Constructing the model
The plasma clinical chemistry, metabolomics and proteomics results of a human nutritional intervention study were visualized in a 3-dimensional space (Figure 1), where the three axes represented the 'overarching processes' oxidation, metabolism and inflammation. These processes were shown to be affected by the treatment [10]. For these processes p = 7, 115 and 18 parameters respectively, changed significantly (p < 0.05) due to the intervention (Additional file 1). Each subject is represented in this 3-D space (both prior to and after treatment), based on the unique values of these 140 metabolites and proteins. The method is based on the construction of three PLS-DA models, one for each of the above processes. A comparison between scores within a PLS-DA model is meaningful, but scores can not directly be compared between PLS-DA models. In order to compare the scores per process, the scores of the three models were subsequently scaled around zero (the average of the dietary-mix group) and one (average of the placebo group). The difference between 0 and 1 was used to separate out relatively healthy subjects from less healthy subjects, assuming that people become healthier using the anti-inflammatory mix. This is further described in the methods section.
Analysis and visualization of the effects of the anti-inflammatory mix showed a separation between the treated and control groups, suggesting an average treatment effect. This is logical, since the axes were constructed using all parameters that changed significantly for the average study group. The unique feature of this visualization model, however, is that the changes were visualized and directly correlated to the relevant biological processes (metabolism, oxidation and inflammation). Overall, a treatment related shift was observed along the predefined inflammation and metabolism axes, and to a lesser extent along the oxidation axis. This is emphasized by the 2-D projections in Figure 1. Note that the process oxidation contained fewer parameters (Additional file 1).
In addition to the group treatment effect, strong differences in inter-individual responses were visualized (i.e., the length of the vector is different, but the direction is similar).
Dietary compounds have a marked effect on the response
As the health space model is partly based on metabolic profiles, plasma metabolites derived from the dietary intervention itself might interfere with the division into groups. These metabolites would thus report on the diet effect in plasma rather than on the resulting change in metabolism or related health status. Therefore, we rebuilt the health space without the dietary metabolites (omega-3 PUFAs, 2,3,4,-trihydroxybutanoic acid and vitamin E).
After excluding the dietary metabolites, we built two models, one using metabolomics and proteomics data (Figure 2) and a second additionally including PBMC transcriptomics data (Figure 3). Both models made it possible to distinguish the placebo- and dietary intervention groups, showing that the health space could be used for different data types. However, since the classification of transcripts in the selected functional groups (metabolism, oxidation and inflammation) requires improvement and the selection of the processes oxidation, inflammation and metabolism as axes is based on proteomics and metabolomics data [10], we further focus on the health space based only on significantly changed proteins and metabolites.
In the model, based on metabolomics and proteomics data and excluding the dietary metabolites, it was possible to distinguish the placebo group from the dietary-mix treated group (Figure 2). Oxidative stress did not significantly contribute to separating the two groups in the new health space. This indicated that the dietary compounds vitamin E and 2,3,4-trihydroxybutanoic acid were the only oxidative parameters that distinguished the placebo from the dietary-mix group in the initial model including dietary compounds. As in the initial health space model, the intensity of the response (length of the vector) differed between subjects. In contrast to the initial model, the contribution of the different processes (direction of the vector) varied between subjects; some people only responded on the inflammation axis, whereas others only responded on the metabolic stress axis. Subject 33 for instance reacted primarily on the metabolism axis. Subject 1 responded on the oxidative axis. Subject 25 showed a combined response on the oxidation and metabolism axes (see also Figure 4). Subject 14 was already within the "improved health area" before the dietary-mix treatment. He had very high C20:5-ChE, C22:6-ChE and poly-unsaturated TG's, which have large loadings in the metabolic PLS-DA model. This is why the fish oil in the dietary-mix scarcely improved Subject 14's metabolic status. Subject 1 responded differently on dietary-mix as compared to the other subjects, as shown by the vector facing in a different direction (Figure 2). This subject had a much higher concentration of plasma VCAM-1 and a lower concentration of M-CSF than average, both after taking the placebo and the dietary mix. He also had relatively high alanine transaminase (ALT) and albumin levels. After taking the placebo, subject 19 showed very different levels of inflammatory parameters compared to the other subjects. The levels of Beta2-microglobulin and uric acid were much higher than average and were even higher after the placebo period than after the dietary-mix period.
The intervention response of the subjects can be grouped
The individual response was defined as the individual difference between the placebo and the anti-inflammatory mix on the three axes. The individual response to the dietary treatment on the three different processes was clustered in two main groups (Figure 4). Subjects in group 1 mainly showed changes in the inflammation and oxidation parameters. Subjects in group 2 had primarily a metabolic response. Note that oxidation did not separate between the dietary-mix group and the placebo groups, but is important in defining the clusters. The two responder groups we identified did not significantly differ in age, BMI, order of treatment, waist-hip ratio, pre-weight, post-weight or in bodyweight change. The molecules primarily responsible for the separation were C20:5-cholesterol ester, C22:6-cholesterol ester, indole-3-propionic acid, uric acid and arachidonic acid. Indole-3-propionic acid (IPA) had a significant interaction effect. This means that the levels of IPA were affected by the anti-inflammatory treatment in group 1 but not in group 2.