Prediction of impacts of mutations on protein structure and interactions: SDM, a statistical approach, and mCSM, using machine learning

Abstract

Next‐generation sequencing methods have not only allowed an understanding of genome sequence variation during the evolution of organisms but have also provided invaluable information about genetic variants in inherited disease and the emergence of resistance to drugs in cancers and infectious disease. A challenge is to distinguish mutations that are drivers of disease or drug resistance, from passengers that are neutral or even selectively advantageous to the organism. This requires an understanding of impacts of missense mutations in gene expression and regulation, and on the disruption of protein function by modulating protein stability or disturbing interactions with proteins, nucleic acids, small molecule ligands, and other biological molecules. Experimental approaches to understanding differences between wild‐type and mutant proteins are most accurate but are also time‐consuming and costly. Computational tools used to predict the impacts of mutations can provide useful information more quickly. Here, we focus on two widely used structure‐based approaches, originally developed in the Blundell lab: site‐directed mutator (SDM), a statistical approach to analyze amino acid substitutions, and mutation cutoff scanning matrix (mCSM), which uses graph‐based signatures to represent the wild‐type structural environment and machine learning to predict the effect of mutations on protein stability. Here, we describe DUET that uses machine learning to combine the two approaches. We discuss briefly the development of mCSM for understanding the impacts of mutations on interfaces with other proteins, nucleic acids, and ligands, and we exemplify the wide application of these approaches to understand human genetic disorders and drug resistance mutations relevant to cancer and mycobacterial infections.

Statement for a Broader Audience

Genetic or somatic changes in genes can lead to mutations in human proteins, which give rise to genetic disorders or cancer, or to genes of pathogens leading to drug resistance. Computer software described here, using statistical approaches or machine learning, uses the information from genome sequencing of humans and pathogens, together with experimental or modeled 3D structures of gene products, the proteins, to predict impacts of mutations in genetic disease, cancer and drug resistance.

1. INTRODUCTION

Next‐generation sequencing methods have not only allowed an understanding of genome sequence variation during the evolution of organisms1 but have also provided invaluable information about human genetic disorders and cancer.2, 3 At the same time, genome sequences have facilitated our understanding of the emergence of resistance to drugs in tumors as well as antibiotics in infectious disease.4 A major challenge is to distinguish mutations that are “drivers” of disease or drug resistance from “passengers” that are neutral in these respects or even selectively advantageous to the organism. This requires understanding of the impacts of missense mutations in gene expression and regulation, and on protein function. Missense mutations can disrupt function not only by modulating protein stability but also through disturbing interactions with other biological molecules, including proteins, nucleic acids or polysaccharides, small molecule ligands, and metal ions.

Experimental approaches to understanding changes in stability and interactions between wild‐type and mutant proteins are clearly most accurate but are also time‐consuming and costly. Over the past three decades, this has encouraged the development of computational techniques, not only sequence‐based methods using support vector machines (INPS),5 neural networks,6 and decision trees (iPTREE‐STAB and MuStab)7, 8 but also structure‐based techniques using potential‐energy‐based approaches,9, 10, 11 machine learning algorithms,12, 13, 14, 15, 16, 17, 18, 19 or combinations of them.20 The development and validation of these computational methods have also been supported by databases documenting experimentally defined changes in free energy between the wild‐type and mutant proteins.21, 22, 23

Here, we briefly review various computational tools used to predict the impacts of mutations on protein function through impairment of stability and interactions with other molecules including proteins, nucleic acids, and small molecules. We then focus on two structure‐based approaches that were originally developed in Cambridge in the Blundell lab. The first of these, site‐directed mutator (SDM),24, 25 published in 1997, is based on conformationally constrained, environment‐specific substitution tables (ESSTs).26 The second more recent approach published in 2013, mCSM, uses graph‐based signatures to represent the wild‐type structural environment in order to use machine learning to predict the effect of mutations on stability. The mCSM family of software has been further developed by Douglas Pires and David Ascher after they left the Blundell laboratory;27 the current state of the mCSM software can be accessed at the Ascher Laboratory in Melbourne, Australia (https://biomedicalsciences.unimelb.edu.au/sbs-research-groups/biochemistry-and-molecular-biology-research/ascher-laboratory-structural-biology-and-bioinformatics).

In 2014, we brought together SDM and mCSM for predicting the impacts of mutations on protein stability as DUET.18 Here, we focus on developments that are more recent. These include elaboration of SDM to consider the effects of depth from the surface and the residue packing density on the impacts of mutations on protein stability28 and broadening the application of mCSM to consider impacts on protein function not only of residue mutations that impair interactions with other macromolecules19 but also small molecule ligands, which are critical for understanding drug resistance.16 We exemplify recent uses of these approaches to understand human genetic disease, mutations in cancer from the Cancer Gene Census of COSMIC,29 and resistance mutations occurring in mycobacterial and other infections.

2. SEQUENCE‐BASED PREDICTION OF IMPACT OF MUTATION ON PROTEIN FUNCTION

The availability of extensive exome sequencing data provides a wealth of previously unknown mutations that are either causal or lead to increased susceptibility to disease by affecting protein function, both outcomes providing challenges to human health and design of new medicines. Computational approaches to prediction of impacts of the mutations can be based on understanding sequence conservation, for example, SIFT,30, 31 PROVEAN,32 PANTHER‐PSEP,33 MutationAssessor,34 MutPred,35 LRT,36 FATHMM,37 and DEOGEN.38 More recently, machine learning has become more frequently used, and be trained on data describing annotation of functions, or sequence and structural features; for example, PolyPhen‐2,39 MutationTaster,40 SNAP,41 SNPs&GO,42 nsSNPAnalyzer,43 PhD‐SNP,44 SuSPect,45 DEOGEN2,46 and PMut.47 Various ensemble‐based methods, which exploit a gamut of available independent predictors, have also been developed. They include REVEL,48 MetaLR/support vector machine (SVM),49 Eigen,50 CADD,51 DANN,52 KGGSeq,53 Condel,54 and PON‐P.55 In addition, independent scoring functions based on sequence conservation have been developed to quantify the impact of mutations: they include PhyloP,56 SiPhy,57 GERP++,58 and GV‐GD.59

Detailed descriptions of different variant predictors have been reviewed elsewhere.60, 61, 62, 63, 64 Computational prediction tools rely on established databases like VariBench,65 VariSNP,66 dbSNP,67 ClinVar,68 UniProtKB,69 and PhenCode70 for variation benchmark data sets to develop, train, and validate the software.

3. STRUCTURE‐BASED PREDICTION OF THE IMPACT OF MUTATION ON PROTEIN STABILITY

Structure‐based methods use a combination of sequence and structural information. They can be grouped into statistical, physical, and empirical methods based on the potential energy functions. Statistical potentials include SDM,24, 25, 28 PoPMuSiC 2.0,9 CUPSAT,71 AUTO‐MUTE,72 I‐Mutant 2.0,13 DFIRE,73 Hoppe,74 PROTS,75 PROTS‐RF,76 Delaunay tessellation based four‐body statistical scoring function,77 MAESTRO,20 and Rosetta.11 Methods based on empirical energy functions include FoldX,78 ERIS,79 PEAT‐SA,80 and LIE.81 More accurate but computationally expensive physical‐energy‐based methods include CC/PBSA82 and EGAD.83 Machine‐learning approaches include mCSM15 described below, as well as Pro‐Maya, a mutant stability predictor that uses available data on mutations.84 iStable85 and DUET,18 described below, employ machine learning to provide a consensus prediction based on more than one independent predictor. A review with further information about the software tools is available.86 Databases that support the development and validation of the software include ProTherm,21 Platinum,22 and SKEMPI,23 which contain data on changes in protein stability, ligand affinity, and protein–protein interactions upon mutation, respectively.

4. SITE‐DIRECTED MUTATOR

We first focus on SDM,24, 25, 28 a statistical method developed in the Blundell lab that exploits protein sequence and structural data to predict the impacts of mutations on protein stability (Figure 1). SDM uses ESSTs derived from the analysis of amino acid substitutions that are tolerated within families of homologous proteins of known 3‐D structure.26 In a recent comparison study among 14 most efficient mutant prediction tools, SDM228 was ranked in the top five that are least biased toward stabilizing and destabilizing mutants.87 SDM2 is available at http://structure.bioc.cam.ac.uk/sdm2. Below we briefly describe the method.

Figure 1.

Site‐directed mutator (SDM) method. (a) Sequence‐structure alignments from TOCCATA for protein families are used to calculate the ESSTs. (b) 216 ESSTs are calculated using a combination of eight side‐chain hydrogen bonding patterns, nine main‐chain conformations based on ϕ and ψ dihedral angles, and three residue‐occluded surface packings. (c) The folding–unfolding free energy diagram represented as a thermodynamic cycle for site‐directed mutagenesis. The difference in free energy of unfolding of the wild‐type residue j, and mutant residue k, is related to the free energy changes associated with the mutation in the unfolded and folded states. Using ESSTs, the difference in stability score upon mutation [as shown in Equation (2)] is calculated analogously with Equation (1). As an example, the structure of p53 (PDB http://firstglance.jmol.org/fg.htm?mol=2OCJ) is used to illustrate the mutation at residue position 282 from arginine to tryptophan. Parts of the figure were derived from the graphical abstract taken from Ref. [27]

ESSTs are constructed using the TOCCATA database (Skwark, Torres, Ochoa‐Montano, and Blundell, manuscript in preparation) that contains 2,054 protein sequence‐structure alignments of homologous families taken from SCOP88 and CATH89 domain classification databases and represents a total of 12,038 structures (Figure 1a). In early versions, structural features including the main‐chain conformational angles, relative solvent accessibility (RSA), and hydrogen bonding patterns were used to define a set of local structural environments for the purpose of calculating ESSTs.26 These structural environments are characterized by distinct amino acid substitution propensities and have been successfully used to identify templates in homology modelling90 and finding key catalytic residues,91 as well as prediction of impacts on protein stability upon mutation.24 More recently, in SDM2,28 we introduced a combination of residue‐occluded surface packing (OSP)92, 93 and residue depth.94, 95 We have shown that the residue conservation progressively increases with residue depth and packing density and can be used to classify disease and non‐disease mutations.96 As default, SDM2 uses 216 ESSTs defined by the combination of nine main‐chain conformations, three residue occluded surface packings, and eight hydrogen bonding patterns [Figure 1(b)]. ESSTs defined using OSP showed an improvement in the quality of prediction compared to the previous version of SDM that used 54 ESSTs with RSA.25

The reversible folding and unfolding processes, represented as a thermodynamic cycle, are used to estimate the energetic contribution for a mutation of residue type j in the wild‐type protein to residue type k (Figure 1c). The relationship between the difference in free energy of unfolding of the wild type and mutant and the respective free energy changes associated with the mutation of residue j to k in the unfolded and folded states is expressed as:

$$\mathrm{∆∆}G=\left(∆G_k^{U-F}-∆G_j^{U-F}\right)=\left(∆G_{\mathit{jk}}^U-∆G_{\mathit{jk}}^F\right)$$ (1)

As it is time‐consuming to use experiments to measure free energies of denaturation of mutant $∆G_k^{U-F}$ and wild‐type proteins $∆G_j^{U-F}$, SDM uses ESSTs to calculate the difference in the stability scores (∆∆S) of the unfolded ($∆S_{\mathit{jk}}^U$) and folded ($∆S_{\mathit{jk}}^F$) state for the wild‐type and mutant protein structures, respectively, as:

$$\mathrm{∆∆}S=\left(∆S_{\mathit{jk}}^U-∆S_{\mathit{jk}}^F\right)-\left(∆S_{\mathit{jk}}^{\text{Disrupt}}-∆S_{\mathit{jk}}^{\text{Cavity}}\right)$$ (2)

where

$$∆S_{\mathit{jk}}^F=-\ln \left\{\frac{P\left(r_k/R_j,{\epsilon }_{\mathrm{wt}}\right)}{P\left(r_j/R_j,{\epsilon }_{\mathrm{wt}}\right)}\right.\times \left.\frac{P\left(r_k/R_k,{\epsilon }_{\mathrm{mut}}\right)}{P\left(r_j/R_k,{\epsilon }_{\mathrm{mut}}\right)}\right\}$$ (3)

where P(r_k/R_j, ε_wt) is the conditional probability for replacement of residue type R_j in the wild‐type environment ε_wt by residue r_k in an undefined environment and P(r_j/R_k, ε_mut) is the conditional probability for replacement of residue type R_k in the mutant environment by residue r_j in an undefined environment. Reference‐state probabilities P(r_j/R_j, ε_wt) and P(r_k/R_k, ε_mut) are introduced to normalize probabilities combined from different substitution tables. The difference in stability score of the unfolded state $∆S_{\mathit{jk}}^U$ is calculated in a similar way using conformationally constrained substitution tables representing non‐hydrogen‐bonded, surface‐exposed amino acids whose main‐chain dihedral angles fall outside regular secondary structure regions of the Ramachandran phi–psi map. In addition to the previously defined disruption penalty $∆S_{\mathit{jk}}^{\text{Disrupt}}$ term for buried residues (24), a cavity penalty term $∆S_{\mathit{jk}}^{\text{Cavity}}$ for the substitution of buried bulky hydrophobic residues (Phe, Leu, and Ile) by Ala or Val is added.28 All residues with RSA <17% are considered to be buried.

5. MACHINE LEARNING USING MUTATION CUTOFF SCANNING MATRIX

The family of mCSM computer programs uses the graph‐based approach based on Cutoff Scanning Matrix (CSM)97 to predict the impact of point mutations not only on protein stability but also on protein–protein, protein‐nucleic acid, and protein‐ligand affinities.15 Feature vectors, known as mCSM signatures, defined as inter‐atomic distance patterns around the mutated residue, are used to describe the structural environment in the wild‐type protein. Supervised machine learning is used to train predictive models using mCSM signatures (Figure 2). The mCSM predictive models for protein stability, protein–protein, and protein‐nucleic acid interactions are trained using thermodynamic data sets taken from ProTherm,21 SKEMPI,23 and ProNIT21 databases, respectively. mCSM is available at http://biosig.unimelb.edu.au/mcsm/.

Figure 2.

Mutation cutoff scanning matrix (mCSM) method. Starting from the mutant site of the wild‐type protein structure, the method calculates an atom contact graph based on a given distance criterion that defines the wild‐type residue environment. The cumulative distribution obtained from the distances of all pairwise atoms in the contact graph is used to construct the mCSM signature. Pharmacophore count vectors are used to describe the frequencies of atom classes found in the wild‐type and mutant residues. The pharmacophore change vector is used to quantify the difference between the mutant and wild‐type pharmacophore count vectors. mCSM signatures and the pharmacophore change vectors, along with their respective experimental thermodynamic stabilities and affinity data, were used to train predictive models using supervised machine learning techniques to predict the effects of mutations on stability and affinity

The residue environment is defined using graph‐based distance patterns, constructed by collecting atoms within a radial cutoff distance of 30 Å from the centroid of the wild‐type residue atoms. Atoms represent the nodes of the contact graph and edges that connect them are defined by the cutoff distance. The cumulative distribution obtained from the distances of all pairwise atoms in the contact graph forms the major component of mCSM signature that represents the environment of the wild‐type residue. The cumulative distributions were segmented using three types of atom classification as described previously.98

The mCSM signature captures the environment of the wild‐type residue only and, therefore, homology models of the mutants are not required as they are for SDM. mCSM uses a pharmacophore count vector to describe the frequencies of eight atom classes (hydrophobic, positive, negative, hydrogen acceptor, hydrogen donor, aromatic, sulfur, and neutral) based on PMapper program.99 The difference between the mutant and wild‐type pharmacophore count vectors is used to define the pharmacophore change vector, which is then appended to mCSM signature. Experimental conditions including pH, temperature, as well as the RSA of wild‐type residue are added to the mCSM signature.

Extensions to the mCSM method have been developed in Cambridge to predict the impact of mutation on protein–ligand (mCSM‐lig)16 and protein–protein interactions (mCSM‐PPI2),27 and more recently in Fiocruz, Brazil and in Melbourne, Australia by Douglas Pires and David Ascher for protein–nucleic acid (mCSM‐NA),19 antibody–antigen (mCSM‐AB)17 interactions and protein conformations and dynamics in combination based on normal mode dynamics (DynaMut).100

6. DUET

The use of ESSTs to define substitution probabilities in local structural environments and mCSM structural signatures to define long‐range interactions represent two independent complementary approaches to predict the impacts of mutations on protein stability. DUET18 was developed to leverage the individual strengths of SDM and mCSM into a consensus machine‐learning‐based predictor using a SVM101 trained with sequential minimal optimization. The SVM‐based supervised machine‐learning algorithm uses inputs as complementary features, including residue secondary structural annotation from SDM, pharmacophore change vector from mCSM, as well as the individual predictions from SDM and mCSM. The regression model tree is used prior to the machine learning in order to optimize SDM prediction results in combination with residue RSA. The method was trained on the test set of 2,297 mutant stability data taken from ProTherm database. The performance of the predictive model was tested on two independent blind sets that included thermodynamic stability data for 42 mutations in tumor suppressor p53 protein. In all cases, the method achieved higher performance compared to SDM and mCSM.18 DUET is available at http://structure.bioc.cam.ac.uk/duet.

We have recently performed a comparison of DUET and the latest version of the well‐established force‐field‐based mutant‐stability predictor FoldX 5.0.78 The comparison was done using the largest ProTherm benchmark containing 2,648 mutant stability data. DUET achieved a significantly higher Pearson's correlation coefficient (r) of 0.68 between the predicted and the experimental stability change compared to 0.51 obtained using FoldX (Figure 3). The root mean square error obtained for DUET and FoldX is 1.09 and 1.71, respectively.

Figure 3.

Scatterplot comparing DUET and FoldX stability prediction using the ProTherm benchmark containing 2,468 mutant thermodynamic stability data. Data points for DUET and FoldX are shown as dots (in blue) and triangles (in red), respectively. Pearson's correlation coefficient (r) for DUET and FoldX is shown at the lower right corner of the plot

7. EXAMPLES: UNDERSTANDING MUTATIONS IN GENETIC DISEASE, CANCER AND DRUG/ANTIMICROBIAL RESISTANCE

SDM2, mCSM, and DUET approaches have been used to understand the effects of mutations on human cancer‐related genes in the COSMIC database,102 inhibition of inosine‐5′‐monophosphate dehydrogenase in Mycobacterium tuberculosis,103 isoniazid and rifampicin resistance in Mycobacterium tuberculosis,104 phosphodiesterase somatic mutations implicated in cancer and retinitis pigmentosa,105 protein presenilin 2 linked to familial Alzheimer's disease,106 rifampin resistance in Mycobacterium leprae,107 ESX‐5 type VII‐secreted protein implicated in the transmission of Mycobacterium tuberculosis,108 carbapenem resistance in Acinetobacter baumannii,109 drug resistance in epidermal growth factor receptor,110 POT1 gene implicated in maintaining the telomere homeostasis,111 human Atg8 gene involved in autophagy,112 K13 propeller domain associated with artemisinin drug resistance,113 missense mutation in SLC6A1 associated with Lennox–Gastaut syndrome,114 and TN1 Gene mutations involved in coat plus syndrome.115

Some of these applications are ongoing, a most important example of which is COSMIC. Experimental structural data relevant to COSMIC are limited. Most genes in the Cancer Gene Census (the most frequently mutated cancer genes with mutations in COSMIC) are multidomain, but many of the reported structures for the gene products are single domain or functional substructures of the complete protein. We have already spent 2 years predicting models and the impacts of the 172,000 unique mutations reported for the 700 gene products of the Cancer Census (Alsulami, Torres, and Blundell, Unpublished). This has underlined the importance of full 3D models for the multidomain and multicomponent systems and software so that protein–protein, protein–nucleic acid, protein–ligand and other interactions can be considered. These factors are also important for understanding mutations in genetic disease and drug resistance in cancer therapeutics and antimicrobials.

An example of understanding genetic disease is the application of the SDM and mCSM software to understanding von Hippel–Lindau Syndrome, which leads to the development of clear cell renal carcinoma and is caused by mutations in the VHL gene that encodes protein pVHL (Figure 4). We earlier sought to identify a patient's risk of developing clear cell renal carcinoma from understanding the impacts of the mutations not only on pVHL protein stability but also on the affinity for the binding partners, HIF‐1α, and Elongin B and C.116, 117 At that time of the first paper,116 SDM was the only structure‐guided program available from our lab. We showed that the molecular mechanisms of renal cell carcinoma (RCC) and pheochromocytoma (PCC) in VHL disease appear to be decoupled: RCC can arise from disruption of HIF‐1α interactions or binding at the elongin B interface, while PCC is driven by mutations at binding site of elongin C. In the second publication when mCSM software117 was available, we used DUET18 to estimate impacts on stability and mCSM software to assess the impacts on protein–protein interactions. The observation of important mutations in protein–protein interactions has proved typical of many human Mendelian diseases, although mutations that affect stability are very common.118, 119, 120, 121

Figure 4.

Clear cell renal carcinoma in von Hippel–Lindau disease. (a) Ternary complex of pVHL with Elongin B/C, critical for pVHL stability and function. (b) Protein–protein interactions mediated by arginine (R107), mutation of proline at position 86 to arginine alters complementarity of charge between subunits and destabilize protein–protein interactions, essential to function. Hydrogen bond interactions are shown as springs in green

The software SDM and mCSM can also be used to begin to understand the emergence of drug resistance. Looking at single‐point coding mutations in Mycobacterium tuberculosis, strong correlations have been observed between the structural features of mutations and their link to antibiotic sensitivity.4 This suggests that prediction of drug resistance mutations before they arise, using our software with genomic sequences and structures, could be useful in guiding drug development. This idea has been applied in our efforts to develop small‐molecule inhibitors of GuaB2 to treat Mycobacterium tuberculosis infections.122 Computational saturation mutagenesis on the crystal structure of GuaB2 with VCC234718 using SDM, mCSM, and mCSM‐lig suggested that a resistance mutation would likely occur at Y487, altering interactions with the inhibitor but not NAD, while not disrupting protein stability or the interactions between the homo‐tetramer units. In fact, Y487C was observed to be an important resistance mutation. Other molecules were identified that did not make these interactions and were active against the Y487C mutant. Further work is required to show that inhibitors that are active against Mtb in vitro through the inhibition of GuaB2 have therapeutic benefits in vivo.

These and many other analyses highlight the importance of considering the structural environment of a mutation in order to understand the molecular and biological consequences. In evolution, they may indicate new functions emerging. For genetic disease, they should help identify mechanisms of disease emergence, while for the emergence of drug resistance they may help identify molecules that are less likely to lead to the emergence of resistance.

The authors recognize that there are limitations of most computational approaches that attempt to predict mutational effects on the whole organism. As demonstrated above, we have moved in the direction of recognizing that mutations affect higher‐level interactions rather than just the stability of individual proteins, by considering the impacts of mutations on the interactions between various components of macromolecular assemblies. In cases of resistance to drugs and genetic disease, the correlations are evident at a whole organism level. However, the level of expression in vivo, how this exceeds the threshold required for function in the organism and how alterations in stability will affect protein levels and protein solubility are more difficult to assess. Furthermore, for diploid organisms, it is unclear whether a single mutation will show a phenotype. None of these factors is currently considered. Hence, while predictions of mutational effects on stability are undoubtedly important, translation of these to predict disease phenotypes will likely require individual disease‐specific training sets to calibrate predictions of programs like SDM2, mCSM, and DUET. Thus, studying the impact of mutations on in vivo protein expression, as well as within the context of recessive and dominant mutants must in future become central to the system‐level understanding for the whole organism.

Pandurangan AP, Blundell TL. Prediction of impacts of mutations on protein structure and interactions: SDM, a statistical approach, and mCSM, using machine learning. Protein Science. 2020;29:247–257. 10.1002/pro.3774