High risk genetic variants of human insulin receptor substrate 1(IRS1) infer structural instability and functional interference

Abstract Insulin receptor substrate 1(IRS1) is a signaling adapter protein encoded by the IRS1 gene. This protein delivers signals from insulin and insulin-like growth factor-1(IGF-1) receptors to the phosphatidylinositol 3-kinases (P13K)/protein kinase B (Akt) and Extracellular signal-regulated kinases (Erk) - Mitogen-activated protein (MAP) kinase pathways, which regulate particular cellular processes. Mutations in this gene have been linked to type 2 diabetes mellitus, a heightened risk of insulin resistance, and an increased likelihood of developing a number of different malignancies. The structure and function of IRS1 could be severely compromised as a result of single nucleotide polymorphism (SNP) type genetic variants. In this study, we focused on identification of the most harmful non-synonymous SNPs (nsSNPs) of the IRS1 gene as well as prediction of their structural and functional consequences. Six different algorithms made the initial prediction that 59 of the 1142 IRS1 nsSNPs would have a negative impact on the protein structure. In-depth evaluations detected 26 nsSNPs located inside the functional domains of IRS1. Following that, 16 nsSNPs were identified as more harmful based on conservation profile, hydrophobic interaction, surface accessibility, homology modelling, and inter-atomic interactions. Following an in-depth analysis of protein stability, M249T (rs373826433), I223T (rs1939785175) and V204G (rs1574667052) were identified as three most deleterious SNPs and were subjected to molecular dynamics simulation for further insights. These findings will help us understand the implications for disease susceptibility, cancer progression, and the efficacy of therapeutic development against IRS1 gene mutants. Communicated by Ramaswamy H. Sarma


Introduction
Single Nucleotide Polymorphisms (SNPs) are genetic variations that occur at a single base location in a DNA sequence.These variations play a significant role in the connection of diseases and phenotypic changes across individuals.SNPs account for approximately 90% of an individual's genetic diversity (Dakal et al., 2017).SNPs that alter the amino acids have neutral or deleterious effects on the protein structure and function.This type of SNPs is known as non-synonymous SNPs (nsSNPs).Half of all human disease-related genetic variants have been attributed to nsSNPs (Cargill et al., 1999;Yue & Moult, 2006).Due to the high concentration of genetic variants in the human genome, extensive studies are required to understand how they affect drug development and patient susceptibilities to diseases (Cargill et al., 1999).Protein functions that are altered by the SNPs can interfere with interaction with miRNA, transcription factors, splicing, and translation of mRNA (Alexander et al., 2010;Stenson et al., 2009).Moreover, if SNPs cause amino acid alterations in a conserved region, plausibly that will affect protein function, structure, stability, charge, molecular contacts, hydrophobicity, geometry, molecular dynamics, intra or inter protein interactions, and disease risk severely (Chasman & Adams, 2001;Kucukkal et al., 2015;Petukh et al., 2015).The SNP-related study is also essential for phenotypic characterization.Additionally, it provides us with thorough information on treatments and drug development for diseases linked to a particular amino acid substitution.SNP research is time-consuming, expensive, and labor-intensive; nevertheless, bioinformatics tools can quickly and cheaply identify harmful SNPs.Before conducting mutation analysis in the lab (in vivo or in vitro), this process can be used to identify the most dangerous nsSNPs among all of the SNPs in the database.
The aim of the study was to identify the most deleterious nsSNPs of IRS1 and predict their effect on protein structure, function, treatment responsiveness, and predisposition of diseases.After performing primary nsSNP screening, we depicted their conservation profiles, position in functional domains, molecular interactions, stability analysis of protein structures, and post-translational modifications.Afterward, molecular dynamics (MD) simulation was performed in order to gain a better understanding of the effect that mutations have on physiological conditions.
Here, PolyPhen-2 makes comparison-based predictions about the potential effects of amino acid substitution on protein structure and function.The protein sequence, database ID, and accession number were entered into the server.It classified databases based on probability scores as potentially harmful (scoring.45to.93), probably harmful (score >9.3), and benign (score < 0.4) (Thusberg et al., 2011).
PhD-SNP discriminated between polymorphism that was 'diseased or neutral' by single point mutations of amino acid sequences (Capriotti et al., 2006).PANTHER estimated specific non-synonymous coding SNPs in order to investigate their influence on protein function.The protein sequence of the IRS1 gene and amino acid replacements were used as input data.This technique foresaw whether the outcomes would be harmful or indifferent (Diseased or Neutral).PANTHER also produced a wealth of information about the phylogeny, function, and genetic variations that regulate the function of the protein (Tang & Thomas, 2016).
SNPs & GO technique involved prediction of detrimental single point protein mutation using Predicting disease associated variations using gene ontology terms.The protein sequence (FASTA format) as well as amino acid substitutions were used as input purposes through setting of target variations and functional gene ontology terms.This technique produced results for every protein variant and classified them as either diseased or neutral (Capriotti et al., 2013).
Effects of SNP in various genomic areas were analyzed using Predict SNP2.As an input format, it was distinctive in that it featured the name, number, and single nucleotide variant of individual chromosomes.Approximately 50 queries might be processed simultaneously by this program.By utilizing the various characteristics of genetic variation in different genomic regions, the program accurately evaluated the effects of the SNPs.Predict SNP2 displayed output data in percentage form, with red and green representing harmful and benign outcomes, respectively (Bendl et al., 2016).
SNAP2 predicted the functional effects of mutations on protein by determining the impact (effect) of single amino acid substitution.FASTA format, which was obtained from UniProtKB (http://www.uniprot.org/uniport/),was used as input data and makes predictions about whether the results will be positive or negative (Effect or Neutral) (Bromberg et al., 2008).

Evolutionary conservation analysis
Conserved region of IRS1 was identified by ConSurf (https:// consurf.tau.ac.il/).ConSurf identified conserved regions from amino acid sequences of three dimensional structure of protein.FASTA format was used as input query.The protein structure was divided into three groups, each of which was assigned a conservation score between 1 and 9.With different color schemes for each score, these were variable (1-2), average (3-7), and conserved (8-9) in nature (Glaser et al., 2003).Here, we selected score 8-9 as conserved positions.Moreover, we also explored the conservation profile of rs IDs using Ensembl (https://www.ensembl.org/).Ensembl provided conservation profile of the selected nsSNPs in 10 closely related primates such as Pan paniscus, Pan troglodytes, Gorilla gorilla, Pongo abelii, Nomascus leucogenys, Chlorocebus sabaeu, Macaca fascicularis, Macaca mulatta and Microcebus murinus.

Secondary structure analysis
NetSurfP-2.0 (https://services.healthtech.dtu.dk/service.php?NetSurfP-2.0)predicted the surface accessibility, secondary structure, disorder, and phi (U)/psi (W) dihedral angles in an amino acid sequence.Primary protein sequence was employed as input, and the resulting visualization highlighted accessible regions of the genome in red (exposed) and blue (buried).The thickness of a green line that indicated potential pathogenic residues was assessed by another metric (Klausen et al., 2019).

Analysis of protein stability variation
Protein stability variation was calculated using predictions of protein stability changes upon mutations (MUpro) (http:// mupro.proteomics.ics.uci.edu/).Mutation name, location, original amino acid, and substitution amino acid were chosen as the input for the query.In addition, plain amino acid sequence of the protein was submitted, however headers were removed.Protein structure files may occasionally be used as input queries.Consistency in protein stability was predicted.Support Vector Machine (SVM) and sequence data were used to calculate the DG score that was then applied to estimate the value and sign of the energy shift.This algorithm predicted an increase or decrease in protein stability, depending on the score (Cheng et al., 2006).

HOPE analysis
Have Our Protein Explained (HOPE) (https://www3.cmbi.umcn.nl/hope/)predicted the structural implications of point mutations in a protein sequence.Proteins in FASTA format were given as input data.As it studies the mutant, it may detect inheritable disorders.This tool examined the three dimensional (3D) structure and picture of the target protein, the structure and properties of replaced amino acids, and variations in the conservation profiles of the wild-type and mutant types.The mutations that are thought to be conserved disrupt the interaction between these two protein sequence domains, which in turn affects the protein's functionality (Venselaar et al., 2010).

Domain identification
Most domains are independent functional parts of proteins that can be put together in different ways to make different kinds of proteins.As well as, domains help to understand the protein function.Pfam (Protein family databases) (https://pfam.xfam.org/) was used to choose the IRS1 domains.The Uniprot accession ID was used as the query input.By extracting certain start and end positions from the entire amino acid sequence, this method detected IRS1 domains (Mistry et al., 2021).
InterPro (http://www.ebi.ac.uk/interpro/) examines the functional effects of proteins and predicts the relevant sites and domains.As a query input, the protein sequence is represented in FASTA format.Result indicated protein family membership classifying them into family, domain, homologous superfamily, unintegrated sequences & predictions through coloring them with specific amino acid start & end point (Blum et al., 2021).
Another useful tool is PROSITE (https://prosite.expasy.org/prosite.html),which may be used to identify protein families, functional sites, and domains with patterns and profiles.As input data, UniProt id and FASTA format were provided.Two domains with distinctive amino acid sequences were found.It created unique domain images (Hulo et al., 2006).

Homology modeling
Homology modeling of the mutants were detected using HHpred (http://protevo.eb.tuebingen.mpg.de/hhpred).HHPred finds remote protein homology, which implies hidden Markov model profile comparison (HMMs).The FASTA sequence of the wild type residue and the FASTA sequence of the mutant, in which an amino acid had been substituted, were used as the inputs.Prior to submission, structural and domain databases as well as proteomes were chosen.Then, the hit-list, model selection option was selected.Prior to that, sequences that shared a perfect match with the natural protein submission were chosen.The modeler was subsequently notified, and the entire protein PIR was sent for structure modeling.It could construct pairwise or multiple query-template alignments by selecting a set of templates from the search option, and the software MODELLER created 3D structure models of proteins based on these alignments (S€ oding et al., 2005;Webb & Sali, 2016).
The quality of the protein structure was then validated using SAVES v6.0 (https://saves.mbi.ucla.edu/).To execute the program, an HHpred-downloaded .pdbformat query was provided as input.Then, structure validation was assessed using five programs ERRAT, VARIFY 3D, and PROCHECK.ERRAT is a unique method for distinguishing between correctly and erroneously identified sections of protein structures based on atomic interaction characteristics (Colovos & Yeates, 1993) .VERIFY 3D estimated similarity of an atomic model (3D) with its own amino acid sequence by placing a structural class based on its surroundings and comparing the results to good structures (Bowie et al., 1991;L€ uthy et al., 1992).PROCHECK (Programs to check the Stereochemical Quality of Protein Structures) (http://www.csb.yale.edu/userguides/datamanip/procheck/manual/index.html)provides an aggregated check on protein stereochemistry.This includes postscript charts listing residue by residue.This produced high-quality structures compared to others of the same resolution.Through plot statistics, PROCHECK developed a Ramachandran plot and calculated the quality score of the protein structure (Laskowski et al., 2018).

Structure refinement
The GalaxyRefine which is well recognized for protein structure refinement (http://galaxy.seoklab.org/refine)has re-fined the structural quality of protein by reconstructing side chains and conducting side-chain repacking.This strategy not only enhanced the quality of the local structure but also has the potential to improve the global structure.As inputs, both the wild-type and mutant variants of the protein structures were provided.Five refined models were created by this tool based on MolProbity, clash score, Rama favored scores, and so on.In order to select the most optimally refined model, a lower MolProbity and Clash score in conjunction with a higher Rama favored score were prioritized (Heo et al., 2013).

Three dimensional interatomic interactions analysis
DynaMut-2 (http://biosig.unimelb.edu.au/dynamut2/), a comprehensive tool for flexibility analysis and visualization.This tool predicted the impact of missense mutations on protein dynamics and stability at tertiary level.In order to conduct protein dynamics analysis, the input query needed to be either a file in the .pdbformat.The impact of mutations on protein dynamics and stability could be examined by supplying a specific amino acid substitution together with the position number where that point mutation took place.The location of the wild-type residue along the chain was specified.The performance of protein stability change is estimated by the values of (DDG Stability ).The results showed G � 0 as stabilizing and G < 0 as destabilizing.Following this procedure, mutants with the highest degree of instability were chosen for MD simulation (Rodrigues et al., 2018).

Molecular dynamics simulation
For MD simulation analysis, GROningen MAchine for Chemical Simulations (GROMACS) -version 2020 was utilized.Here, GROMOS96 43a1 force-field was applied for protein energy minimization (EM).After EM, 100 ps NVT simulation and NPT simulation were conducted.Finally, a 100 ns long MD run was conducted.The results of the simulation were analyzed via Root Mean Square Deviation (RMSD), Root Mean Square fluctuation (RMSF), Radius of Gyration (Rg) and Solvent Accessible Surface Area (SASA) data.

Result
The outline of the entire study is given in Figure 2.

Fifty-nine out of 1142 nsSNPs were predicted as risky variants
About six bioinformatics tools such as PolyPhen-2, PhD-SNP, PANTHER, SNP & GO, Predict SNP 2, and SNAP-2 were used to predict the functional impact of most deleterious SNPs (Figure 4).
SNAP2 was also employed to investigate the influence of amino acid substitutions on protein function.This tool supported the results of the five bioinformatics tools by  displaying effect, which indicates they were all determined to be harmful.As a result, the 59 most harmful nsSNPs were evaluated for the following stage of filtering.

Twenty-six out of 59 nsSNPS were located inside functional domains of IRS1
Before the analysis of the structural impact of single point mutations within the protein, domains were selected to filter out most deleterious SNPs.Owing to the fact that domains served as the structural, functional, and evolutionary building blocks of protein.Therefore, the alteration of residues in a domain that might change regular physiology.It is possible that the mutation will mess up the interaction of these two domains, which will have a knock-on effect on the protein function.Domain mutations lead to more severe and disease-causing mutations further down the chain.Hence, protein domains are identified through Pfam (http://pfam.xfam.org/),PROSITE (https://prosite.expasy.org/prosite.html),InterPro (http://www.ebi.ac.uk/interpro/).These tools identified Pleckstrin homology (PH) domain (7-114) and Insulin receptor substrate (IRS) domain (160-262) inside the IRS1 protein (Figure 5).We found 59 nsSNPs filtered through six bioinformatics tools related to functional impact analysis.Here, 26 deleterious nsSNPs were found inside functional domains.

Most of the amino acid substitutions led to the loss of hydrophobic interactions
HOPE (Have Our Protein Explained) (https://www3.cmbi.umcn.nl/hope/)analyzed the impact of point mutations in protein structure.This server made a prediction on the impact of 16 tested mutations on the structure of proteins by comparing the sizes and structures of wild and mutant amino acids, as well as the effects of changes in charges associated with ionic interactions with other molecules (Table 3 and Figure 7; Supplementary file 2).Out of 16 mutants, 14 mutants (M249T, D242N, E235K, R227S, I223T, F221Y, F220I, G215R, R213H, R213G, V204G, K79N, E50K, R20H) might affect the protein structure & two mutants (A246S, C69R) were possibly damaging.Within those 14 mutants which might affect the function of protein, four mutants (E235K, I223T, F221Y, F220I) were located near a highly conserved position.Twelve mutants (�85%) reduced hydrophobic interactions due to amino acid substitutions.

R20H, C69R and G215R mutants alter surface accessibility significantly
NetSurfP-2.0 server predicted secondary structure, surface accessibility, disorders in an amino acid sequence of protein.This showed a comparison of wild & mutants 'Relative Surface Accessibility' (Blue-Buried, Red-Exposed) in percentages and their 'Absolute Surface accessibility' in Å (dihedral angles).This server also performed an analysis on how changes in line thickness equaled the probability of a disordered residue.Surface accessibility and the structural stability of the protein was   exposed (D242N, R227S, G215R, R213H, R213G, K79N, C69R, E50K) to the surface for both wild type and mutant amino acids (Table 4) (Figure 7).R20H mutant altered relative surface accessibility from exposed to buried where C69R and G215R altered it from buried to exposed.

Homology modeling and quality evaluation of wild type and the mutant 3D protein structures
HHpred was used to generate 3D structure of the protein.One IRS1 type and 16 mutant models (M249T, A246S, D242N, E235K, R227S, I223T, F221Y, F220I, G215R, R213H, R213G, V204G, K79N, C69R, E50K, R20H) were generated using the FASTA sequence of the particular protein.SAVES v6.0 (https:// saves.mbi.ucla.edu/)server was used to check the quality of protein structure through ERRAT.Verify 3D measured the compatibility of a 3D atomic model with its own 1D amino acid sequence.Both the wild type and the mutant structures passed the verification by ERRAT and Verify 3D (Table 5).PROCHECK estimated the number of residues in the most favored regions and generated a 'Ramachandran plot' for the wild type model and each of the 16 mutants.

Structure quality refinement
The Galaxy Refine web server produced improved local and global structure quality.Wild type and 16 mutants were submitted as input data.The best refined models were selected from the output based on lower MolProbity and Clash scores as well as higher Rama favored scores (Table 6).

Molecular dynamics (MD) simulations of the mutants and wild type IRS1
The RMSD value of wild type IRS1 and mutants were different through the 100 ns simulation (Figure 8).Specially, after 25 ns the protein backbones demonstrated significant differences (Figure 8a).All mutants showed peak mobility in different regions/domains in dynamic situations that are not similar to the Wild type patterns (Figure 8b).V204G notably altered protein compactness (Rg) and SASA (Figure 9a,b).

Discussion
Insulin Receptor Substrate 1 (IRS1) is related with impaired insulin receptor signaling including phosphatidylinositol 3 0kinase activity which is stimulated by insulin (Figure 1; Almind et al., 1996).Therefore, an association of IRS1 mutation may lead down to insulin resistance.This generates risk  -Garcia et al., 2010).The polymorphisms related to IRS1 are also associated with onward susceptibility to Polycystic Ovary Syndrome (PCOS).As the insulin resistance is one of the major factor behind high degree polycystic ovary syndrome and non-alcoholic fatty liver diseases (Baranova et al., 2011).Genetic polymorphisms regarding receptor substrates which involves Insulin-like growth factor (IGF), and IGF binding protein (IGFBP) have an influence on the association of colorectal cancer as those receptor substrates mediate important roles in cell growth & proliferation (Slattery et al., 2004).IRS proteins act as adaptors which cause the transmission of signals from various receptors.Post translational modification such as phosphorylation of IRS-1 at YXXM motifs located in the outer domain can generate docking sites for PI3Kp85 binding which activate AKT kinase.If any mutant affects the phosphorylation of IRS1, AKT, ERK, and STAT3 induced by insulin, it generates multiple diseases and cancers by affecting the migration of cells, normal glucose, uptake and cell response to radiation (Gibson et al., 2007;Senthil et al., 2002;Vuori & Ruoslahti, 1994).Mutations in Prolactin, leptin, growth hormone, vascular endothelial growth factor (VEGF), tropomyosin receptor kinase B (TrkB), anaplastic lymphoma kinase (ALK), insulin like growth factor (IGF1), and integrins of a binding protein may develop cancer as it causes increased expressions which has already been detected in numerous cell lines (Bergmann et al., 1996;Han et al., 2006;Hoang et al., 2004;Koda et al., 2005;Ravikumar et al., 2007;Rocha et al., 1997;Schnarr et al., 2000).Previous studies have shown that computationally predicted harmful missense mutations are critical for disease development in Galactosemia, Parkinson's disease, Gaucher's disease, dysfibrinogenemia, thrombophilia, hypofibrinogenemia, and Canavan disease (Agrahari et al., 2019;Ali et al., 2017;George Priya Doss & Zayed, 2017;Sneha et al., 2018;Thirumal Kumar et al., 2018;Zaki et al., 2017a).Hence, in this current study, we have explored the deleterious effects of harmful mutations in the IRS1 protein.
Depending on the probability of influence on protein function and structure, deleterious 59 ns-SNPs of IRS1 were detected using six bioinformatics tools.About 26 nsSNPs were sorted from 59 nsSNPs on the basis of domain mutations.They were all found in evolutionarily conserved regions of the protein.Hence, these 26 nsSNPs might hinder the domain related functions of the protein.According to MUpro, 16 of those 26 nsSNPs are highly destabilizing.Moreover, the HOPE server predicted 14 nsSNPs out of the 16 nsSNPS might affect the structure and 2 nsSNPs possibly damaging.
After running an analysis on the primary amino acid sequence, we explored the effects of the nsSNPS on the protein secondary structure.Here we implemented NetSurfP-2 for exploring the effects on secondary structure.NetSurfP-2 identified 9 nsSNPs were buried and 7 were exposed to the surface for wild-type structure however on each side 8 nsSNPs were found buried and exposed to mutant structure depending on changes of relative and absolute surface accessibility.Two of them switched from being buried to being exposed, and one of them switched from being exposed to being buried.This is how amino acid substitutions changed from wild-type to mutants.
The effect of the deleterious nsSNPs also investigated on 3D protein models.HHpred predicted 17 IRS1 structures, which included the "wild type" and 16 "mutants".Quality of the predicted models were decent.Finally, M249T, I223T, and V204G were predicted as destabilizing for the structure according to DynaMut-2, Mcsm, SDM, and DUET.These techniques predicted the changes in protein stability caused by all of the mutations by assessing the delta G scores.In the end, these top three damaging mutants were put through a molecular dynamics simulation in order to move on to the next stage of the process, which was the sorting of the most detrimental ns-SNP of the IRS1 gene.
MD simulation showed that RMSD values for wild type IRS1 and mutants differed after �25 ns.The differences between RMSD values support that the structures were getting dissimilar in dynamic conditions.Moreover, the mobility/fluctuation peaks in different regions/domains that were not similar to the wild type mobility indicated the restrictions of usual IRS1 functions (Figure 8b).Interestingly, V204G significantly altered protein compactness (Rg) and SASA (Figure 9a,b).V204G also change the RSA and ASA values of the protein (Table 4, Supplementary file 3), hence, Figure 9b showed less SASA.None of the mutations have been explored in the patients with IRS1 related disorders, notably the pathogenic variants M249T (rs373826433), I223T (rs1939785175), and V204G (rs1574667052).A potentially significant genotype-phenotype correlation will highlight the development and application of this computational technique for future clinical assessments of genetic variants (Zaki et al., 2017b).
Finally, the findings of our study have established that some of the deleterious IRS1 nsSNPs could be linked to diseases as a result of their influence on protein structures.As a result, the above-mentioned mutants can be explored in different populations for IRS1-related disorders such as insulin resistance, different carcinomas, lung cancer, polycystic ovarian syndrome, and type 2 diabetes mellitus via simple PCR or sequencing-based diagnostics.Further in vitro research can include the impacts of these mutants on exogenous insulin therapy and predict the potential genotype-phenotype connections of these variants related to diseases.Changes in cell signaling patterns due to these mutations will also decipher their contribution to oncogenesis.Additionally, this study opens the larger-scale in vivo research options in the future to develop effective IRS1 inhibitors for treating osteosarcoma (Garofalo et al., 2015).

Conclusion
This study evaluated potentially harmful SNP variants of human IRS1.Our analyses suggested that M249T, I223T, and V204G mutants of IRS1 may promote insulin resistance or other IRS1 related disorders.These mutants can be focused in subsequent Genome Wide Association Studies (GWASs).
Further in vitro and in vivo explorations might develop new therapeutics and potential biomarkers.

Figure 2 .
Figure 2. Schematic representation of the entire study.

Figure 3 .
Figure 3. Distribution of SNPs of IRS1 gene into different classes as obtained from the NCBI dbSNP database.

Figure 5 .
Figure 5. Location of deleterious non-synonymous SNPs inside IRS1 functional domains.Five nsSNPs were located inside the Pleckstrin homology (PH) domain whereas 21 nsSNPs were situated inside IRS-type Phosphotyrosine-binding (PTB) domain.

Figure 6 .
Figure 6.Conservation profile of IRS1 amino acid residues.These mutants were located in highly conserved positions.

Figure 7 .
Figure 7. Changes of Inter-atomic interactions in the wild type and highly 9 deleterious mutants predicted by DynaMut-2, DUET, m-CSM, SDM.

Figure 8 .
Figure 8.The (a) Root mean square deviation (RMSD) and (b) Root mean square fluctuation (RMSF) graphs of the native and mutant IRS1 structures in dynamic conditions.Here, the color scheme are wild (blue), I223T (Yellow), M249T (Green) and V204G (Pink).V204G, I223T and M249T showed at least 2 major peaks or enhanced mobility of several regions in the IRS1 protein,.

Table 3 .
Comparison of wild type and the effect of mutations on protein structure, size, ionic interactions, and hydrophobicity interaction.
factors such as obesity, hypertension, and hypertriglyceridemia with diseases including Non-Insulin Dependent Diabetes

Table 4 .
Comparison of Relative Surface Accessibility (RSA) and Absolute Surface Accessibility (ASA) between wild type and the mutants.

Table 6 .
Selected models and their refinement scores.