Machine Learning for Prioritization of Thermostabilizing Mutations for G-Protein Coupled Receptors

Abstract

Although the three-dimensional structures of G-protein coupled receptors (GPCRs), the largest superfamily of drug targets, have enabled structure-based drug design, there are no structures available for 87% of GPCRs. This is due to the stiff challenge in purifying the inherently flexible GPCRs. Identifying thermostabilized mutant GPCRs via systematic alanine scanning mutations has been a successful strategy in stabilizing GPCRs, but it remains a daunting task for each GPCR. We developed a computational method that combines sequence-, structure-, and dynamics-based molecular properties of GPCRs that recapitulate GPCR stability, with four different machine learning methods to predict thermostable mutations ahead of experiments. This method has been trained on thermostability data for 1231 mutants, the largest publicly available data set. A blind prediction for thermostable mutations of the complement factor C5a receptor 1 retrieved 36% of the thermostable mutants in the top 50 prioritized mutants compared to 3% in the first 50 attempts using systematic alanine scanning.

Significance

This study is a significant technical advance that demonstrates the potential of combining sequence (evolutionary coupling), structure (network analysis), and dynamics (energy of an ensemble of conformations)-based properties of α-helical transmembrane proteins as features to describe their thermostability. We have combined these features with four different “classification” machine learning algorithms and trained models on the largest public data set of G-protein coupled receptor thermostability mutants. We have performed a blind test of the algorithms on the complement C5a receptor 1, referred to C5aR henceforth, and efficiently recover the significantly thermostable positions with fewer “trial attempts” than sequential analysis. We also identify that network centrality features describing protein structure are the strongest predictors of potential thermostable positions.

Introduction

G-protein coupled receptors (GPCRs), the largest superfamily of drug targets, reside in membranes, where they coordinate agonist stimulation with activation of different G-proteins and/or β-arrestin signal transduction pathways. Apo and ligand-bound GPCRs exhibit a dynamic equilibrium among multiple functional conformational states (1) broadly classified as active and inactive states. The active state conformations of class A GPCRs are characterized by the large movement of the intracellular regions of transmembrane (TM) helices TM6, TM5, and TM7, as illustrated in Fig. 1 A using the crystal structures of adenosine receptor A_2AR. The relative population of these conformation states are modulated by the binding of ligands and G-proteins or other intracellular transducers (2, 3, 4). Therefore, to enable structure-based ligand design for GPCRs, it is imperative to solve the three-dimensional structures of GPCR conformations with various ligands and/or intracellular transducer proteins bound. However, the conformational flexibility of GPCRs poses significant challenges in stabilizing and purifying these proteins for structural studies. The breakthrough in protein engineering technologies in the past decade has resulted in a bounty of GPCR structures in inactive and active states that have provided many new functional insights (5).

Figure 1.

Analysis of the experimental thermostability data on 1231 thermostable GPCR mutants. (A) The overlay of the crystal structures of A_2AR in the antagonist-bound inactive (gray), agonist-bound active-intermediate (green), and agonist- and G-protein-bound fully active state (red). The intracellular view shows large movements in TM5, TM6, and TM7 upon activation. (B) Analysis of the location of the thermostable mutant positions in all the five experimental thermostability data sets. (C) Location of the thermostable mutants in the TM regions that are common in β₁AR, A_2AR inactive and active-intermediate states, and NTSR1 active-intermediate states. (D) Classification of the thermostable mutants by the amino acid properties. To see this figure in color, go online.

One such protein engineering technology that yielded multiple GPCR structures is via thermostabilization, pioneered by Tate and co-workers (6). This method involves performing systematic alanine scanning mutations to identify thermostabilizing mutant positions in the GPCR. These positions can be subsequently combined to yield a thermostable mutant that exhibits conformational homogeneity, enabling purification of a selected GPCR conformational state. Although this strategy has been successful for purifying multiple class A GPCRs in various conformational states (7, 8, 9, 10), deriving an optimal combination of thermostable mutations quickly escalates the cost and scale of experiments/research and development. Therefore, this procedure is only accessible to a handful of laboratories worldwide and requires smarter methods to democratize this process. Computational methods that can reliably predict the thermostabilizing mutations of a GPCR ahead of experiments greatly reduce the burden by prioritizing only promising alanine mutations. Our prior studies on properties of thermostable mutant GPCRs showed that energy function calculated using all-atom force fields and a conformational ensemble rather than a single structural model recapitulates thermostability and recovers more thermostabilizing mutations in the predictions (11).

Building on this success, in this work, we trained four different machine learning (ML) classifier algorithms (referred to as classifiers henceforth in the article) on the largest publicly available thermostability data for 1231 GPCR mutants from four different class A GPCRs. The classifiers were trained using a combination of 26 features calculated from 1) amino acid sequence co-variation analysis of all class A GPCRs, 2) GPCR structural features extracted using network analysis, and 3) thermodynamic energies calculated from an ensemble of conformations of the GPCRs that include a membrane potential. Using the trained classifiers (referred to as models henceforth) we performed a blind prediction of thermostabilizing mutations for the human complement C5a receptor, C5aR. Please note that the alanine scanning mutation experiments on C5aR experiments were performed at Pfizer (Groton, CT) concurrent with the predictions. Comparison of the predicted thermostable mutants to the experiments showed that we recovered 14 out of 34 thermostable mutants within the top 50 prioritized mutations, whereas systematic alanine scanning mutation has a probability of only capturing 1 mutant in 50 attempts. 4 out of the 14 thermostable mutants are strong thermostable mutants. We present a robust machine learning workflow that is extensible to other class A GPCRs. This makes the predictions of thermostabilizing mutations for any class A GPCR accessible to the larger community.

Methods

Experimental thermostability measurements

To train the classifiers, we have used the thermostability data on 1231 single alanine scanning mutants on four class A GPCRs (two sets of data for inactive state and active-intermediate state stabilization for A_2AR) measured by Tate and co-workers. We have thermostability data only for 200–250 mutations in each of the four GPCRs tested. We calculated features that describe GPCR structural stability using structural models that correspond to these residues for which thermostability data is available in the training data set. The experimental thermostability measurements were performed on antagonist cyanopindolol-bound turkey β1-adrenergic receptor (9), agonist 5′-N-ethylcarboxamidoadenosine-bound human adenosine receptor A_2AR (10), antagonist ZM-241385-bound A_2AR (8), agonist NTS1-bound rat neurotensin receptor NTSR1 (7), and antagonist ZD7155-bound human angiotensin receptor type 1 (AT₁R). Briefly, the experimental thermostability was measured by heating the detergent-solubilized mutant to an elevated temperature (∼28–32°C) and cooling to 0°C, and the amount of correctly folded receptor was determined by a radioligand (either an agonist or antagonist or both) binding assay. In the work presented here, we have defined the thermostability of the wild-type receptor as 100%, and any mutants that showed 100% or higher ligand binding compared to the wild-type receptor are defined as thermostable. For the β₁AR, A_2AR, and AT₁R receptor, the mutants were heated without any ligand present. In the case of NTSR1, two types of experiments were performed, one in which the mutants were heated in the presence of the agonist neurotensin, and the other experiment heated the receptor in the absence of neurotensin; the data that is used in our machine learning models only included data from heating without the presence of agonist.

Thermostability measurements for C5aR

The wild-type human C5aR receptor gene was cloned into mammalian cell expression vector pcDNA3.1 with the hemagglutinin signal sequence followed by a Flag epitope tag at the N-terminus and a 10× His-tag at the C-terminus. A library of 283 mutant C5a receptors was constructed with individual residues from Val35 to Leu311 of C5aR mutated to alanine or leucine (if the original residue is an alanine) by PCR-based site-directed mutagenesis. Each clone of the mutation was transient transfected in 6 ml of human HEK-293 Expi cells, and 6 million cells harvested after 2 days expression were pelleted in aliquots and frozen for thermostability screening. Six million cells expressing wild-type C5aR or mutant C5aR were suspended into 650 μL of phosphate-buffered saline supplemented with protease inhibitor cocktail. Thermostability of C5aR mutants was assessed in a high-throughput three-temperature assay screening, and all experiments were run in triplicate. For non-heat-treated samples, aliquots of cell suspension were kept on ice, whereas other aliquots of cell suspension were heated at 43°C (apparent melting temperature (Tm) determined previously) and 46°C for 15 min, respectively, and the samples were then cooled on ice for 5 min. The following radioligand binding assay was done in a 96-well format; the functional expression level of each C5aR mutant was assessed on non-heat-treated samples, and the thermal stability was assessed on thermal-treated samples in their cellular lipid environment. In each 200 μL of radioligand binding assay reaction, 20 μL of C5aR cellular suspension was incubated in binding buffer (75 mM Tris-HCl pH 7.4, 1 mM EDTA, 5 mM MgCl₂, 100 mM NaCl) with 100 nM [³H]compound 1 (PF-06733901) at room temperature for 1 h. Nonspecific binding of [³H]PF-06733901 was determined in the presence of 2.5 μM unlabeled ligand. Unbound radioligand was then removed from the reaction on PerkinElmer FilterMate (Waltham, MA), and radioactivity was counted by liquid scintillation on a PerkinElmer MicroBeta. Data were analyzed using Prism software (GraphPad).

A machine learning model for predicting GPCR thermostabilizing mutations

Using the largest set of experimental thermostability data publicly available, we have employed trained classifier models for predicting thermostabilizing mutations ahead of experiments. Our goal is to use machine learning methods to recover the maximum number of thermostable mutants in fewer mutant trials. We used supervised “classification” learning methods to identify the relationship between our feature set and the thermostability class. The four classifier algorithms used in this study are as follows: random forest (12) (RF), cost-sensitive RF (13) (CSRF), adaptive boosting (14) (AdaBoost) (AB), and gradient boost (15) (GB). The RF and CSRF methods use a collection of decision trees to make a prediction. CSRF penalizes trees that make a false negative prediction. AB and GB are boosting algorithms that improve predictions by providing higher weights to previously misclassified data points. A more detailed and relevant description of these decision tree algorithms are given in the Supporting Materials and Methods. We developed a workflow as shown in Fig. 2 to train and optimize classifiers. The two major sections of this workflow consist of the following: 1) preparing the experimental data for machine learning and 2) calculating features describing the thermostability of GPCRs.

Figure 2.

Workflow developed for optimizing, validating, and testing the machine learning methods. The figure shows the atomic level features that we have calculated to recapitulate the GPCR structural thermostability. Properties based on co-variation of amino acid sequence positions, structural properties such as degree, betweenness, and closeness centrality calculated from network analysis, and thermodynamic energies calculated from structural ensemble. The right side of the figure shows the algorithm we used for balancing the experimental thermostability that was the target data for training the machine learning models. To see this figure in color, go online.

Preparing experimental data for ML optimization

Partitioning the experimental data into training and testing ensembles

To train the classifiers, we have used the thermostability measured by Tate and co-workers on every residue in four different class A GPCRs listed above. Experimentally, we classify a mutant as thermostable if the experimental thermostability value is at or above the wild-type threshold, which is a score of 100. To convert this experimental data into labels for training classifiers, we used threshold ranking. If a mutation has a measured thermostability score of 100 (the score for the wild-type receptor) or more, it is considered thermostable and given a label of “1.” Any score below 100 is considered nonthermostable and given a label of “0.” These labels are used to classify our computationally calculated data set. We divided the five sets of experimental thermostability data on four GPCRs (A_2AR has two sets of data, one for the antagonist-bound inactive state and the other for the agonist-bound active-intermediate state (Fig. 1 A)) into five ensembles, E1–E5, shown in Table S1. In each of the five ensembles, we trained the classifiers using four data sets and held out the fifth data set as an unknown test set.

Balancing the experimental training data

85% of the data in each ensemble E1–E5 was used for training the classifiers, whereas 15% of the data was held out from training for classifier model validation. The experimental thermostability data on 1231 mutants contains 256 thermostable mutants and 975 nonthermostable mutants and, hence, is not balanced. Therefore, to balance the training data, we used the Synthetic Minority Oversampling Technique (SMOTE)-Tomek sampling (16) method to both undersample the majority class (nonthermostable mutants) and oversample the minority class (thermostable mutants) simultaneously. SMOTE oversamples the minority class using its k nearest neighbors (17). Tomek removes outliers from the majority class by using distance measurements between the two classes. SMOTE forms new minority class (thermostable mutants) data by interpolating between several minority class data points that lie together (16). The effect of balancing the data is shown in the Results.

Generating features for the machine learning algorithms

The quality of predictions made using machine learning algorithms requires “features” that accurately describe the outcome (in this case, thermostability of GPCR mutants). Identifying informative, discriminating, and independent GPCR features that best describe its thermostability is a crucial step for effective ML classification algorithms. As listed in Table S2, we have used features extracted from class A GPCR amino acid sequence analysis, structural analysis, and atomistic energy calculated from conformation ensembles. For assessing the importance of the features used in this study, we performed feature ranking using the importance measure, “mean decrease in impurity (MDI).” MDI was calculated using the inherent feature of importance attribute calculation in each of the classifiers. Nodes on a decision tree represent a condition for a single feature that will split data into separate classes. The measure that decides this condition is called impurity, usually referred to as GINI impurity for classification trees. For a decision tree, the effect of a feature on the impurity can be calculated. For a forest, this effect can be calculated across all trees to get the average impurity for a feature.

Sequence-based properties related to stability

To provide a comprehensive set of features to train the classifiers, we included amino acid sequence-based property, evolutionary coupling score (18) that quantifies the extent to which each residue is conserved or undergoes a correlated mutation. The higher the evolutionary coupling score, the stronger is the role of the residue in preserving the structure or function of the GPCR. The evolutionary score is calculated using a multiple sequence alignment, which we generated using 300 nonolfactory class A GPCR sequences and GPCRDB (19) web server. We performed evolutionary co-variation analysis using the “EV couplings” web server (http://EVfold.org) (18) with the multiple amino acid sequence alignment of 300 nonolfactory class A GPCRs as input, generated using the structure-based sequence alignment toolkit from “GPCRDB” (19). This analysis calculates the co-variation of any pair of amino acid positions in the class A GPCR sequence alignment. We converted the evolutionary coupling scores to z-scores.

Structural features that describe the thermostability

We have abstracted structural properties using a graphical network model of the GPCR crystal structures that are available or homology models if the crystal structures are not available. We used the “Protein Contact Atlas” web server (https://www.mrc-lmb.cam.ac.uk/rajini/index.html) and calculated residue-based structural properties, such as degree, centrality, betweenness, and closeness, from a single input of the structure or structural model of the GPCRs (20). The solvent-accessible surface area (referred to as solvated area) for each residue was calculated using the Parameter OPtimized Surfaces (POPS) (21) algorithm within “Protein Contact Atlas” (https://www.mrc-lmb.cam.ac.uk/rajini/index.html). Together, these structural properties capture the noncovalent interactions between residues in the protein framework. The score for each of these structural features was also converted to z-scores.

Energy features calculated from force field function

In our previous studies of GPCR thermostability, we have shown that the van der Waals and dihedral energy components calculated using the all-atom force field function from CHARMM leads to a 30% recovery of the thermostable mutants for several class A GPCRs in the top 50 predicted mutants (11). In this study, we have used all the energy components calculated using atomic force field functions, CHARMM27 (22) and ROSETTA (23), to describe the thermostability of GPCRs. The CHARMM energy function includes the valence bond energy, the angle energy, and the dihedral angle energy. It also includes nonbond energy terms, such as Coulombic energy and van der Waals energy. In this study, we have also included other energy terms derived from the statistical analysis of protein structures in the protein data bank, such as energy term favoring preferred Ramachandran backbone dihedral angles and statistics-based term favoring salt bridges available in the ROSETTA force field (23). We also included a feature that classifies an amino acid as hydrophobic or hydrophilic. We have included energy components from two force fields because the force field function and their respective parameters come from different sources. The ROSETTA force field parameters come from mining data using the protein structures from the protein data bank. However, we checked if any of the features from the CHARMM force field is correlated to any term from the ROSETTA force field by calculating the Pearson’s correlation coefficient. As expected and seen in Table S3, the correlation is poor (less than 0.29), and hence, these features are not correlated.

For the GPCRs used in the training set shown in ensembles E1–E5 in Table S1, there were crystal structures available. However, at the time we started this work, we did not have a crystal structure for the blind test case of C5aR, and hence, we have not used the C5aR crystal structure that has been published since (24). Starting from the respective crystal structure for each GPCR (see Table S4 for a list of crystal structures and their respective protein data bank identities), we used the LiticonDesign method (11,25,26) to generate a small ensemble of conformations for each GPCR. We have described the LiticonDesign method in detail in our previous work (11,26, 27, 28). Briefly, the LiticonDesign method for predicting thermostable GPCR mutants involves two steps: 1) using a starting structural model of the GPCR, the method generates a small ensemble of conformations that allows for the perturbations in the GPCR conformation caused by mutations and 2) an all-atom energy function to calculate the stability of the conformations that takes into account the difference in the structural stability of the mutants and the wild-type to score the positive thermostable mutants. Starting from an initial receptor structure or homology model in the case of C5aR, all of the seven TM helices are rotated simultaneously about their respective helical axes by ±5°. Thus, 2⁷ = 128 conformations are generated for the wild-type receptor. Then, we perform a mutation of each residue to alanine (alanine in the wild-type is mutated to leucine) in each of the 128 conformations and repack the side-chain conformations of all of the residues using SCWRL4.0 (29), followed by steepest descent energy minimization using the CHARMM27 force field for 1000 steps. Each of the energy components listed in Table S2 was calculated for the wild-type and the mutant for each of the 128 conformations and averaged over the 128 conformations for the wild-type and the mutant separately. Each energy component of the overall stability score was calculated as the difference in the average mutant energy of that particular energy component to the average wild-type energy. For example, the van der Waals energy component of the stability was calculated as stability score (vdw) = <E_vdw, _mutant > − < E_{vdw, wild-type} > for a given single mutant in which the respective energies are averaged over the 128 conformations. To give equal weight to every amino acid type in our scoring, we calculated the z-score from the stability score for every amino acid type. We will refer to this z-score as the stability score from here on in the article. The stability score for each energy component and for the mutant was also calculated using the Rosetta force field mpframework_fa_2007 as implemented in the Rosetta software (23). The different energy components listed in Table S2 were all calculated for each mutant.

Homology model generated for C5aR

We used the crystal structures for all the GPCRs used to train the machine learning models (Table S4). For the blind test case C5aR, we generated a homology model because this work was done before the report of the crystal structure of C5aR (24). C5aR showed low sequence homology with existing crystal structures at the time of this modeling effort. Crystal structures of the inactive state of μ, κ, and δ- opioid receptors, nociception receptor, and chemokine receptors CCR5 (30), CXCR1 (31), and CXCR4 (32) showed 17–23% sequence homology to human C5aR. Therefore, we used the GPCR-I-TASSER (33) method to build the homology models with multiple templates. Specifically, the human C5aR sequence was first threaded through the GPCR structure library to identify putative structures of fragment templates. After which, the fragments are assembled into full-length models by replica-exchange Monte Carlo simulations, which are assisted by a GPCR and membrane-specific force field and spatial restraints collected from mutagenesis experiments. All models were further refined by fragment-guided dynamics simulation to eliminate inter-TM steric clashes that improve the model quality. This led to C5aR structural models with a high confidence score (C-score = 2.3), especially for TM1–7 and ECL1–3 regions, allowing for further use in thermostability mutation predictions.

Properties that discriminate the experimental data on thermostable mutants

We performed principal component analysis (34,35) and linear discriminant analysis (LDA) (36,37) on the 26 features listed in Table S2. These two methods reduce the multidimensional space into fewer dimensions. The vectors obtained from these analyses best separate the thermostable from the nonthermostable mutant data. Based on principal component analysis and LDA of the entirety of the training subset of proteins, we were unable to separate classes based on these single value deconvolution methods (Fig. S1). LDA results suggest that the feature space cannot distinguish classes 0 and 1 linearly. These results further clarified our thinking to use ML methods.

Results

A comparison of the amino acid positions of thermostabilizing mutations in different receptors shows that there is no significant conservation between the amino acid position and thermostability. Thus, transferring thermostabilizing mutations between GPCRs works only if the receptor sequences are highly conserved (for example, between the turkey β₁AR and human β₁AR (38)). Therefore, a robust mutagenesis prediction method will help scientists reduce time and cost, focusing the number of trials on more promising amino acids mutations.

Thermostable mutants cluster on the intracellular region of TM helices 5 and 6

Four GPCRs have been subjected to comprehensive alanine scanning mutagenesis by Tate and co-workers, and the resulting mutants were tested for their thermostability using a radioligand binding assay (39). The turkey β₁-adrenoceptor (β₁AR) (9), human adenosine A_2A receptor (A_2AR) (8), and human AT₁R were stabilized in an inactive state that bind antagonists and with similar affinity to the wild-type receptors. The rat neurotensin receptor (NTSR1) (7) and A_2AR (10) were also thermostabilized in an active-intermediate state (Fig. 1 A) that binds agonists with a similar affinity to the wild-type receptor. A total of 1231 alanine/leucine mutants were tested for thermostability, out of which 256 mutants are thermostable. The criteria for thermostability we use here is if the mutant has the same or increased thermostability as the wild-type receptor. The mutants with >130% of the wild-type are strong thermostable mutants. Analysis of the positions of thermostable mutations show that they are distributed throughout the TM helices and loop regions with a higher percentage located in TM5, TM6, and intracellular loop 3 (Figs. 1 B and S2 A). The thermostabilizing mutations common to β₁AR and both conformational states of A_2AR and NTSR1 are clustered in the intracellular region of TM5, TM6, and intracellular loop 3 (Figs. 1 C and S2 B). The highly conserved residue in each TM helix is numbered as the 50th residue according to the Ballesteros-Weinstein residue numbering scheme for GPCRs (40). The clustering of mutations in this mobile domain of the GPCR highlights that GPCRs have evolved to retain flexibility in their dynamic equilibrium, and the mutation away from natural sequence effectively shunts this conformational flexibility. Understandably, the mutation of this most conserved residue in each TM helix decreases the stability of the receptor (Fig. S3). This may be related to the mechanism of receptor activation, which involves significant changes in the orientation of TM5 and TM6 with respect to TM3 (Fig. 1 A). However, the mutation of other conserved residues such as Asn or Tyr of the NPxxY motif on TM7 does not always destabilize the receptor (Fig. S3 B). GPCRs are in a dynamic equilibrium between inactive and active states, so thermostabilization of an inactive state may arise through mutations that prevent the transition to active states or destabilize the active states. Other mutations may stabilize the receptor through increasing contacts between residues, promoting the formation of interhelical backbone hydrogen bonds and/or entropic effects (25,41). Mutation of aliphatic hydrocarbon amino acids, such as Val, Ile, Leu, Met and Cys, more often lead to thermostabilization than other types of amino acids (Fig.1 D; (25,41)).

Merits of the features in recovering thermostable mutants

We assessed the merit of each of the 26 features listed in Table S2 to predict thermostable mutants by using the z-scores of each feature to test how well they recover the thermostable mutations in the 1231 mutant data set. The number of thermostable mutations recovered in the top 50 predicted mutants using each of the sequence, structural, and energy-based features as well as their sum are shown in a radial plot (Fig. 3 A). All the features recover more thermostable mutants in the top 50 predicted mutants compared to the random predictions, demonstrating the merit of these properties in describing the GPCR thermostability. The sum of the z-scores of all the 26 features recovers nearly threefold the number of thermostable mutants compared to the individual features because the individual features recover a distinct set of thermostable mutants. This suggests that the sum of the properties captures the GPCR thermostability better than any single property.

Figure 3.

(A) Merits of the sequence, structure, and dynamical-based properties used for training the classifier models. The recovery of thermostable mutants greater than 100% of the wild-type stability using each of the features. (B) The five figures here show the fraction of thermostable mutants recovered using the predicted prioritized list of mutants using different machine learning models on the five ensembles of data used for training and testing the four machine learning models in this work (for breakdown of the GPCR systems included in training and testing, see Table S1). The aqua colored curve in this figure is the recovery rate using the sum of the features without any machine learning. The curve titled sequential is the one representing the experimental test done in the consecutive order of the residue position in the amino acid sequence of the GPCR tested. To see this figure in color, go online.

Performance of the four classifiers in training and validation

The four classifiers, namely, RF, CSRF, AB (also known as AdaBoost), and GB, were trained and validated on five data ensembles shown in Table S1. To achieve the optimal parameters that will allow the model to generalize to new data, we performed hyperparameterization using the 85% of the balanced data for training and the 15% of the data held out for validation. We performed a grid search for each parameter in each classifier method only during the training of the models. We used the Matthews correlation coefficient (MCC) (42) as a metric to assess the performance of each of the classifier models because it is a composite function of the true positives, true negatives, false positives, and false negatives. The optimal parameters were chosen based on the best MCC yielded for the validation data. The best combination hyperparameters for each decision tree classifier are shown in Table S5 and were optimized for increasing the MCC for the validation data. Fig. S4 A shows the average of the four MCC ratios obtained with the optimized parameters for each of the four classifier models tested on the 15% held-out validation data. Fig. S4 B shows the average value of the MCC ratios across four models for each of the test proteins described in the data ensembles E1–E5 in Table S1. The RF, AdaBoost, and GB models perform better than the CSRF for the validation data, whereas the CSRF model performs well for the test set. For the problem in hand, the more important measure is how many thermostable mutants do we recover early on in the alanine scanning mutations. Fig. 3 B shows the fraction of thermostable mutations retrieved as a function of the prioritized classifier model-predicted prioritized alanine mutation list as compared to the systematic alanine scanning mutation list. We observe that 35–40% of the thermostable mutants are recovered in the top 50 prioritized mutant list compared to less than 10% in the systematic alanine scanning mutations. We note that the systematic alanine scanning mutations list contains only those residues for which we have reliable experimental thermostability results.

Blind predictions of thermostable mutations for C5aR

For the blind predictions of thermostable mutants in C5aR, we used the four classifiers trained on the thermostability on all five GPCR data sets (β1AR + A_2AR-inactive + A_2AR-active + NTSR1 + AT₁R). Using the four models, we performed blind predictions on the thermostabilizing single point mutants for the inactive state of wild-type C5aR GPCR. We predicted the thermostable mutants using the four optimized classifier models not previously trained or validated on C5aR. The systematic alanine mutagenesis experiments were performed simultaneous to the predictions to identify thermostable mutants for the antagonist-bound inactive state of C5aR using the experimental procedure described in the Methods.

We generated homology-based three-dimensional structural models for C5aR followed by the LiticonDesign method (11,26,27) to generate a small conformation ensemble starting from the homology model. We calculated the energy features listed in Table S2 for all the alanine scanning mutations using ROSETTA (23) and CHARMM 27 (22) force fields. We calculated the evolutionary coupling scores for C5aR using nonolfactory class A GPCRs multiple sequence alignment and the structural features using the homology model. These features were used in the models to calculate the probability score for mutating each residue in C5aR to alanine (or mutate to leucine if the wild-type is alanine). We compared our blind predictions to the experimental thermostabilities that we obtained from our collaborators at Pfizer. Fig. 4 A shows the fraction of thermostable mutants recovered from the Ala scanning mutation list prioritized by each ML model (green, blue, red, purple) compared to performing the sequential alanine scanning mutations in C5aR (black). The dots plotted in Fig. 4 A are the positions at which strong thermostable mutants (>130% wild-type ligand-binding density) are recovered. All four models recover the thermostable mutants at a higher recovery rate than the systematic alanine scanning mutation trials, but RF and adaptive boost models recover a consistently higher number of thermostable mutations, even for as few as 25 mutation trials (Table S6). The GB model recovers 14 thermostable mutants containing 4 strong thermostable mutants in the top 50 prioritized list. Thus, we demonstrate that sequence-, structure-, and dynamics-derived features are useful tools for efficiently prioritizing thermostabilizing mutations. Fig. 4 B shows the thermostable mutants recovered by the classifier models on a heat map on the structure of C5aR. The color range from green to red shows the different strengths of measured thermostability, and the thickness of the cartoon shown in Fig. 4 B is indicative of the average probability score of each residue being a thermostable mutant as predicted using the four optimized classifier models. It is evident from the figure that the highly thermostable residues (shown in red in Fig. 4 B) also show a high calculated probability score.

Figure 4.

(A) Recovery rate of thermostable mutants from the prioritized list of alanine scanning mutations for the blind test on C5aR, prioritized using different machine learning models. The x axis shows a prioritized list of mutations predicted by the four machine learning models. The black curve titled “sequential” is the recovery rate when performing systematic alanine scanning mutations along the sequence of human C5aR. (B) Two views of the heat map of thermostability of the residues in C5aR as measured in this study. The measured thermostability varies from green to red as shown in the heat map scale. The gray cartoon color indicates nonthermostable residue positions upon mutating them to Ala. The front view from TM helices TM5–TM7 is shown on the left, and the rotated view showing TM1–TM4 is shown on the right. The thickness of the cartoon is proportional to the calculated probability of mutating each of these residues positions to lead to thermostable mutants or not. The higher the probability of thermostabilizing mutations, the larger is the thickness of the backbone cartoon representation. Two C5aR mutations, L256A and I298A, with maximum and minimum experimental thermostability are shown with BW residue numbers as superscripts. The predicted probability of these two mutations are 0.85 and 0.29, respectively, represented by the difference in cartoon thickness. This correlation between experimental and predicted thermostability suggests that the calculated probability is an effective measure in separating the thermostable mutations of C5aR from the nonthermostable ones. To see this figure in color, go online.

Prediction of thermostable mutants for agonist-bound active-intermediate state of the adenosine A_2AR

Magnani et al. (10) identified single point thermostable mutants of the agonist-bound (5′-N-ethylcarboxamidoadenosine) active-intermediate state and the antagonist-bound (ZM-241385) inactive state of A_2AR (8) using alanine scanning mutagenesis across the entire receptor sequence. They identified mutations that retained the ligand binding at higher temperatures as compared to the wild-type (8). The majority of thermostabilizing mutations are distinct to either the inactive or active-intermediate state of the receptors, with a few common mutation positions (Fig. S5). Mutants present in TM helices TM1–TM4 (shown in magenta, Fig. S5) stabilize the antagonist-bound inactive state, whereas mutants that specifically stabilize the agonist-bound active-intermediate state are predominantly found in TM5–TM7 (shown in yellow, Fig. S5). This shows that the thermostable mutants are not easily generalizable, even for the same receptor in different conformational states.

We tested the optimized models in predicting the thermostabilizing mutations for the agonist-bound active-intermediate state of A_2AR (A_2AR^∗). We trained the classifiers on the experimental thermostabilizing data for the inactive state of β₁AR, the inactive state of A_2AR, the inactive state of AT₁R, and the agonist-bound active-intermediate state of NTSR1. Fig. 5 A shows the recovery curve for the agonist-bound A_2AR^∗ using the classifiers trained on the ensemble of thermostability data. The four models recover 17–20 thermostable mutants, out of which 11–14 are strong thermostable mutants in the top 50 prioritized alanine mutants as shown in Table S7. The probability of predicting the thermostabilizing residue positions in A_2AR in the active-intermediate state is depicted by the thickness of the backbone cartoon in the structure of A_2AR shown in Fig. 5 B. The red spots in Fig. 5 B indicative of high thermostable mutants also show a higher calculated probability of being thermostable. Thus, the models show a much better recovery compared to using a linear combination of features to predict thermostability. This test validates that our ML prioritized list recovers thermostable mutants much faster than systematic alanine scanning.

Figure 5.

(A) Recovery rate of thermostable mutants from the prioritized list of alanine scanning mutations for the agonist-bound active-intermediate state of A_2AR (A_2AR^∗), prioritized using different machine learning models. The black curve titled “sequential” is the recovery rate when performing systematic alanine scanning mutations along the sequence of human A_2AR. (B) Two views of the heat map of the measured thermostability of the residues in the active-intermediate conformational state of A_2AR. The measured thermostability varies from green to red as shown in the heat map scale. The gray cartoon color indicates nonthermostable residue positions upon mutating them to Ala. The front view from TM helices TM5–TM7 is shown on the left, and the rotated view showing TM1–TM4 is shown on the right. The thickness of the cartoon is proportional to the calculated probability of mutating each of these residue positions to lead to the thermostable mutants or not. The higher the probability of thermostabilizing mutations, the larger is the thickness of the backbone cartoon representation. Two A_2AR^∗ mutations, T224A and Y288A, with maximum and minimum experimental thermostability are shown with their BW residue numbers as superscripts. The calculated probability of these residues being thermostable are 0.70 and 0.12, respectively, represented by the difference in cartoon thickness. To see this figure in color, go online.

Relative importance of features describing thermostability

The most important features for thermostability prediction using a classifier model can be determined via feature selection (model agnostic) or directly from the classifier model by feature ranking. The feature ranking was generated using the importance measure, MDI, otherwise known as the “GINI importance” (defined in the Methods). Across the three classifier models tested (RF, AdaBoost, and GB), the “solvated area” feature is the most important, as shown in Fig. 6 A, and removal of this feature from the list shows poor recovery of thermostable mutants (Fig. 6 B). We also tested the performance of the sum of van der Waals and dihedral energy components of the CHARMM force field that we used in our previous study LiticonDesign (11). As seen in Fig. S6, the green curve shows a better recovery than the sequential test, but using all the 26 features described in this work surely shows much better recovery.

Figure 6.

(A) Ranking the weights of the 26 features used that describe the thermostability. On the y axis is the normalized weighted consensus score derived from feature ranking for each machine learning model. The normalization was done with respect to the solvated area feature. (B) The retrieval rate of thermostable mutants using 25 features after removing the high weighted feature—solvated area for C5aR. To see this figure in color, go online.

Discussion

In this study, we have used an innovative combination of molecular level features calculated from GPCR sequence co-variation analysis, structural features calculated from network property analysis, and thermodynamic energy calculation of GPCRs to recapitulate the thermostability of GPCRs. As shown in Fig. 3, the sum of these features, without any machine learning methods involved, recovers more thermostable mutations than a random prediction or systematic alanine scanning mutation analysis along the sequence of the GPCR. These 26 features when combined with four different classification-based machine learning methods provide robust classifier models for the prediction of thermostabilizing mutations as demonstrated in the blind test case of C5aR. The classifier models performed better than random predictions or systematic alanine scanning mutations along the amino acid sequence of C5aR and, hence, are useful in prioritizing alanine scanning mutations for experimentalists. We have demonstrated that the features calculated from a homology model of the GPCR structures works well for the blind test case C5aR.

We have used the largest publicly available thermostability data set (1231 mutants) for training the four classifier models. These data are for the inactive state of β₁AR, inactive state of A_2AR, inactive state of AT₁R, and A_2AR^∗ and NTSR1. As anticipated in many biological data sets, our thermostability data was imbalanced by negative data, which complicates the development of a classification model generalizable to new data (43). Using SMOTE-Tomek (16) to balance our training data resulted in a significant improvement in MCC.

Features that play a dominant role in recapitulating the thermostability of GPCRs

As seen in Fig. 6, the solvated area feature is important to recapitulate GPCR stability. Here, we analyzed the structural location of the residues that are correctly predicted to be thermostable by the solvated area feature in all the GPCRs studied here, including C5aR. There are a total of 144 thermostable mutations in all the GPCRs studied here, which are characterized by a high solvated area. The 144 residues are all facing the solvent and/or membrane bilayer (Fig. S7). Most of these residues are located in the edge of the TM or in the loop regions (Fig. S7). Some of the outward facing residues are also located in the middle of the TM region, facing the membrane. We also observe that mutation of the amino acids Leu, Val, Ile, Met, and Cys to Ala make up the largest percentage of thermostable mutations in the GPCRs studied here. Our previous molecular dynamics simulation studies have shown that the thermostable mutants have a higher number of interhelical van der Waals interactions compared to the wild-type (25,41,44, 45, 46, 47, 48). Although counterintuitive, mutating the larger Ile, Leu, and Met to smaller Ala leads to small perturbations in the interhelical packing, thus enhancing the main chain van der Waals packing between the TM helices. This is also recapitulated by the features fa-atr, fa_rep, and fa_pair, as shown in Fig. 6. For the blind prediction case, C5aR, the features solvated_area, fa-atr, fa_rep, and fa_pair that recapitulate interhelical interactions, show an increase in van der Waals attraction and a decrease in repulsion for the Leu to Ala mutations.

Predictions for thermostabilizing mutants of different conformation states

We have demonstrated a high recovery rate of thermostable mutants for GPCRs in different conformational states. Machine learning models trained on all data show a recovery of over 30% of thermostable mutants (and 20% of “highly” thermostable mutants) within the top 50 prioritized mutations. We note that the structural and energetic features describing the stability of the agonist-bound state were calculated using the structural models for the active-intermediate state rather than inactive states. We infer that it is important to calculate the features using the structural model of the conformation state for which thermostabilization is desired. Particularly, we believe individual features may capture distinct aspects of the phenomenon of thermostability, but the right ensemble of features is critical for better predictive ability.

Author Contributions

N.V. and S.A. designed the research work. S.M., S.G., S.A., and M.S. contributed to machine learning methods, developing features, and performing the calculations. X.C., X.Y., M.C.G., K.F.F., Y.C., V.S., and X.Q. contributed to experimental testing on C5aR. C.G.T provided the thermostability data for 1231 mutants. N.V., S.M., S.G., S.A., M.S., X.C., and X.Y. wrote the manuscript.

Editor: Rohit Pappu.