sequence space that is accessible via design. For challenging problems, such as the design of functional proteins, this may not be acceptable. Here, we present a method for introducing backbone flexibility into protein design calculations and apply it to the design of diverse helical BH3 ligands that bind to the anti-apoptotic protein Bcl-xL, a member of the Bcl-2 protein family. We demonstrate how normal mode analysis can be used to sample different BH3 backbones, and show that this leads to a larger and more diverse set of low-energy solutions than can be achieved using a native high-resolution Bcl-xL complex crystal structure as a template. We tested several of the designed solutions experimentally and found that this approach worked well when normal mode calculations were used to deform a native BH3 helix structure, but less well when they were used to deform an idealized helix. A subsequent round of design and testing identified a likely source of the problem as inadequate sampling of the helix pitch. In all, we tested 17 designed BH3 peptide sequences, including several point mutants. Of these, eight bound well to Bcl-xL and four others showed weak but detectable binding. The successful designs showed a diversity of sequences that would have been difficult or impossible to achieve using only a fixed backbone. Thus, introducing backbone flexibility via normal mode analysis effectively broadened the set of sequences identified by computational design, and provided insight into positions important for binding Bcl-xL.     

Computational de novo design of a four-helix bundle protein—DND_4HB

The de novo design of proteins is a rigorous test of our understanding of the key determinants of protein structure. The helix bundle is an interesting de novo design model system due to the diverse topologies that can be generated from a few simple α-helices. Previously, noncomputational studies demonstrated that connecting amphipathic helices together with short loops can sometimes generate helix bundle proteins, regardless of the bundle''s exact sequence. However, using such methods, the precise positions of helices and side chains cannot be predetermined. Since protein function depends on exact positioning of residues, we examined if sequence design tools in the program Rosetta could be used to design a four-helix bundle with a predetermined structure. Helix position was specified using a folding procedure that constrained the design model to a defined topology, and iterative rounds of rotamer-based sequence design and backbone refinement were used to identify a low energy sequence for characterization. The designed protein, DND_4HB, unfolds cooperatively (T_m >90°C) and a NMR solution structure shows that it adopts the target helical bundle topology. Helices 2, 3, and 4 agree very closely with the design model (backbone RMSD = 1.11 Å) and >90% of the core side chain χ1 and χ2 angles are correctly predicted. Helix 1 lies in the target groove against the other helices, but is displaced 3 Å along the bundle axis. This result highlights the potential of computational design to create bundles with atomic-level precision, but also points at remaining challenges for achieving specific positioning between amphipathic helices.     

De novo protein design. II. Plasticity in sequence space.

It is generally accepted that many different protein sequences have similar folded structures, and that there is a relatively high probability that a new sequence possesses a previously observed fold. An indirect consequence of this is that protein design should define the sequence space accessible to a given structure, rather than providing a single optimized sequence. We have recently developed a new approach for protein sequence design, which optimizes the complete sequence of a protein based on the knowledge of its backbone structure, its amino acid composition and a physical energy function including van der Waals interactions, electrostatics, and environment free energy. The specificity of the designed sequence for its template backbone is imposed by keeping the amino acid composition fixed. Here, we show that our procedure converges in sequence space, albeit not to the native sequence of the protein. We observe that while polar residues are well conserved in our designed sequences, non-polar amino acids at the surface of a protein are often replaced by polar residues. The designed sequences provide a multiple alignment of sequences that all adopt the same three-dimensional fold. This alignment is used to derive a profile matrix for chicken triose phosphate isomerase, TIM. The matrix is found to recognize significantly the native sequence for TIM, as well as closely related sequences. Possible application of this approach to protein fold recognition is discussed.     

XENet: Using a new graph convolution to accelerate the timeline for protein design on quantum computers

Graph representations are traditionally used to represent protein structures in sequence design protocols in which the protein backbone conformation is known. This infrequently extends to machine learning projects: existing graph convolution algorithms have shortcomings when representing protein environments. One reason for this is the lack of emphasis on edge attributes during massage-passing operations. Another reason is the traditionally shallow nature of graph neural network architectures. Here we introduce an improved message-passing operation that is better equipped to model local kinematics problems such as protein design. Our approach, XENet, pays special attention to both incoming and outgoing edge attributes. We compare XENet against existing graph convolutions in an attempt to decrease rotamer sample counts in Rosetta’s rotamer substitution protocol, used for protein side-chain optimization and sequence design. This use case is motivating because it both reduces the size of the search space for classical side-chain optimization algorithms, and allows larger protein design problems to be solved with quantum algorithms on near-term quantum computers with limited qubit counts. XENet outperformed competing models while also displaying a greater tolerance for deeper architectures. We found that XENet was able to decrease rotamer counts by 40% without loss in quality. This decreased the memory consumption for classical pre-computation of rotamer energies in our use case by more than a factor of 3, the qubit consumption for an existing sequence design quantum algorithm by 40%, and the size of the solution space by a factor of 165. Additionally, XENet displayed an ability to handle deeper architectures than competing convolutions.     

Unsupervised Inference of Protein Fitness Landscape from Deep Mutational Scan

The recent technological advances underlying the screening of large combinatorial libraries in high-throughput mutational scans deepen our understanding of adaptive protein evolution and boost its applications in protein design. Nevertheless, the large number of possible genotypes requires suitable computational methods for data analysis, the prediction of mutational effects, and the generation of optimized sequences. We describe a computational method that, trained on sequencing samples from multiple rounds of a screening experiment, provides a model of the genotype–fitness relationship. We tested the method on five large-scale mutational scans, yielding accurate predictions of the mutational effects on fitness. The inferred fitness landscape is robust to experimental and sampling noise and exhibits high generalization power in terms of broader sequence space exploration and higher fitness variant predictions. We investigate the role of epistasis and show that the inferred model provides structural information about the 3D contacts in the molecular fold.