Differentiating between cancer and normal tissue samples using multi-hit combinations of genetic mutations

Dash, Sajal; Kinney, Nicholas A.; Varghese, Robin T.; Garner, Harold R.; Feng, Wu-chun; Anandakrishnan, Ramu

doi:10.1038/s41598-018-37835-6

Differentiating between cancer and normal tissue samples using multi-hit combinations of genetic mutations

Article
Open access
Published: 30 January 2019

Volume 9, article number 1005, (2019)
Cite this article

Download PDF

You have full access to this open access article

Scientific Reports

Differentiating between cancer and normal tissue samples using multi-hit combinations of genetic mutations

Download PDF

Sajal Dash¹,
Nicholas A. Kinney^2,4,
Robin T. Varghese^2,4,
Harold R. Garner^2,4,
Wu-chun Feng^1,3 &
…
Ramu Anandakrishnan ORCID: orcid.org/0000-0003-0422-3984^2,4

12k Accesses
16 Citations
16 Altmetric
Explore all metrics

An Author Correction to this article was published on 09 December 2019

This article has been updated

Abstract

Cancer is known to result from a combination of a small number of genetic defects. However, the specific combinations of mutations responsible for the vast majority of cancers have not been identified. Current computational approaches focus on identifying driver genes and mutations. Although individually these mutations can increase the risk of cancer they do not result in cancer without additional mutations. We present a fundamentally different approach for identifying the cause of individual instances of cancer: we search for combinations of genes with carcinogenic mutations (multi-hit combinations) instead of individual driver genes or mutations. We developed an algorithm that identified a set of multi-hit combinations that differentiate between tumor and normal tissue samples with 91% sensitivity (95% Confidence Interval (CI) = 89–92%) and 93% specificity (95% CI = 91–94%) on average for seventeen cancer types. We then present an approach based on mutational profile that can be used to distinguish between driver and passenger mutations within these genes. These combinations, with experimental validation, can aid in better diagnosis, provide insights into the etiology of cancer, and provide a rational basis for designing targeted combination therapies.

CoMEt: a statistical approach to identify combinations of mutually exclusive alterations in cancer

Article Open access 08 August 2015

Cancer subtype identification using somatic mutation data

Article Open access 16 May 2018

Cancer classification in the genomic era: five contemporary problems

Article Open access 19 October 2015

Introduction

Experimental studies and mathematical models suggest that carcinogenesis is likely a result of different combinations of a small number of carcinogenic mutations (hits)^{1,2,3,4,5,6,7}. Mathematical models estimate that the number of such hits varies from two to eight^{1,2,3,4,5,6,7,8}. Yet, our collective computational and experimental efforts and the accumulation of cancer genomic data have failed to identify, for most cancers, the specific combinations of mutations triggering carcinogenesis.

Current computational efforts to find carcinogenic mutations generally focus on identifying individual “driver mutations”, based on mutational frequency and signatures^9,10,11,12. These driver mutations have been shown to be associated with an increased risk of cancer. However, they can not generally cause cancer by themselves. For example, 72% of women with an inherited BRCA1 mutation are likely to get cancer by age 80. However, even for women with the BRCA1 mutation, none are likely to get cancer before age 20, and 28% of them may never get cancer¹³. The Li Fraumeni syndrome is another example where germline P53 mutations is associated with early onset cancer predisposition (e.g. soft tissue and bone sarcomas). However, cancer penetrance is less than 20% for children while approaching 80% by age 70, indicating that multiple hits are required for carcinogenesis^14,15,16,17. The relationship between most other known genetic markers and increased cancer risk is far weaker^18,19. The limited early cancer incidence in individuals with germline mutations suggests that additional genetic defects acquired over an individual’s lifetime are necessary for carcinogenesis. Therefore, current computational approaches focused on identifying individual genes that are cancer drivers, cannot find the specific combinations of mutations responsible for individual instances of cancer. Several factors, other than genetic mutations, have also been implicated in carcinogenesis, such as epigenetic modifications²⁰, tumor environment²¹, and adaptive evolution²². However, carcinogenesis is primarily a result of genetic mutations²³.

The goal of this work is to develop a method for identifying combinations of genetic mutations that are most likely responsible for individual instances of cancer. This goal is fundamentally different from identifying the most frequent driver mutations, and represents the first computational study to specifically identify multi-hit combinations. Our approach consists of first identifying likely combinations of genes with carcinogenic mutations. We then present a method, based on the mutational profile of these genes, for identifying likely carcinogenic mutations within these genes. Although it is theoretically possible to search for combinations of individual mutations using our method, the problem becomes computationally intractable, since most genes contain hundreds of somatic mutations. In addition, in the much larger set of somatic mutation combinations many carcinogenic combinations will be rarely represented, further increasing the challenge of identifying these combinations. Therefore, we chose to first identify combinations of genes with somatic mutations, and then present an approach for identifying likely carcinogenic mutations within these genes.

We mapped the problem of finding these combinations to the extensively studied weighted set cover (WSC) problem²⁴. Finding the optimal solution to the corresponding WSC problem is computationally intractable due to the exponentially large number of possible sets of multi-hit combinations. However, there exist approximation algorithms for finding near-optimal solutions^24,25. We adapted one such algorithm to find a set of multi-hit combinations that maximizes the number of tumor samples that contain one of these multi-hit combinations while minimizing the number of normal samples that contain any of these combinations. The number of candidate set covers is an exponentially large quantity due to the large number of possible combinations. We applied the above algorithm to find a set of 2-hit combinations using somatic mutation data from the cancer genome atlas (TCGA). For the 17 cancer types with at least 200 matched tumor and blood-derived normal samples in TCGA, the algorithm identified a set of 197 2-hit combinations. For a separate set of Test samples, these combinations were able to differentiate between tumor and normal samples with 91% sensitivity (95% Confidence Interval (CI) = 89–92%) and 93% specificity (95% CI = 91–94%) on average, for the 17 cancer types. The results are consistent across different randomly selected Training and Test sets. Despite this high accuracy, our analysis of the results shows that many of the 2-hit combinations are likely to be two-gene subsets of three or more-gene combinations. We discuss how carcinogenic and non-carcinogenic mutations within the gene combinations can be distinguished. We also discuss how the multi-hit combinations can be used to develop targeted combination therapy.

Identifying gene combinations is important for two reasons. First, it brings us closer to the understanding of carcinogenesis and the complexity of cancer biology. Second, the identification of the specific combination responsible for a given instance of cancer can help us design more effective combination therapies for treating the disease. Combination therapies can be more effective than single target treatments; however, most current therapeutic combinations have been based on trial and error^26,27. Identifying the precise combination of genomic anomalies responsible for individual instances of cancer provides a more rational basis for designing combination therapies.

In the Methods section, we present our approach for finding genes with mutations responsible for cancer. We describe the map** of the problem to the weighted set cover (WSC) problem and the WSC approximation algorithm used to identify the multi-hit combinations. In the Results section, we show that our approach can identify a set of multi-hit combinations that can differentiate between tumor tissue and normal tissue samples with over 90% sensitivity and specificity. This result is robust to different randomly selected training and test sets. We discuss how these combinations can be further analyzed to distinguish carcinogenic and non-carcinogenic mutations within genes and how they may be used to design targeted combination therapies.

Results

We implemented a weighted set cover algorithm to identify 2-hit combinations of cancer causing genes with mutations using a randomly selected Training set of tumor and normal tissue samples (see Methods). The set of combinations distinguish between tumor and normal tissue samples with over 90% sensitivity and specificity. This result is robust to different Training and Test set partitions of the available tumor and normal tissue samples. Although the identified combinations contain many genes previously implicated in cancer, our approach has also identified several potentially novel cancer genes. Our results suggest that some of the combinations identified are 2-hit subsets of 3+ hit combinations.

A set of 2-hit combinations can differentiate between tumor and normal tissue samples with high accuracy

We implemented the weighted set cover algorithm described in Methods, for identifying a set of 2-hit combinations with the goal of maximizing accuracy (sensitivity and specificity) in differentiating between tumor and normal samples. Using a randomly selected Training set (see Methods), we identified a set of 2-hit combinations for each of the seventeen cancer types with at least two hundred matched tumor and blood-derived normal samples.

When tested against a separate randomly selected Test set, the identified set of combinations were able to differentiate between tumor tissue samples and normal tissue samples, for their respective cancer types, with greater than 90% specificity and sensitivity on average. Table 1 shows the sample sizes, sensitivity, and specificity for the Training and Test sets for each of the seventeen cancer types. Sensitivity varies from 83% to 100% and specificity varies from 86% to 100%, depending on cancer type.

Table 1 2-hit combinations can differentiate between tumor and normal tissue samples with over 90% sensitivity and specificity.

Full size table

The number of combinations identified varies from 8–20 for the 17 cancer types (Table 1). In total, 197 combinations were identified (Tables S2–S18). The top three 2-hit combinations are summarized in Fig. 1. The combinations include 256 unique genes with 138 genes occurring in more than one combination.

Results are robust to different Training and Test sets

To test the robustness of the above results, we randomly re-partitioned the available samples into two more alternative Training and Test sets. Figure 2 shows specificity and sensitivity of the algorithm across the seventeen cancer types considered here, for three different sets of partitions. The average difference in sensitivity between any two pairs of train-test partitions is less than 4.2% and the average difference in specificity is less than 4.1%. The largest difference in sensitivity is 12% (BLCA) and the largest difference in specificity is 13% (KIRP). In addition, the most frequently occurring combinations in the tumor samples were the same between any two train-test partitions for 14 of 17 cancer types, representing 65% of tumor samples (Fig. 3). However, there were significant differences between the less frequently occurring combinations with only 39 common combinations, out of 197 total combinations, across the three sets of combinations for the three training-test partitions (Fig. S2). Clearly, the samples included in the Training set affect the set of combinations identified. This is to be expected since 42% of the combinations occur in less than 5% of the samples for each cancer type (Fig. S4). Different partitions of the tumor samples will result in different sets of these rare combinations being included in the Training set, resulting in different combinations being identified. In addition, since the approximation algorithm used here identifies a near-optimal solution, changes in the Training set can result in different near-optimal combinations being selected by the algorithm.

The combinations identified include novel cancer genes

The genes comprising the 2-hit combinations identified above fall into three categories. (1) Confirmed cancer genes based on the Catalog of Somatic Mutations in Cancer (COSMIC) database²⁸. (2) Non-COSMIC genes that have been implicated in cancer based on experimental evidence. (3) Genes that have not been experimentally implicated in cancer. Table 2 summarizes, from Tables S2–S18, the 31 genes that comprise the top three most frequently occurring 2-hit combinations for each of the cancer types studied. Of these genes, nine are confirmed cancer genes (e.g. APC, IDH1, KRAS, PTEN, RB1, and TP53), thirteen have been experimentally implicated in cancer (e.g. HLA-C, IGHG1, and KCNB1), and nine have not previously been implicated in cancer (e.g. TUBBP12).

Table 2 Genes in the top three most frequently occurring 2-hit combinations.

Full size table

The genes in the last category have not been extensively studied, and represent potentially novel cancer genes. For example, TUBB8P12 (Tubulin Beta 8 Pseudogene 12) occurs in the top three 2-hit combinations in 15 of the 17 cancer types. However, TUBB8P12 has not been previously identified as frequently mutated in cancers. There are two possible reasons why we have identified TUBB8P12 as a potential cancer gene while previous bioinformatics studies have not. The first reason is that, we considered low frequency somatic mutations, identified using matched tumor and blood derived normal samples, that were not included in many of the previous studies^{9,12,S5 shows the distribution for all seventeen cancer types. 64.5% tumor samples contain multiple combinations, suggesting that the 2-hit combinations might represent subsets of three or more hits.}

Full size image

Genes within combinations are not correlated

Analysis of genes within each combination shows that they are not correlated. For each of the genes in a combination we construct a vector of 0’s and 1’s. The length of the vector is equal to the number of normal samples, and the value in the i^th position of that vector represents whether the i^th normal sample has a protein-altering mutation (as determined by the Variant Effect Predictor (VEP)) in that location or not. Then we computed Pearson’s correlation coefficient³² using stats. pearsonr routine from python module scipy. stats between two vectors representing two different genes. The Pearson correlation coefficient is less than 0.25 for the gene pairs within each combination (Fig. S1). If the genes within a combination were correlated it would have suggested that the combination is a result of some common underlying cause, such as being a passenger mutation or due to structural chromosomal modification, and unlikely to be causative. We also examined the chromosomal location of genes within each combination (Fig. 5). Only two of the 197 combinations contain genes within the same chromosome, suggesting that the genes within combinations are not due to a chromosomal abnormality that may affect multiple genes within a chromosome.

Discussion

Here we discuss how the multi-hit combinations identified above can be used to identify carcinogenic (driver) and non-carcinogenic (passenger) mutations within genes. We also illustrate how these combinations may be used to design a combination therapy targeting the specific genetic mutations responsible for individual instances of cancer.

Distinguishing between driver and passenger mutations

The method used to identify multi-hit combinations uses a mutation frequency based approach to preferentially select driver genes instead of passenger genes, i.e. the selected genes have a significantly higher mutation frequency in tumor samples compared to normal samples. For each gene, the mutation frequency in normal samples is considered to be approximately representative of the background mutation frequency for the gene. However, within these genes not all mutations are carcinogenic.

The combinations found above provide a starting point for examining a smaller subset of genes more closely to identify specific carcinogenic mutations within these genes. In identifying the multi-hit combinations, we did not take into consideration the location of mutations within genes. Clearly there are locations within a gene where certain mutations are unlikely to affect the function of the gene product. Such mutations can result in false positives and contribute to the large number (65%) of tumor samples containing multiple combinations (Fig. 4). Consider for example, the 2-hit combination of mutations in IDH1 and MUC6 in brain lower grade glioma (LGG) tumor samples. Of the 479 LGG tumor samples, 134 (28%) contain mutations in both IDH1 and MUC6, while 5 (1.5%) of 333 normal tissue samples contain a mutation in both these genes (Fig. 6). Comparing the mutations within these genes for normal and tumor samples may reveal which are carcinogenic and which are not. In this example, every one of the tumor samples contains a missense mutation at R132 in IDH1 and no other mutations, while the normal samples do not contain any mutations at this position (Fig. 6). Mutations at R132 in IDH1 have previously been implicated in cancer³³. On the other hand, the IDH1 mutations seen in the normal samples are unlikely to be carcinogenic. Similarly, mutations at F1989 of MUC6, which occur most frequently in both tumor and normal samples are unlikely to be carcinogenic (Fig. 6). Excluding such non-carcinogenic mutations can reduce the number of false positives and further increase accuracy of our algorithm. In our future work we will develop an automated method to compare and contrast the individual gene loci, so that all of these mutations within genes can be identified. To further improve accuracy of our algorithm, variants that are likely to be carcinogenic can be weighted higher than those that are unlikely to be carcinogenic.

Some of the genes identified by our approach may not be causative (passenger mutations) even though they may be correlated to cancer incidence. Functional analysis can be used to identify genes in the above set of combinations that are unlikely to be driver genes, even though they may be frequently mutated in tumors^11,34,35. For example, the affect of specific mutations on gene expression levels can be analyzed to determine if the mutation is likely to have a functional effect. In addition we can analyze the pathways affected by the gene combinations (Tables S19–S22). Studies show that combinations of driver gene mutations generally affect mutually exclusive pathways³⁶. Therefore, one of the genes in a multi-hit combination affecting the same pathway may include passenger mutations. Although in most cases multiple different pathways are affected by the gene combinations, Tables S19–S22 shows that in some cases (e.g. MUC6 and MUC12 in BRCA) the same pathway is affected by both genes in the combination. Further analysis would be required to determine if the mutations within one of these genes are passenger mutations.

The search algorithm can be run iteratively to incrementally refine the list of multi-hit combinations by excluding these passenger mutations. The input to our algorithm is a list of genes with mutations for each sample. Genes with only passenger mutations can be excluded from this list to minimize the inclusion of passenger mutations in the resulting multi-hit combinations.

A rational basis for combination therapy

The combinations identified above, with further refinement and clinical validation, may represent a more rational basis for targeted combination therapy, instead of the current “marriages of convenience”²⁷ with limited biological rationale²⁶. A more rational strategy may also reduce the risk of expensive failures such as the phase III trial of imfinzi plus tremelimumab. The combination of therapies for a given patient could be designed to target specific carcinogenic combinations of gene mutations found in the patient. Although only 30 of the 256 genes in the combinations identified above were formally identified as “cancer genes” in the catalog of somatic mutations in cancer (COSMIC), many of the other genes were previously implicated in cancer (Table 2). Therapies that target many of the genes in both these categories may be available or under development. For example, the combination of mutations in TP53 and IGHG1 occur in 41% of HNSC tumor samples in TCGA. Several drugs that can restore TP53 function, deplete mutant TP53 or affect downstream targets are currently in pre-clinical development³⁷. siRNA targeted silencing of IGHG1 has been shown to inhibit cell viability and promote apoptosis, which might therefore act as a potential target in cancer gene therapy^38,39. For patients with this combination of mutations, a combination therapy targeting both these genes may be more effective in combination, than separately.

Conclusions

Cancer is many different diseases, although the symptoms may be similar. These different diseases are a result of different combinations of genetic defects (hits). In this study we have developed a method for identifying combinations of genes with mutations that may be responsible for different instances of cancer. Our method is fundamentally different from current approaches which identify individual genes, instead of combinations of genes, in which mutations increase the likelihood of carcinogenesis.

The problem of identifying a set of multi-hit combinations that can differentiate between tumor and normal samples was mapped to the extensively studied weighted set cover (WSC) problem. We adapted a WSC algorithm to the problem of identifying multi-hit combinations. The algorithm was applied to a training set of somatic mutation data from the cancer genome atlas (TCGA) to identify a set of 2-hit combinations for the 17 cancer types with at least 200 matched tumor tissue and blood-derived normal samples. The resulting 2-hit combinations were able to differentiate between tumor and normal tissue samples in a separate test set with over 90% sensitivity and specificity on average. Accuracy of the results were robust to different random partitionings of the available data between training and test sets. The resulting set of combinations include potential novel cancer genes, not previously implicated in cancer.

We show how carcinogenic and non-carcinogenic mutations within genes could be identified, by comparing the occurrence of different mutations in tumor and normal samples. We also illustrate how the combination of mutations responsible for an individual instance of cancer can be used to design a combination therapy targeting the specific genes responsible for that instance of cancer.

Methods

Our approach for identifying sets of multi-hit combinations consists of two steps (Fig. 7). First, we identified somatic mutations from whole exome sequencing data for tumor and normal tissues with matched blood-derived normal samples from The Cancer Genome Atlas (TCGA). Somatic variants called from matched tumor tissue and blood-derived normal samples can detect low-frequency variants, which would not be detected when using tumor samples alone. Second, we use a weighted set cover algorithm to identify multi-hit combinations that can differentiate between tumor and normal samples with high sensitivity and specificity. The problem of identifying a set of multi-hit combinations is computationally intractable; however, there exist algorithms for finding a near-optimal approximate solution. We used a variant of one such algorithm to identify a set of multi-hit combinations for each cancer type, using a randomly selected subset of the available tumor and normal tissue samples (the Training set). The accuracy (sensitivity and specificity) of the resulting multi-hit combinations was evaluated using the remaining tumor and normal tissue samples (the Test set).

Somatic mutations calculated from the cancer genome atlas (TCGA) data

The primary input to our algorithm is somatic mutation data for tumor and normal tissue samples. TCGA contains a set of such data for tumor tissue samples with matched blood-derived normal samples, in mutation annotation format (MAF) datasets⁴⁰. These somatic mutations were identified using the commonly used and well documented Mutect2 software. For normal tissue samples we identified a set of 333 normal tissue samples with matched blood-derived normal samples. We calculated somatic mutations for these normal tissue samples using the same Mutect2 protocol used for the tumor tissue samples. We use the Variant Effect Predictor (VEP) to determine the location (intron, exon, UTR) and effect of these variants (synonymous, non-synonymous, missense, nonsense). The specific commands and parameters used are included in Supporting Information (SI). In our analysis we only consider protein-altering variants (non-synonymous, nonsense, and insertion/deletions in exons), as predicted by VEP. We found 6733 tumor samples with ~10⁷ pre-calculated protein-altering somatic variants in the MAF files for the 17 cancer types with at least 200 matched tumor and blood-derived normal samples. In addition, we found 333 matched normal tissue samples in TCGA, in which we identified ~10⁶ protein-altering somatic mutations using the Mutect2/VEP protocol detailed in SI.

The algorithm presented below is based on the somatic mutation data described above, which does not include possible germline mutations that may contribute to carcinogenesis. However, carcinogenic germline mutations are in general relatively rare. For example, BRCA1 is one such rare exception where it occurs as a germline mutation in 5–10% of breast and ovarian cancer patients with a BRCA1 mutation^41,42. However, the other 90–95% of cases with the BRCA1 mutations are somatic variants. Therefore, the following algorithm should still be able to identify mutations in such genes as carcinogenic, although the possible presence of germline mutations may limit the accuracy of the algorithm.

Map** the problem of finding multi-hit combinations to a weighted set cover problem

Our goal is to identify a set of multi-hit combinations of gene mutations, such that at least one combination occurs in each tumor sample while minimizing the number of normal samples containing any of the combinations. Identifying this set of carcinogenic multi-hit combinations can be mapped to the extensively studied weighted set cover (WSC) problem. The WSC problem can be described as follows. For a universal set of elements and a collection of wighted subsets of this universal set, find a minimum weight collection of subsets such that all elements of the universal set are covered. The problem of identifying a set of multi-hit combinations that optimally differentiates between tumor and normal samples can be mapped to the WSC problem as follows.

1.
Let, \(T=\{{t}_{1},{t}_{2},\ldots ,{t}_{{N}_{t}}\}\) be a set of N_t tumor samples, and \(N=\{{n}_{1},{n}_{2},\ldots ,{n}_{{N}_{n}}\}\) be a set of N_n normal samples. We consider T as the universal set in the WSC problem.
2.
Let C = {c₁, c₂, …, c_M} be a set of M possible combinations. We construct a subset for each of these combinations by taking the tumor samples containing that combination. \({T}^{{c}_{i}}\) represents the subset associated with combination c_i, i.e. \({T}^{{c}_{i}}=\{{t}_{1}^{{c}_{i}},{t}_{2}^{{c}_{i}},\ldots \}\), where all tumor samples in \({T}^{{c}_{i}}\) contain the combination c_i. Union of all the subsets \({T}^{{c}_{i}}\) constructs the universal set T.
3.
Assign a weight w_i to each combination c_i (subset \({T}^{{c}_{i}}\) in the WSC problem) such that the weight represents the inverse likelihood of the combination being carcinogenic. w_i is described below. Combinations with lower weights have higher likelihood to be carcinogenic.
4.
Find a set of combinations \({C}^{\ast }=\{{c}_{1}^{\ast },{c}_{2}^{\ast },\ldots \}\) such that all the samples in T are covered and the total weight \(W=\sum {w}_{j}^{\ast }\) is minimized.

The goal of the algorithm is to maximize sensitivity TP/N_t and specificity TN/N_n, where TP is the number of true positives, TN is the number of true negatives, N_t is the number of tumor samples, N_n is the number of normal samples (Fig. 8). Therefore, we assign a weight to each combination as the inverse of the accuracy metric, \({w}_{i}={(\frac{\alpha TP+TN}{{N}_{t}+{N}_{n}})}^{-1}\), where 0 ≤ α ≤ 1 is a scaling factor. The scaling factor is used to balance the optimization of sensitivity and specificity simultaneously. We use the scaling factor 0.1 to reflect the fact that the WSC solution for the Training set always has a true positive rate of 1.0, i.e. every tumor sample in the Training set contains at least one combination.

Algorithm for finding an approximate solution to the weighted set cover problem

The computational complexity for finding an optimal solution to the WSC problem scales exponentially with problem size, making it computationally intractable. For the problem of finding a set of multi-hit combinations, let G = 20000 be the number of genes and h = 8 be the maximum number of hits. Then, the number of possible combinations \(M={\sum }_{c=2}^{h}(\frac{G}{c})\approx 6\times {10}^{29}\). The number of possible subsets of these combinations is 2^M. The optimal solution would be a subset of combinations with the minimum weight. Though a brute-force search could find the optimal solution, the size of the search-space makes the task computationally impossible. However, many approximate algorithms have been developed and analyzed for solving set cover and weighted set cover problems. We use the approximation algorithm illustrated in Fig. 9. The algorithm iteratively performs the following steps until all tumor samples have been selected:

1.
Compute weights for each possible combination of genes using only the unselected samples (initially all samples)
2.
Greedily choose the combination with the lowest weight
3.
Select all samples containing this combination and exclude from subsequent iterations.

The computational complexity for this algorithm is O(NM), where N is the number of tumor samples and M is the number of possible multi-hit combinations, compared to 2^M for the brute force algorithm. Even with this approximation, the computational complexity of O(4 × 10³¹) for the number of samples N = 200 is still impractical with currently available computational technology. Therefore, to be able to find a solution within available computational resource we limit the number of hits to two. For h = 2, computational complexity is O(4 × 10¹⁰). In a future study we will optimize and parallelize the algorithm to make it practical to identify more than two hits.

Data and Source

Data and source can be found at the following bitbucket repository: (https://bitbucket.org/sajal000/multihit-combinations).

Change history

09 December 2019
An amendment to this paper has been published and can be accessed via a link at the top of the paper.

References

Tomasetti, C., Marchionni, L., Nowak, M. A., Parmigiani, G. & Vogelstein, B. Only three driver gene mutations are required for the development of lung and colorectal cancers. Proc Natl Acad Sci USA 112, 118–123 (2015).
ADS CAS PubMed Google Scholar
Zhang, X. & Simon, R. Estimating the number of rate limiting genomic changes for human breast cancer. Breast Cancer Res Treat 91, 121–124 (2005).
PubMed Google Scholar
Luebeck, E. G. & Moolgavkar, S. H. Multistage carcinogenesis and the incidence of colorectal cancer. Proc Natl Acad Sci USA 99, 15095–15100 (2002).
ADS CAS PubMed PubMed Central Google Scholar
Little, M. & Wright, E. A stochastic carcinogenesis model incorporating genomic instability fitted to colon cancer data. Mathematical biosciences 183, 111–134 (2003).
MathSciNet CAS PubMed MATH Google Scholar
Ashley, D. The two “hit” and multiple “hit” theories of carcinogenesis. Br J Cancer 23, 313 (1969).
CAS PubMed PubMed Central Google Scholar
Armitage, P. & Doll, R. The age distribution of cancer and a multi-stage theory of carcinogenesis. Br J Cancer 8, 1 (1954).
CAS PubMed PubMed Central Google Scholar
Nordling, C. A new theory on the cancer-inducing mechanism. Br J Cancer 7, 68 (1953).
CAS PubMed PubMed Central Google Scholar
Anandakrishnan, R. Estimating the number of genetic mutations (hits) required for carcinogenesis based on the distribution of somatic mutations. PLOS Comp Bio In Review (2018).
Tian, R., Basu, M. & Capriotti, E. Contrastrank: a new method for ranking putative cancer driver genes and classification of tumor samples. Bioinformatics 30, 572–578 (2014).
Google Scholar
Tamborero, D., Gonzalez-Perez, A. & Lopez-Bigas, N. Oncodriveclust: exploiting the positional clustering of somatic mutations to identify cancer genes. Bioinformatics 29, 2238–2242 (2013).
CAS PubMed Google Scholar
Dees, N. D. et al. Music: identifying mutational significance in cancer genomes. Genome Res 22, 1589–1598 (2012).
CAS PubMed PubMed Central Google Scholar
Kumar, R. D., Swamidass, S. J. & Bose, R. Unsupervised detection of cancer driver mutations with parsimony-guided learning. Nat Genet 48, 1288–1294 (2016).
CAS PubMed PubMed Central Google Scholar
Kuchenbaecker, K. B. et al. Risks of breast, ovarian, and contralateral breast cancer for BRCA1 and BRCA2 mutation carriers. JAMA 317, 2402–2416 (2017).
CAS PubMed Google Scholar
Mai, P. et al. Li-Fraumeni syndrome: report of a clinical research workshop and creation of a research consortium. Cancer Genet 205, 479–87 (2012).
PubMed PubMed Central Google Scholar
Pantziarka, P. Primed for cancer: Li Fraumeni Syndrome and the pre-cancerous niche. Ecancermedicalscience 9, 541 (2015).
PubMed PubMed Central Google Scholar
Guha, T. & Malkin, D. Inherited TP53 mutations and the Li-Fraumeni syndrome. Cold Spring Harb Perspect Med 7, a026187 (2017).
PubMed PubMed Central Google Scholar
Amadou, A., Waddington Achatz, M. & Hainaut, P. Revisiting tumor patterns and penetrance in germline TP53 mutation carriers: temporal phases of Li-Fraumeni syndrome. Curr Opin Oncol 30, 23–29 (2018).
CAS PubMed Google Scholar
Grant, R. C. et al. Prevalence of germline mutations in cancer predisposition genes in patients with pancreatic cancer. Gastroenterology 148, 556–564 (2015).
CAS PubMed Google Scholar
Kinzler, K. W. & Vogelstein, B. Lessons from hereditary colorectal cancer. Cell 87, 159–170 (1996).
CAS PubMed Google Scholar
Stahl, M. et al. Epigenetics in Cancer: A hematological perspective. PLoS Genet 12, e1006193 (2016).
PubMed PubMed Central Google Scholar
Schneider G, R. R. S. D. & Schmidt-Supprian, M. Tissue-specific tumorigenesis: context matters. Nat Rev Cancer 17, 239–53 (2017).
PubMed PubMed Central Google Scholar
Almassalha, L. et al. The greater genomic landscape: The heterogeneous rvolution of cancer. Cancer Res 76, 5605–9 (2016).
CAS PubMed PubMed Central Google Scholar
Vogelstein, B. et al. Cancer genome landscapes. Science 339, 1546–58 (2013).
ADS CAS PubMed PubMed Central Google Scholar
Chvatal, V. A greedy heuristic for the set-covering problem. Mathematics of Operations Research 4, 233–235 (1979).
MathSciNet MATH Google Scholar
Feige, U. A threshold of ln n for approximating set cover. Journal of the ACM (JACM) 45, 634–652 (1998).
MATH Google Scholar
Al-Lazikani, B., Banerji, U. & Workman, P. Combinatorial drug therapy for cancer in the post-genomic era. Nature biotechnology 30, 679 (2012).
CAS PubMed Google Scholar
Ledford, H. Cocktails for cancer with a measure of immunotherapy. Nature 532, 162–164 (2016).
ADS CAS PubMed Google Scholar
Pleasance, E. et al. A comprehensive catalogue of somatic mutations from a human cancer genome. Nature 463, 191–196 (2010).
ADS CAS PubMed Google Scholar
**, J., Wang, M. & Li, A. Discovering mutated driver genes through a robust and sparse co-regularized matrix factorization framework with prior information from mRNA expression patterns and interaction network. BMC Bioinformatics 19, 1–14 (2018).
Google Scholar
Spencer, D. H. et al. Performance of common analysis methods for detecting low-frequency single nucleotide variants in targeted next-generation sequence data. J Mol Diag 16, 75–88 (2014).
CAS Google Scholar
Sandmann, S. et al. Evaluating variant calling tools for non-matched next-generation sequencing data. Sci Rep 7, 43169 (2017).
ADS CAS PubMed PubMed Central Google Scholar
Pearson, K. Mathematical contributions to the theory of evolution. iii. regression, heredity, and panmixia. Philosophical Transactions of the Royal Society of London. Series A, containing papers of a mathematical or physical character 187, 253–318 (1896).
ADS Google Scholar
Liu, X. & Ling, Z.-Q. Role of isocitrate dehydrogenase 1/2 (IDH 1/2) gene mutations in human tumors. Histology and Histopathology 30, 1155–1160 (2015).
CAS PubMed Google Scholar
Merid, S. K., Goranskaya, D. & Alexeyenko, A. Distinguishing between driver and passenger mutations in individual cancer genomes by network enrichment analysis. BMC Bioinformatics 14, 308 (2014).
Google Scholar
Berger, A. et al. High-throughput phenoty** of lung cancer somatic mutations. Cancer Cell 30, 214–228 (2016).
CAS PubMed PubMed Central Google Scholar
Leiserson, M. D., Reyna, M. A. & Raphael, B. J. A weighted exact test for mutually exclusive mutations in cancer. Bioinformatics 32, 736–745 (2016).
Google Scholar
Parrales, A. & Iwakuma, T. Targeting oncogenic mutant p53 for cancer therapy. Front Oncol 5, 288 (2015).
PubMed PubMed Central Google Scholar
Pan, B., Zheng, S., Liu, C. & Xu, Y. Suppression of IGHG1 gene expression by siRNA leads to growth inhibition and apoptosis induction in human prostate cancer cell. Mol Biol Rep 40, 27–33 (2013).
CAS PubMed Google Scholar
Xu, Y. et al. IgG silencing induces apoptosis and suppresses proliferation, migration and invasion in LNCaP prostate cancer cells. Cell Mol Biol Lett 21, 27 (2016).
PubMed PubMed Central Google Scholar
Weinstein, J. et al. The cancer genome atlas pan-cancer analysis project. Nat Genet 48, 1288–1294 (2016).
Google Scholar
Copson, E. R. et al. Germline BRCA mutation and outcome in young-onset breast cancer (POSH): a prospective cohort study. Lancet Oncol 19, 169–180 (2018).
CAS PubMed PubMed Central Google Scholar
Berchuck, A. et al. Frequency of germline and somatic BRCA1 mutations in ovarian cancer. Clin Cancer Res 4, 2433–2437 (1998).
CAS PubMed Google Scholar
Zhang, H., Meltzer, P. & Davis, S. Rcircos: an R package for Circos 2D track plots. BMC Bioinformatics 14, 244 (2013).
PubMed PubMed Central Google Scholar
Cerami, E. et al. The cbio cancer genomics portal: An open platform for exploring multidimensional cancer genomics data. Cancer Discovery 2, 401–404 (2012).
PubMed Google Scholar
Gao, J. et al. Integrative analysis of complex cancer genomics and clinical profiles using the cbioportal. Sci. Signal. 6, pl1–pl1 (2013).
PubMed PubMed Central Google Scholar
Song, Y. & C.J., Z. Effect of hydralazine on demethylation status and expression of APCgene, proliferation and apoptosis of human cervical cancer cell lines. Chinese journal of pathology 36, 614–8 (2007).
CAS PubMed Google Scholar
Wang, T. et al. Increased nucleotide polymorphic changes in the 5′-untranslated region of δ-catenin (CTNND2) gene in prostate cancer. Oncogene 28, 555–564 (2009).
CAS PubMed Google Scholar
Dhillon, S. Ivosidenib: First global approval. Drugs 78, 1509–1516 (2018).
CAS PubMed PubMed Central Google Scholar
Liu, X., Jakubowski, M. & Hunt, J. KRAS gene mutation in colorectal cancer is correlated with increased proliferation and spontaneous apoptosis. Am J Clin Pathol 135, 245–52 (2011).
CAS PubMed Google Scholar
Matsuyama, T. et al. MUC12 mRNA expression is an independent marker of prognosis in stage II and stage III colorectal cancer. Int J Cancer 127, 2292–9 (2010).
CAS PubMed Google Scholar
Vincent, A. et al. Epigenetic regulation (DNA methylation, histone modifications) of the 11p15 mucin genes (MUC2, MUC5AC, MUC5B, MUC6) in epithelial cancer cells. Oncogene 26, 6566–76 (2007).
CAS PubMed Google Scholar
Yim, E. et al. Rak functions as a tumor suppressor by regulating PTEN protein stability and function. Cancer Cell 15, 304–14 (2009).
CAS PubMed PubMed Central Google Scholar
Arima, Y. et al. Rb depletion results in deregulation of E-cadherin and induction of cellular phenotypic changes that are characteristic of the epithelial-to-mesenchymal transition. Cancer Res 68, 5104–12 (2008).
CAS PubMed Google Scholar
Vannini, I. et al. Role of p53 codon 72 arginine allele in cell survival in vitro and in the clinical outcome of patients with advanced breast cancer. Tumour Biol 29, 145–51 (2008).
CAS PubMed Google Scholar
Ma, J. et al. 15-lipoxygenase-1/15-hydroxyeicosatetraenoic acid promotes hepatocellular cancer cells growth through protein kinase b and heat shock protein 90 complex activation. Int J Biochem Cell Biol 45, 1031–41 (2013).
CAS PubMed Google Scholar
Ben-Arie, A., Hagay, Z., Ben-Hur, H., Open, M. & Dgani, R. Elevated serum alkaline phosphatase may enable early diagnosis of ovarian cancer. Eur J Obstet Gynecol Reprod Biol 86, 69–71 (1999).
CAS PubMed Google Scholar
Natrajan, R. et al. Amplification and overexpression of CACNA1E correlates with relapse in favorable histology Wilms’ tumors. Clin Cancer Res 12, 7284–93 (2006).
CAS PubMed Google Scholar
Ritterhouse, L. L. et al. Ros1 rearrangement in thyroid cancer. Thyroid 26, 1 (2016).
Google Scholar
Tan, E., Richard, C., Zhang, H., Hoskin, D. & Blay, J. Adenosine downregulates DPPIV on HT-29 colon cancer cells by stimulating protein tyrosine phosphatase(s) and reducing ERK1/2 activity via a novel pathway. Am J Physiol Cell Physiol 291, 433–44 (2006).
Google Scholar
Paul, N. et al. α5β1 integrin recycling promotes Arp2/3-independent cancer cell invasion via the formin FHOD3. J Cell Biol 210, 1013–31 (2015).
CAS PubMed PubMed Central Google Scholar
An, Q. et al. Heterogeneous breakpoints in patients with acute lymphoblastic leukemia and the dic(9; 20)(p11-13; q11) show recurrent involvement of genes at 20q11.21. Haematologica 94, 1164–9 (2009).
CAS PubMed PubMed Central Google Scholar
Verheyden, S. et al. Role of the inhibitory KIR ligand HLA-Bw4 and HLA-C expression levels in the recognition of leukemic cells by natural killer cells. Cancer Immunol Immunother 58, 855–65 (2009).
CAS PubMed Google Scholar
Mundhada, S., Luthra, R. & Cano, P. Association of HLA class i and class ii genes with bcr-abl transcripts in leukemia patients with t(9; 22) (q34; q11). BMC Cancer 4, 25 (2004).
PubMed PubMed Central Google Scholar
Fleming, J., Ginsburg, E., Oliver, S., Goldsmith, P. & Vonderhaar, B. Hornerin, an s100 family protein, is functional in breast cells and aberrantly expressed in breast cancer. BMC Cancer 12, 266 (2012).
CAS PubMed PubMed Central Google Scholar
Coma, M. et al. Impaired voltage-gated K+ channel expression in brain during experimental cancer cachexia. FEBS Lett 536, 45–50 (2003).
CAS PubMed Google Scholar
Qin, Y., Tang, X. & Liu, M. Tumor-suppressor gene NBPF1 inhibits invasion and PI3K/mTOR signaling in cervical cancer cells. Oncol Res 23, 13–20 (2016).
PubMed PubMed Central Google Scholar
Tsai, L. et al. The sodium-dependent glucose cotransporter SLC5A11 as an autoimmune modifier gene in SLE. Tissue Antigens 71, 114–126 (2007).
PubMed Google Scholar

Download references

Author information

Authors and Affiliations

Department of Computer Science, Virginia Tech, Blacksburg, VA, USA
Sajal Dash & Wu-chun Feng
Biomedical Sciences, Edward Via College of Osteopathic Medicine, Blacksburg, VA, USA
Nicholas A. Kinney, Robin T. Varghese, Harold R. Garner & Ramu Anandakrishnan
Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA, USA
Wu-chun Feng
Gibbs Cancer Center and Research Institute, Spartanburg, SC, USA
Nicholas A. Kinney, Robin T. Varghese, Harold R. Garner & Ramu Anandakrishnan

Authors

Sajal Dash
View author publications
You can also search for this author in PubMed Google Scholar
Nicholas A. Kinney
View author publications
You can also search for this author in PubMed Google Scholar
Robin T. Varghese
View author publications
You can also search for this author in PubMed Google Scholar
Harold R. Garner
View author publications
You can also search for this author in PubMed Google Scholar
Wu-chun Feng
View author publications
You can also search for this author in PubMed Google Scholar
Ramu Anandakrishnan
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

R.A. and S.D. designed the algorithm, implemented the software, analyzed the results and wrote the manuscript. R.A., R.V. and N.K. prepared the input data, R.A. and S.D. analyzed the results. R.A., H.G. and W.F. contributed towards the research design.

Corresponding author

Correspondence to Ramu Anandakrishnan.

Ethics declarations

Competing Interests

The authors declare no competing interests.

Additional information

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Material

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Dash, S., Kinney, N.A., Varghese, R.T. et al. Differentiating between cancer and normal tissue samples using multi-hit combinations of genetic mutations. Sci Rep 9, 1005 (2019). https://doi.org/10.1038/s41598-018-37835-6

Download citation

Received: 01 August 2018
Accepted: 14 December 2018
Published: 30 January 2019
DOI: https://doi.org/10.1038/s41598-018-37835-6
Springer Nature Limited

This article is cited by

Definition of a novel breast tumor-specific classifier based on secretome analysis
- Philémon Sirven
- Lilith Faucheux
- Vassili Soumelis
Breast Cancer Research (2022)
Identifying multi-hit carcinogenic gene combinations: Scaling up a weighted set cover algorithm using compressed binary matrix representation on a GPU
- Qais Al Hajri
- Sajal Dash
- Ramu Anandakrishnan
Scientific Reports (2020)

Associated content

Top 100 in Cancer

Collection 11 March 2020

Differentiating between cancer and normal tissue samples using multi-hit combinations of genetic mutations

Abstract

Similar content being viewed by others

CoMEt: a statistical approach to identify combinations of mutually exclusive alterations in cancer

Cancer subtype identification using somatic mutation data

Cancer classification in the genomic era: five contemporary problems

Introduction

Results

A set of 2-hit combinations can differentiate between tumor and normal tissue samples with high accuracy

Results are robust to different Training and Test sets

The combinations identified include novel cancer genes

Genes within combinations are not correlated

Discussion

Distinguishing between driver and passenger mutations

A rational basis for combination therapy

Conclusions

Methods

Somatic mutations calculated from the cancer genome atlas (TCGA) data

Map** the problem of finding multi-hit combinations to a weighted set cover problem

Algorithm for finding an approximate solution to the weighted set cover problem

Data and Source

Change history

09 December 2019

References

Author information

Authors and Affiliations

Contributions

Corresponding author

Ethics declarations

Competing Interests

Additional information

Supplementary information

Supplementary Material

Rights and permissions

About this article

Cite this article

Share this article

This article is cited by

Definition of a novel breast tumor-specific classifier based on secretome analysis

Identifying multi-hit carcinogenic gene combinations: Scaling up a weighted set cover algorithm using compressed binary matrix representation on a GPU

Search

Navigation