Guidelines for correlation coefficient threshold settings in metabolite correlation networks exemplified on a potato association panel

Correlation network analysis has become an integral tool to study metabolite datasets. Networks are constructed by omitting correlations between metabolites based on two thresholds—namely the r and the associated p-values. While p-value threshold settings follow the rules of multiple hypotheses testing correction, guidelines for r-value threshold settings have not been defined. Here, we introduce a method that allows determining the r-value threshold based on an iterative approach, where different networks are constructed and their network topology is monitored. Once the network topology changes significantly, the threshold is set to the corresponding correlation coefficient value. The approach was exemplified on: (i) a metabolite and morphological trait dataset from a potato association panel, which was grown under normal irrigation and water recovery conditions; and validated (ii) on a metabolite dataset of hearts of fed and fasted mice. For the potato normal irrigation correlation network a threshold of Pearson’s |r|≥ 0.23 was suggested, while for the water recovery correlation network a threshold of Pearson’s |r|≥ 0.41 was estimated. For both mice networks the threshold was calculated with Pearson’s |r|≥ 0.84. Our analysis corrected the previously stated Pearson’s correlation coefficient threshold from 0.4 to 0.41 in the water recovery network and from 0.4 to 0.23 for the normal irrigation network. Furthermore, the proposed method suggested a correlation threshold of 0.84 for both mice networks rather than a threshold of 0.7 as applied earlier. We demonstrate that the proposed approach is a valuable tool for constructing biological meaningful networks.


Background
Correlation-based network analysis (CNA) has become an integral tool for studying the coordinated behavior of metabolite profiles in plant sciences. Metabolite correlation networks (CN) are constructed by correlating each two pairs of metabolites across a set of different conditions or by exploiting the natural variability of mapping populations or collection of varieties, as they provide large sample size stabilizing the correlation and reducing the error rate, e.g. in a tomato introgression line mapping population [1,2], a variety collection of sparkling wines [3], a diverse collection of Arabidopsis accessions [4], and a maize association panel [5]. The coordinated behavior of metabolites across diverse varieties gives insights into their genetic communalities. Each pairwise correlation is represented by a correlation-coefficient r ranging from -1 to 1. In addition, the significance of each correlation is evaluated by computing a p-value ranging from 0 to 1. In a metabolite CN, nodes represent metabolites and the edges between them represent the estimated correlation coefficients.
To construct metabolite CNs, threshold restrictions are applied to the correlation coefficients and their associated p-values to identify spurious correlations between metabolites. Subsequently, the non-significant correlations or edges, respectively, are removed from the network. Threshold settings for the associated p-values follow the rules of multiple hypotheses testing, such as a false discovery rate (FDR) [6]. However, guidelines for threshold settings of the correlation coefficient have not been well defined yet.
Ideally, edges between nodes in a CN reflect metabolic fluxes through a metabolic pathway. However, the relationship between metabolic pathways and the correlations between metabolites is not straight forward. Factors, such as short-term fluctuations caused by plant variability or internal noise may render into weak correlations. Systematic changes of the steady-state as well as aspects of cellular organization also need to be taken into account. Furthermore, the involvement of metabolites in multiple pathways and their extensive crosstalk within, makes it difficult to clearly affiliate metabolites to metabolic pathways in CNs. Thus, the correlation coefficient threshold cannot be universally set and must be instead adjusted to the system of study in order to extract meaningful biological data [7]. As a result, different studies have applied different selective threshold settings, e.g. Hu et al. constructed metabolite CNs for Osteoarthritis and control patients to identify significantly changing correlations between networks [8]. There the authors set a threshold for edges based on the p-value only. Via topological analysis of the difference network, they managed to highlight key metabolites that played an important role in governing the connectivity and information flow of the network. In [9] the authors used a moderate correlation coefficient threshold of 0.6 which enabled them to identify genes affecting free amino acids. In [10] a threshold of 0.7 was applied to highlight the differences between metabolite networks of fed and fasted mice. Yet again in [11] a rigorous correlation coefficient threshold of 0.8 was employed identifiying metabolic patterns for freezing tolerance in two Brachypodium Sylvaticum ecotypes. The selection of the correlation coefficient threshold, which allows meaningful biological interpretation, depends on the network topology rather than on the strength of the correlation coefficient itself. That being said, network properties associated with node connectivity alter (or better said stop altering) once a certain threshold has been reached. In other words, the selection of the correlation coefficient threshold depends on the distribution of the network's numbers of edges at varying p-values, similar to the idea of choosing a p-value threshold based on the p-value distribution in the Benjamini-Hochberg FDR multiple hypotheses testing correction [6].
We have recently profiled the tuber of a potato association panel that was grown under normal irrigation (NI) and recovery (REC) treatments for 42 metabolites and also measured a set of 45 morphological traits of the plant. [12]. For each treatment, one CN was constructed, where the correlation coefficient threshold for both CNs was set to Pearson's |r|≥ 0.4. In addition, the profiles were utilized for a genome wide association study (GWAS). Via the application of set theory to networks, a difference network was constructed, highlighting the difference set of the REC network over the NI network (REC ⊄ NI). In this perspective, the correlation between the metabolite fumaric acid and the morphological trait plant vigor was shown to be specific to the REC network. Next, we analyzed the single nuclear polymorphisms associated with fumaric acid and identified a gene coding for a RING domain protein on chromosome 1 in the potato genome and a gene coding for a zinc finger protein ZAT2 on chromosome 4. It was demonstrated before that both genes are essential for plants to cope with abiotic stresses [13][14][15][16][17][18].
In the current study, we demonstrate that the estimated correlation coefficient threshold of Pearson's |r|≥ 0.4 was crucial for the establishment of a connection between fumaric acid and plant vigor, and by that for the identification of the aforementioned regulating genes. By in silico manipulations of the tuber CN and monitoring its connectivity between nodes, we define guidelines onto how to identify the proper correlation coefficient threshold for metabolite CNs. Last, we apply the proposed method on a mouse metabolite dataset to prove its efficacy on a dataset of different origin.

Initial networks
We defined the NI and REC-CNs as weighted networks G i = (V i , E i , w) , where V i was the set of nodes corresponding to metabolites and morphological traits found in the dataset of treatment i , E was the set of edges between them, and edge weights ( w : E → R ) corresponded to the Pearson correlation coefficient. Thresholds for both networks were set to Pearson's |r|≥ 0.4 and a q-value ≤ 0.05, removing spurious correlations. At these settings, the NI-CN had |V NI | = 88 nodes and |E NI | = 438 edges connecting them; the REC-CN was composed of |V REC | = 90 nodes and |E REC | = 370 edges. The connection between fumaric acid and plant vigor was present in the REC-CN but not in the NI-CN, as the corresponding correlation coefficients were computed with 0.458 and 0.013, respectively.

Determining the correlation coefficient by testing the number of edges
As described above, the integration of a correlation into a CN depends on two threshold settings, namely the correlation coefficient and the associated p-value. In this and in other studies it was observed that the number of edges in a CN remains stable despite gradually increasing the p-value stringency settings until a certain correlation coefficient has been reached. In other words, the proposed method quantifies the number of edges that would be integrated into the CN dependent on the absolute correlation coefficient and its associated p-value. It does so in descending order at decrements of 0.1 of the correlation coefficient and ranging p-values of 0.05 to 0.01. Once the number of edges starts dropping at a certain correlation coefficient, the threshold is set to this value. We tested for significant changes in edge number by estimating the confidence intervals (CI) at 95% employing the modified Cox method [19] assuming non-normal distributions, such that: where the term corcoef corresponds to the initial correlation coefficient, edge.number. changed is a Boolean variable, edge.number is an array variable of function correlation_ network, computing the number of edges of different CNs at parameters corcoef and p.values, and the error term is defined as CI / 2.
To demonstrate this approach, we constructed a series of NI and REC networks at varying correlation coefficient and p-value settings and quantified the number of edges present ( Fig. 1). The figure shows that the number of edges remained unchanged at varying p-values until the Pearson's correlation coefficient dropped to 0.2 in the NI-CN ( Fig. 1a and b), suggesting that the correlation coefficient threshold ought to be located somewhere in-between 0.2 and 0.3. At a correlation coefficient of 0.2 the edge number ranged between 923 and 948, the mean was estimated with 938.8, the upper limit of the 95% CI was calculated with 951.34, the lower limit with 926.45. For the REC-CN a significant drop was registered at a correlation coefficient of 0.4, suggesting that the threshold was supposed to be located somewhere in-between 0.4 and 0.5 ( Fig. 1c and d). The associated values were estimated with mean = 373.4, the upper limit of the CI = 376.66, the lower limit of the CI = 370.18, the minimum = 370, and the maximum = 376.

Fine-tuning computes the correlation coefficient thresholds with 0.23 and 0.41
Next, we investigated CNs at different r-values in the range of 0.2 to 0.3 for the NI-CN and in the range of 0.4 to 0.5 for the REC-CN (Fig. 2). As before, the r threshold value was defined as the r-value when the minimum or maximum edge number was located outside the confines of the corresponding CI. This behavior occurred for the NI-CN at a correlation coefficient threshold Pearson's |r|≥ 0.23. Here, the edge number ranged between 850 and 853, the mean was calculated with 851.8, the CI lower limit was computed with 850.18 and the CI upper limit with 853.42 (Fig. 2). For the REC-CN this behavior was observed at Pearson's |r|≥ 0.41. At this r-value the edge number ranged between 354 and 359, the mean was computed with 357, the CI upper limit with 359.65 and the CI lower limit with 354.37. Consequently, the r-value threshold for the NI-CN was proposed with 0.23 and for the REC-CN with 0.41.

Bootstrapping analysis confirms correlation coefficient thresholds
To verify the proposed correlation coefficient threshold settings, we employed bootstrapping with replacement; such that 100 NI and REC-CNs were generated with 80% of the samples randomly selected, where one sample could be part of the sample subset more than once. Performing this analysis, we wanted to validate whether the estimated threshold was due to chance or indeed the result of the network topology at the proposed threshold even at a reduced set of samples (80%). As before, the analysis was divided into two cycles. For the first cycle the edge numbers of all networks were   Fig. 3c). We further investigated r-values in-between 0.5 to 0.4 (Fig. 3d), demonstrating a steady increase of median edge numbers the lower the correlation coefficient became. At the targeted r-value of 0.41 the minimum edge numbers was estimated with 337 and the maximum edge number Next, we used the bootstrapped CNs and computed the CI for each correlation coefficient at varying p-values employing the modified Cox test (Fig. 4). For the first cycle, 1% (empirical p = 0.99) of all NI-CNs at an r-value of 0.4 (Fig. 5a) revealed minimum or maximum edge numbers beyond their estimated CI (Fig. 4a); at r = 0.3, the number rose to 54%, equivalent to an empirical p-value of 0.46. At r = 0.2, all networks showed to have minimum or maximum edge numbers beyond their estimated CI (empirical p < 0.01). During the second cycle we inspected in particular the targeted r-value of 0.23, revealing that 98% (empirical p-value 0.02) of all networks had minimum or maximum edge numbers beyond their estimated CI (Fig. 4b) For the bootstrapped REC-CNs, significant changing edge numbers as suggested by the CI started to occur at r = 0.5 (Fig. 4c). Here, 21% of the CNs showed significant changes (empirical p = 0.79); at r = 0.4, 93% of all REC-CNs exerted significant changes (empirical p = 0.07), and at r = 0.3, all networks (empirical p < 0.01) had significantly changing edge numbers. For the second cycle we investigated the number of bootstrapped CNs

Method validation on mouse metabolite dataset
To validate the proposed method on a dataset of different biological background, we utilized the mouse metabolite datasets from Batushansky et al. [10]. There the authors tested the effect of fasting in the hearts of mice. For one of the analyses used in the study, the authors constructed CNs for two conditions, i.e. fed and fasted mice. To highlight the differences between networks intersection edges were identified. Networks were constructed at an absolute r-value threshold of ≥ 0.7. Here, we used the datasets for fed and fasted mice and ran it through our correlation-coefficient threshold pipeline. For the first cycle, our method suggested that the correlation coefficient threshold ought to be located in between r-values 0.8 and 0.9 for both networks (Fig. 5). For the second cycle, the r-value for which the edge number was located within the lower and upper CI was determined with 0.84 for both networks (Fig. 6). At these settings, the Mouse Fed -CN

Discussion
The construction of metabolite CNs is a non-trivial undertaking. In contrast to weighted gene co-expression networks [20], where all edges are kept within the network, the aim of metabolite CNs is to eliminate some of the correlations [21]. As such, if the correlation coefficient threshold is set too high, valuable biological data may be lost, while if the correlation coefficient threshold is set too low the plethora of edges may have a confounding effect, rendering it difficult to identify the most viable biological information associated with the data at hand. It is therefore highly important to set the correlation coefficient threshold appropriately so that meaningful biological conclusions can be derived. However, the correlation coefficient threshold for metabolite CNs cannot be universally set. Instead, it must be determined dynamically in accordance to the studied system. Although different studies have already applied different correlation coefficient thresholds to construct metabolite CNs, e.g. [1, 4,  Here we introduced an approach that allows researchers to select a correlation coefficient threshold suitable to their studied system. In our recent study on a potato association panel we constructed a CN on tuber metabolites and plant morphological traits, where we applied a correlation coefficient threshold of 0.4 [12]. At this setting we identified a critical connection between fumaric acid and plant vigor, which enabled us to identify essential genes aiding the plant to cope with abiotic stresses. Using the same potato dataset, in the current study, we demonstrated how to select the adequate correlation coefficient threshold based on an iterative approach, during which the network topology and specifically its associated edge number was monitored. A series of CNs were constructed, where different absolute correlation coefficients (from 0 to 1) were combined with a set of different p-values. To determine the r-value threshold a significant change had to be detected. We defined a significant change as the excess of the range of edge numbers beyond the confines of the corresponding CI of a CN. Once this criterion was fulfilled, the r-value threshold was set.
The initial analysis with increments of 0.1 between r-values from 0 to 1 suggested a threshold 0.2 for the NI-CN and a threshold 0.4 for the REC-CN (Fig. 1). The second cycle of our proposed method suggested an r-value threshold of 0. 23 1 and 2). Bootstrapping analysis with 100 CNs based on 80% of the samples confirmed the proposed thresholds (Figs. 3 and 4). At these threshold settings the crucial connection between fumaric acid and plant vigor was still present in the REC-CN but not in the NI-CN as suggested in [12]. To validate the proposed method, it was also applied to heart metabolite datasets of fed and fasted mice [10]. The original study proposed a threshold of 0.7 for both CNs. To compare the CNs a network intersection was generated. Our analysis suggested a threshold 0.84 for both CNs. Although our proposed threshold was significantly higher than in the original study, we showed a similar edge intersection. The proposed thresholds for the four different CNs stretched over a wide range of r-values, starting at a correlation coefficient as low as 0.23 for the NI-CN and reaching a correlation coefficient as high as 0.84 for the two mouse CNs. To identify a property that could potentially be key to this finding, we computed a number of network properties for each CN. Network properties derived from node degree (e.g. average degree, edge to node ratio, density, clustering coefficient) varied as much as the proposed correlation coefficient thresholds themselves. However, two other network properties provided interesting results that could potentially be the key elements for correlation coefficient threshold detection, namely: i) the network diameter, which is the maximum geodesic distance between any pair of nodes in a graph; and ii) the network assortativity coefficient, which is the correlation coefficient of degree between pairs of connected nodes [25]. It tells in a concise fashion how nodes are preferentially connected to each other. The diameter for the NI-CN was computed with 6 and its assortativity coefficient with 0.45, for the REC-CN the diameter was 9 and its assortativity coefficient = 0.36, for the Mouse Fed -CN the diameter was 9 and its assortativity coefficient = 0.47, and for the Mouse Fasted -CN the diameter was calculated with 8 and its assortativity coefficient with 0.52. Despite the different topologies of the four CNs these two network properties revealed comparable values. We believe that this finding should be further investigated.

Conclusions
We demonstrated that the approach developed in this study is a valuable tool for the determination of the correlation coefficient threshold for the construction of metabolite CNs. We applied our method to metabolite datasets of different biological background and the thresholds suggested varied from 0.23 over 0.41 to 0.84. Although the newly proposed r-values differed from the values utilized in the original studies, it still allowed us to obtain the same biological conclusions. It is therefore that the network topology of CN determines the biological interpretation, rather than the strength of the correlation coefficient itself. For this reason, we suggest treating CNs as unweighted graphs once the correlation coefficient has been established and non-significant correlations have been removed.

Datasets acquisition and processing
Datasets for metabolites and morphological traits were adopted from [12]. Preprocessing and quantification of metabolites and morphological traits were performed as described therein.