The structure-based cancer-related single amino acid variation prediction

Single amino acid variation (SAV) is an amino acid substitution of the protein sequence that can potentially influence the entire protein structure or function, as well as its binding affinity. Protein destabilization is related to diseases, including several cancers, although using traditional experiments to clarify the relationship between SAVs and cancer uses much time and resources. Some SAV prediction methods use computational approaches, with most predicting SAV-induced changes in protein stability. In this investigation, all SAV characteristics generated from protein sequences, structures and the microenvironment were converted into feature vectors and fed into an integrated predicting system using a support vector machine and genetic algorithm. Critical features were used to estimate the relationship between their properties and cancers caused by SAVs. We describe how we developed a prediction system based on protein sequences and structure that is capable of distinguishing if the SAV is related to cancer or not. The five-fold cross-validation performance of our system is 89.73% for the accuracy, 0.74 for the Matthews correlation coefficient, and 0.81 for the F1 score. We have built an online prediction server, CanSavPre (http://bioinfo.cmu.edu.tw/CanSavPre/), which is expected to become a useful, practical tool for cancer research and precision medicine.


Scientific Reports
| (2021) 11:13599 | https://doi.org/10.1038/s41598-021-92793-w www.nature.com/scientificreports/ tumorigenesis without a systems network biology framework 18,19 . At the proteome level, amino acid substitutions caused by genetic codon transitions may explain the basis of human cancer 20 . Amino acid alterations appear to follow certain rules. For example, arginine has a positive charge that is important for balancing protein and DNA binding; however, arginine is highly mutated in various cancer types. The loss of arginine frequently influences the function of cancer-associated proteins, whereas gaining cysteine, an active and reducing agent, may enhance the capacity of a protein to neutralize reactive oxygen species (ROS) in the tumor environment [21][22][23] . Proteomic changes caused by proteins carrying missense mutations may help cancer cells adapt to environmental pressure 24 . While different types of cancers have unique properties that are not shared among all cancers, these different cancers may share some substitution patterns 25 . For instance, the amino acid substitution spectrum is similar in breast and digestive tract cancers, and is dominated by the alteration of glutamic acid to lysine 25 . Clarification of the relationship between SAVs and cancers using traditional experiments necessitates much time and many resources, so computational prediction methods are sorely needed in cancer biology research. Machine learning has become a favored tool for data analysis, as this offers the capacity for leveraging big data and for analyzing the content of complex problems, clarifying information and content 26,27 . Many predictors have emerged that use machine learning as algorithms for SAVs. The two most important and commonly used categories of development strategies are the genetic-based and protein-based prediction systems. Several large-scale sequencing projects such as the TCGA project are widely utilized for genetic analysis 28,29 , some of which follow the American College of Medical Genetics and Genomics (ACMG) guideline 30 . The multifactorial variant prediction (MVP) is a genetic-specific multifactorial model that integrates 16 in silico predictors with the available clinical evidence 31 . However, while the genetic-based system has established a causal relationship between genes and cancer, this relationship is limited to cause and outcome. A comprehensive assessment of cancer driver mutation prediction models evaluated 33 commonly used prediction tools 32 , the top three (CHASM, CTAT-cancer and DEOGEN2) of which are cancer-or protein-based systems [33][34][35] ; CHASM and CTAT-cancer are designed to incorporate the cancer protein, while DEOGEN2 predicts the deleteriousness of SAV by training the mutations in human inherited disease. Since the protein molecule is intimately involved in cellular processes, using the protein-based system to build a prediction system might be more informative for complex diseases such as cancer. However, although numerous prediction models have been developed, we lack the expertise for constructing accurate prediction tools for cancer and knowing which SAV descriptors to incorporate 36 .
We describe our development of a prediction model, which is capable of recognizing whether a particular SAV is cancer-related or is neutral. This model is not only able to discriminate physical changes for each SAV regarding protein function and structure, but it can also estimate how these changes contribute to cancer progression. We hypothesized that accounting for every kind of SAV might be a vital feature for cancer, so we designed a system that incorporates multiple prediction models and enables the user to extract critical features from cancer-related proteins. Our model provides a novel way forward for cancer research, not only for clinical outcomes but also for recognizing prognostic biomarkers, which we contend is a breakthrough for precision medicine.

Materials and methods
Dataset of SAVs. All SAV data were collected from CanProVar 2.0 37,38 , a human Cancer Proteome Variation database that stores both germline and somatic amino acid variations, including those related to the genesis or development of human cancers based on six sources, including the public databases HPI 39 , COSMIC 40 , OMIM 41 , and TCGA 42 , as well as two large-scale cancer genome resequencing studies 43,44 . CanProVar 2.0 contains 156,671 cancer-related SAVs from mutations that have been reported in cancer samples and 967,017 neutral SAVs from validated coding SNPs in the dbSNP database. In order to determine the exact protein structure of the SAV sequence, CanProVar 2.0 data were mapped to the proteins identified by BLAST 45 from the Protein Data Bank. The search used six criteria, as follows: 1. The e-value of alignment results should be smaller than 1e−50; 2. The sequence identity of alignment results should be greater than 80%; 3. The alignment coverage of the protein structure should exceed 95%; 4. The organism of the aligned target protein should be Homo sapiens; 5. The experimental method ideally uses X-ray diffraction to extract the protein structure for aligning the target protein structures; and 6. The SAV position should be equidistant between the wild-type in the SAV sequence and the aligned target protein. After matching a protein with a 3-dimensional structure, CD-HIT 46 was used to filter out homologous proteins. The CD-HIT cluster algorithm generates sets of protein families and uses the sequence identity cut-off of 0.3 to purify redundant proteins. Subsequently, the remaining 2,867 cancer-related SAVs and 7,562 neutral SAVs were our main training set, which were separated into 20 groups by using the representative wild-type amino acid of SAV. The numbers of cancer-related and neutral SAVs for each wild-type amino acid in the training set are listed in Table 1; δ is the cancer:neutral SAV ratio, which ranges from 0.2245 to 0.6693.
Two independent sets were built, independent sets 30 and 40, which collected the proteins filtered out by CD-HIT in the previous step. Any proteins in the independent sets sharing more than 30 or 40% sequence identity with another protein in the same wild-type amino acid group of a training set were filtered out. The final analysis included 154 cancer-related SAVs and 2,240 neutral SAVs in the independent set 30, and 322 cancer-related SAVs and 3,127 neutral SAVs in the independent set 40. Table 1 lists the numbers of cancer-related and neutral SAVs for each wild-type amino acid in the independent sets 30 and 40. Typically, the 30% sequence identity was usually used as the criteria to remove homologous proteins. However, due to few cancer-related SAVs were collected in the independent set 30, we provide the extra independent set 40 for comparison. www.nature.com/scientificreports/ were combined with other subgroups that had substitution scores exceeding zero based on BLOSUM62. Any subgroups that did not fit these criteria were omitted from the CanSavPre wm , construction and were instead included in the CanSavPre w prediction models. The dataset was split into subgroups to determine the characteristics of specific wild-type amino acid alterations in each of the sequence-based, structure-based and microenvironment-based feature sets. As an illustration, a glycine would be built into prediction models containing for example acidic (e.g., aspartic or glutamic acid) and basic (e.g., arginine) mutated amino acids with distinct SAV features. Thus, our feature selection process effectively detects essential features. The second prediction system ultimately yielded 100 prediction models (see Supplementary Table S1). Each prediction model was a two-level Support Vector Machine (SVM) classifier module. The SVM machine learning method is widely used for classifying protein structure or function in computational biology [47][48][49][50][51][52] . All SVM calculations were performed using LIBSVM (version 3.24) 53,54 , with the radial basis function (RBF) kernel. The first-level SVM comprised 12 SVM classifiers based on four repeats for three feature sets, which are described in the next section. All SAV descriptors in each feature set were fed into the SVM, and a five-fold cross-validation was performed during model training. The parameters (penalty and gamma values of the RBF kernel) were both trained by exponentially increasing the grid search from 2 -15 to 2 15 incorporating best values of informative measures.
The genetic algorithm (GA) 55,56 was used to select features and optimize performance. The basic GA procedures are as follows: N solutions ( S i , i = 1, . . . , N ) are randomly generated as the starting population. Each solution S i is represented as a set of vectors S i = i . The feature vector i is an m-dimensional vector, indicating the binary representations of m features: If f i j = 1 , the jth feature is kept; if f i j = 0 , the feature j th is eliminated. In order to avoid any imbalance between positives and negatives in performance, four informative measures (Eqs. 1-4) for prediction performance were used as the fitness functions, consisting of accuracy (Acc), the Matthews correlation coefficient (MCC), the F1 score (F1), and summation of sensitivity and weighted specificity (Hybrid). They were calculated as follows: where Precision = TP TP+FP , Sensitivity = TP TP+FN , Specificity = TN TN+FP , TP represents true-positives, TN represents true-negatives, FP represents false-positives, FN represents false-negatives and δ is the ratio of the number of cancer-related SAVs to neutral SAVs, which are listed in Table 1.
In the initial population, N solutions are randomly divided into two halves. α and β have the best fitness in each half, and they are defined as α = (� α ) = max S 1 , . . . , S N/2 and β = � β = max S N 2 +1 , . . . , S N . In general, the three basic mechanisms driving the evolutionary processes in one generation consist of the selection, mutation and crossover processes.
Selection operator. In the τ th generation, the selection operators are defined as: Note that for the special case of τ = 0 , α 0 and β 0 are defined as 0 Mutation operator. We apply two types of mutation to the N solutions S i s. In the case of i = 1, . . . , N/2 , every b bit of the vectors is subject to mutation: b =∼ b , if the mutation rate is less than a mutation threshold µ 0 = 0.1 .
In the case of i = N 2 + 1, . . . , N , we randomly choose a bit from the vectors. These bits are then subject to mutation without any mutation thresholds.
Crossover operators. The crossover operations are carried out between S 2p−1 and S 2p , where p = 1, . . . , N/2 and proceed as follows: one-point crossover is performed between 2p−1 and 2p if the crossover rate is less than the crossover threshold µ 1 = 0.5.
The second level of SVM classifiers is used to process the prediction results generated from 12 classifiers in the first level, to produce the final probability distribution of the relationship with cancer-related SAVs or neutral SAVs. The relationship with the largest probability is used as the final prediction. The two-level SVM system is shown schematically in Fig. 1.
Classification of feature sets. SAV descriptors for machine learning were classified into three classes: sequence-based, structure-based, and microenvironment-based feature sets. For the sequence-based feature set, 44 descriptors were extracted from the protein sequence and approximately partitioned into three categories. The first category contained the most commonly used substitution index of wild-type SAV residues to mutations. Three kinds of substitution indices were used; the BLOSUM62 57,58 , PAM250 59 , and the position-specific scoring matrix (PSSM), which was derived from PSI-BLAST 60 . The second category represented the conservation for each residue compared with homologs. The 15 evolutional entropy values derived from PSI-BLAST were used to denote a sliding window containing 7 amino acids on either side of the SAV. The model also calculated average entropy values for window lengths 5 and 15, centered on the SAV, representing local-and wide-ranging sequence conservation, respectively. In the third category, an amino acid composition (AAC) 61 , a 15-residue peptide (with 7 amino acids on either side of the SAV), represented the composition of the neighboring residues. According to the physicochemical properties of residues, we used the following classification schemes 62 Table 2.
The structure-based feature sets contained 13 descriptors extracted from PDB and DSSP 63,64 . The first structure-based descriptor used the B-factor value of the SAV Cα atom; the B-factor value represents those atoms displaced from their mean positions in a crystal structure diminishes the scattered X-ray intensity. This displacement may be the result of temperature-dependent atomic vibrations, or static disorder in a crystal lattice. Our model also uses critical DSSP information regarding solvent accessibility and includes eight DSSP-defined elements in the secondary structure (i.e., H, B, E, G, I, T, S, and others), energy from the acceptor and donor backbone hydrogen bonds, and determines whether or not disulfide bonding exists. These structure-based descriptors are summarized in Table 3.
In the third feature set, the weighted contact number (WCN) model 65 was used to describe the microenvironment properties of SAVs. This weighted contact number model has a local packing density profile, and research has reported a high correlation between the WCN profile and the sequence conservation profile 66  , where r ij was the distance between the atom i and atom j , while N was the number of calculated atoms. In this work, atom i was defined as the Cα atom of SAV, and the different microenvironment properties were represented by calculated different atom types, or the source of atom (4) Hybrid = Sensitivity + δ × Specificity , www.nature.com/scientificreports/ j . The atom type of j could be Cα atoms, nitrogen atoms or oxygen atoms of an amino acid. Moreover, the source of atom j could be located within the same protein chain as SAV or the whole protein, representative of SAV packing density. Alternatively, the source could be derived from another protein chain or molecules such as DNA, RNA, ligands, or metal ions representing protein-protein or protein-molecule interactions. The packing density of SAV may be divided into different classifications representing the microenvironment properties wherein the SAV is located, such as polar, hydrophobic, acidic or basic, according to the physicochemical properties of residues containing the Cα atom j. The same classification schemes were used as described in the sequencebased feature set. The microenvironment-based descriptors are listed in Table 4.

Results
Performance evaluation. Table 5 compares the five-fold cross-validation performances of two prediction systems based on three different feature sets with the prediction performance of the second-level SVM, all of which are optimized by using MCC as the fitness function in GA. In our experiment, the individual prediction model using the sequence-based feature scheme outperformed the other two, while the performance of the model using the microenvironment-based feature was superior to that of the structure-based feature scheme. The outstanding performance of the combined model obtained in the second-level SVM procedure demonstrates that further information is very helpful for understanding and determining cancer-related factors. www.nature.com/scientificreports/ Critically, CanSavPre wm performed better than CanSavPre w in all three individual feature sets and also in the combined set. Specific training and predicting models were built from the specific subgroups according to the wild-type and mutated types of SAV. Using a two-level SVM combining sequence-, structure-and microenvironment-based features, CanSavPre w distinguished between SAVs that were or were not related to cancer, with an accuracy of 79.83%, a Matthews correlation coefficient of 0.45, and F1 score of 0.54. CanSavPre wm is more effective, with an accuracy of 89.73%, a Matthews correlation coefficient of 0.74, and F1 score of 0.81. The fivefold cross-validation performance for each wild-type SAV of the CanSavPre wm system is illustrated in Table 6. Figure 2 illustrates the ROC curve and compares the AUC values of each wild-type SAV in two systems.
Case study: PI3K. The phosphatidylinositol-3-kinase (PI3K) signaling pathway contributes to several cellular processes, including metabolism, proliferation, differentiation and activation. Notably, the PI3K/AKT/mammalian target of rapamycin (mTOR) signaling pathway is one of the most important intracellular pathways and is also one of the most frequently dysregulated pathways in human cancers [67][68][69][70] . Several catalysis subunits exist for PI3K. Those that are encoded by PIK3cδ have been found to induce cell proliferation in colorectal cancer and other types of cancers 71,72 . The amino acid mutation of PI3Kcδ is closely related to oncogenic transformation, and numerous SAVs have been recorded as cancer-related in the COSMIC database, including P57S, Q75K, K111E, P134L, S361F, N380H, L634F, H677R, E713K, A723V, I776T, G890R, and L977I. Figure 3 illustrates the protein structures of PI3K and the p85 α complex (PDB ID: 5DXU) 73 ; 14 amino acids, including one neutral and 13 cancer-related SAVs, are drawn as spheres. All SAVs, except H677R, are correctly predicted by our prediction system. It should be noted that another SAV, R104C, has been marked as a neutral SAV and is also predicted correctly. The predicted results of PI3K are listed in Table 7.   75 , monocytes 76 , follicular dendritic cells 77,78 , and bone marrow stromal cells 79 . Several stimuli regulate the expression of CD23, a critical factor for B-cell activation, growth, and IgE production (OMIM#151445). The D227Y mutation arising from an alteration of the FCER2 gene has been reported in head and neck squamous cell carcinoma (HNSCC) 80 and the colorectal neuroendocrine carcinoma mutational analyses project 81 . D227 is located in one of the conserved double-loops; the interface between CD23 and the carbohydrate protein, Fcε 3-4. Importantly, calcium (Ca 2+ ) is a regulated ligand for CD23 binding affinity and Ca 2+ binding enables loop1 and loop4 to change the conformation and increase the binding affinity. D227 (loop1) and D258 (loop4) form additional salt bridges between CD23 and Fcε 3-4 82,83 . Other bounds are involved in CD23 and Fcε 3-4 binding, while D227Y affects the binding affinity and IgE antitumor functioning (Fig. 4). The boxplots of microenvironment descriptors in Fig. 4 depict the subgroup with the ASP to the TYR mutation. In several descriptors in Fig. 5, the distribution of cancer-related SAVs differs significantly from that of the neutral SAVs, with a 95% confidence interval by two sample z-test (z score > 1.96, p < 0.05). The z score is defined by z = (x1−x2)−(µ1−µ2) where (x 1 − x 2 ) and (µ 1 − µ 2 ) are the observed and expected differences between cancerrelated and neutral SAVs, respectively. σ 1 and n 1 are the standard error and amount for the cancer-related SAVs group, and σ 2 and n 2 are for the neutral SAVs group. The cancer-related SAVs are located in the relatively low packing density region, encompassing Cα atoms, nitrogen, or oxygen located in a single SAV chain, or in whole protein. D227Y in CD23 has a low WCN value in a single SAV chain but a relatively high WCN value in whole protein or other chains (Fig. 5a-c), because D227 is located in the interface of CD23 and This unique surrounding pattern is also found in the case of D227Y in CD23 (Fig. 5d-i). Table 4. List of descriptors in the microenvironment-based feature set.      [84][85][86] . Although most research into the CASQ has focused on cardiac muscle, CASQ in the Ca 2+ signaling pathway is also vital in cancer research 87 , as this pathway is highly correlated with tumor growth and metastasis 88 . Importantly, T189, E194 and D196 can form a pack in the CASQ that harbors Ca 2+89 ; this Ca 2+ binding can be destroyed by the substitution E194G, causing the protein to lose its functionality (Fig. 6).
In the subgroup in which GLU mutates to GLY, none of the microenvironment descriptors of cancer-related or neutral SAVs reveal significant differences at the 95% confidence interval, although several relevant descriptors are found in the case of E194G in the CASQ. That is, E194 exhibits higher WCN values of oxygen in a single SAV chain, as well as higher WCN values of atoms in other molecules, due to the fact that the CASQ is  www.nature.com/scientificreports/ GLU-and ASP-rich, as well as a Ca 2+ buffering protein (Fig. 7a,b). Furthermore, higher WCN values were found in the microenvironment around E194 than in the third quartile of cancer-related SAVs in H-polar (RKEDQN), V-medium (NVEQIL), Z-low polarizability (GASDT), P-high polarity (HQRKNED), F-acidic (DE), and E-basic (HKR) descriptors and were lower than WCN values in the first quartile in E-sulfur-containing (CM) amino acid. Figure 7 depicts the microenvironment descriptor boxplots in the subgroup in which GLU mutates to GLY.

Discussion
Many cancer-related tools are based on genetic or protein sequence information, because of limited protein structure information. Next-generation sequencing technology has led to the large-scale application of genomic information in cancer research and human health [90][91][92][93] . The crucial limitation of this information is that while the technology may determine tumor risk or recurrence, cause and effect remain undetermined. The launch of the Human Proteome Project (HPP) has enriched our understanding of the human proteome blueprint responsible for complex diseases, including cancer. Approximately 20,000 human cancer-related proteome studies have been recorded in PubMed since 2011 94 . However, it remains very difficult to construct a 3-dimensional protein structure from these proteins, which explains why the widely used cancer prediction tools incorporate sequences of genetic information or sequence information of proteins. Importantly, the spatial conformation of protein forms the functional unit.
In this study, we describe how we developed a protein structure-based system, CanSavPre, to predict cancerrelated single amino acid variations. Protein sequence and structure descriptors are used in the model training process. Our prediction system displays excellent performance, and its structural and microenvironmental properties enable us to observe mutating amino acids that generate malfunctioning proteins. Critical descriptors emerge through use of the feature selection procedure. Figure 8 shows the heatmap of selected features in each training group of two prediction systems. For each descriptor, the color represents how many times were selected during four training repeats with different fitness functions. Mutated residues and their functional impact can be characterized by analyzing the selected feature sets. Although further study is needed to reveal the cancer mechanism in most selected features, our results indicate that it is possible to reliably predict cancer-related SAVs.
We also found that it is essential to divide the training data into proper subsets, according to the wild-type and mutated SAVs. Each amino acid with its unique characteristics plays a different role in protein construction. The heatmap in Fig. 8 also indicates that the profiles of selected features in CanSavPre wm are more legible than those in CanSavPre w . The resulting discrepancies influence changes in protein conformation activity, to different degrees. The data from our two-level prediction system optimize the outcomes from a system that uses only www.nature.com/scientificreports/ one level. However, the complexities of protein function in the extensive cellular networks necessitate critical information that identifies cancer-related SAVs. The SAVs in the independent sets 30 and 40 were predicted using the CanSavPre wm system and the prediction performance is illustrated in Table 8. Although a difference in prediction performance is evident between the independent sets and the training set, this is largely because the performance is optimized by a genetic algorithm in the training set. Our purpose was to extract significant features using an optimization procedure. Thus, performance in the training set is the upper bound of prediction. Another reason for the difference in prediction performance between sets might be too few cancer-related SAVs in the training and independent sets. In the training set, the average amount of cancer-related SAVs in the 100 subgroups of the CanSavPre wm system is less Red-framed boxes indicate that two sample z-testing revealed a significant difference between cancerrelated and neutral SAVs, with a 95% confidence interval (z score > 1.96, p < 0.05). Labels of descriptors selected by the genetic algorithm are bolded in the x-axis. Star symbols denote the D227Y cases in CD23. www.nature.com/scientificreports/ than 30; the average amount is much smaller in the independent sets. The performance of independent sets might therefore be distorted, necessitating a greater number of cancer-related SAVs. Nevertheless, compared with DEOGEN2, which is mainly developed for deleterious amino acid variant prediction and publicly available, our prediction system performs better than DEOGEN2 in both two independent sets (Table 8).
Proteins are dynamic molecules with distinct, flexible structures that facilitate allosteric interactions between small molecules or proteins. For our prediction system, we focused on these interactions, i.e., physical contacts between proteins and molecules. A limitation of the crystal-protein structure is that it is capable of revealing only one condition of the protein-protein interaction. The critical purpose of a protein complex is to reflect interactions between single chain proteins. These interactions typically represent functional properties and are expected to be maintained through the genetic algorithm for feature selection. Importantly, technical limitations of the crystal structure prevent the formation of a protein complex in some conditions. Thus, some protein structures only include a part of the complex. These single chain proteins might restrict the range of our prediction system, because their conformations may differ from those of the protein complex. Our research seeks to define the phenomenon underlying cancer-related variations and how differences in conformations influence protein interactions, despite the limited available data on the protein complex.
Our case studies have used several different proteins to illustrate our protein structure-based system. In the first example, the PI3K protein family is well recognized for its association with cancer. The strict rule for our data extraction means that CD-HIT, the cluster database, filters out the homologous proteins and avoids weighting the mutation residues for the training model. Thus, the isoform PI3Kcδ is selected as the representative isoform, even though PI3Kca is one of the most readily recognized isoforms in cancer research 95,96 . The most frequent and pathogenic mutation residues recorded in the TCGA, E542K and E545K, are predicted correctly in our system (see Supplementary Fig. S1). The evidence suggests that these two hotspot mutations induce glycolysis in cervical cancer cells via the β-catenin/SIRT3 signaling pathway 97 . Both glutamic acid mutations (E542K and E545K) are located in the helical domain. Biochemical studies have demonstrated that these mutated residues interact with p85a, so the alteration may affect the inhibitory activity of p85a 98 .
Our prediction system also provides observations of the microenvironment for the SAV residues. The other case studies discussed in our results illustrate how the feature processes cope with the descriptors such as, for example, how the feature selection process manages ASP alteration patterns. This process also yields critical detail regarding ASP alterations (Fig. 4). ASP is a polar amino acid with a negative charge and the carboxyl group in the side chain of aspartic acid allows it to accept hydrogen atoms. Our model calculates the area surrounding the ASP. When the area surrounding the cancer-related ASP is unoccupied space, our model observes fewer basic amino acids compared with neutral ASP environments. This implies that, in this situation, the ASP might assume an interaction role with protein, whereby an ASP mutation can lead to an abnormal protein and different binding activity. This might explain how cellular changes result in tumor activity, as is illustrated in our second case study. In contrast, GLU is an amino acid with high polar residues and a negative charge that is also involved in the hydrogen atom acceptor role, with a distinctly different pattern to that of ASP (Fig. 6). Cancer-related GLUs are located in a denser region within a hydrophobic environment. Although ASP and GLU share similar characteristics, their different microenvironments justify construction of the prediction modules.  89 in the cyan-colored cartoon is drawn by PyMOL. All of the yellow bubbles are Ca 2+ in CASQ. Three Ca 2+ binding residues are highlighted with sticks in deep pink. E194G is a cancer-related SAV. www.nature.com/scientificreports/ We suggest that protein structure and microenvironment features are the cornerstones of cancer research. Slight alterations in protein conformation properties may change the functional activity of molecular interactions. Clarifying protein structure is critical to understanding how a protein mutation can lead to cancer. However, the critical necessity for the protein structures is also the limitation of our prediction systems. Many emerging computational methods are attempting to clarify protein structure. For those proteins that only have sequence information, scientists can use a homology modeling method or an artificial intelligence system to extract the details needed for structural information. Notably, AlphaFold2 is considered to be an excellent solution to the problematic issue of protein folding 99 . As more protein structures become available over time, our model will benefit from the enriched databases, regardless of whether the data are sourced from experimental or predicted methods. www.nature.com/scientificreports/ In conclusion, we have developed a structure-based cancer-related single amino acid variation prediction system. This system not only displays excellent performance, but it also observes how the amino acid substitution influences protein activities. The descriptors provided by our system may offer targets for further research. Moreover, performance is markedly enhanced by the fact that the model includes conformation properties and details of the microenvironments surrounding SAVs. Furthermore, our algorithm detects the best combination of feature vectors for examining specific amino acid variations. Importantly, our model is a user-friendly web-based tool that scientists will find extremely useful when performing cancer research and precision medicine, particularly when investigating rare tumor mutations. The many genetic mutations in human cancers offer us numerous targets and possibilities that may be incorporated into our model, emphasizing its importance in cancer research.