Advancing Personalized Medicine by Analytical Means: Selection of Three Metabolites That Allows Discrimination between Glaucoma, Diabetes, and Controls

Abstract

This paper aimed at devising an intelligence-based method to select compounds that can distinguish between open-angle glaucoma patients, type 2 diabetes patients, and healthy controls. Taking the concentration of 188 compounds measured in the aqueous humour (AH) of patients and controls, linear discriminant analysis (LDA) was used to identify the right combination of compounds that could lead to accurate diagnosis. All possibilities, using the leave-one-out approach, were considered through ad hoc programming and in silico massive data production and statistical analysis. Our proof of concept led to the selection of four molecules: acetyl-ornithine (Ac-Orn), C3 acyl-carnitine (C3), diacyl C42:6 phosphatidylcholine (PC aa C42:6), and C3-DC (C4-OH) acyl-carnitine (C3-DC (C4-OH)) that, taken in combination, would lead to a 95% discriminative success. 100% success was obtained with a non-linear combination of the concentration of three of these four compounds. By discarding younger controls to adjust by age, results were similar although one control was misclassified as a diabetes patient. Methods based on the consideration of individual clinical chemical parameters have limitations in the ability to make a reliable diagnosis, stratify patients, and assess disease progression. Leveraging human AH metabolomic data, we developed a procedure that selects a minimal number of metabolites (3-5) and designs algorithms that maximize the overall accuracy evaluating both positive predictive (PPV) and negative predictive (NPV) values. Our approach of simultaneously considering the levels of a few metabolites can be extended to any other body fluid and has potential to advance precision medicine. Artificial intelligence is expected to use algorithms that use the concentration of three to five molecules to correctly diagnose diseases, also allowing stratification of patients and evaluation of disease progression. In addition, this significant advance shifts focus from a single-molecule biomarker approach to that of an appropriate combination of metabolites.

Keywords: Parkinson’s disease; advanced statistical; aqueous humour; bioanalysis; diagnostic tools; discrimination tools; metabolomic signatures; methods; precision medicine.

Figure 1

Standard pipeline of an intelligent-based classifier (top) and our intelligent-based classifier using LDA as a core platform (bottom).

Figure 2

Accuracy through leave-one-out cross-validation. All possible combinations were tested selecting one metabolite and adding a second, third …; i.e., adding, one by one, metabolites to the classifier. The image shows the best result obtained from all possible combinations. With four molecules (red dot) the peak accuracy was circa 95%; the rest of combinations are represented with blue dots. Adding more metabolites to the four indicated metabolites did not improve precision.

Figure 3

Boxplots for the four selected metabolites. For each of the four metabolites there are concentration differences between the groups that improve classification accuracy by taking them all together instead of using a single metabolite.

Figure 4

Accuracy through leave-one-out cross-validation considering the most discriminative metabolites and selecting the minimum number for non-linear relationship. With six terms (red dot) in the non-linear equation the peak accuracy was 100%. Upon considering all possible combinations (blue dots), six was the minimum number of variables providing full (100%) accuracy.

Figure 5

Boxplots for the six selected variables. On each of the six variables, there are substantial differences between the groups that collectively contribute to discrimination.

Figure 6

A posteriori probability for each subject. Posterior probabilities of individuals being classified into the control, glaucoma, and diabetes groups, respectively, using the LDA learned with leave-one-out cross-validation. Each bar represents a subject, depicted with a different colour. The overall performance of the LDA is 100% (data of members in a given group have a greater posterior probability of 0.5 for the graph in the corresponding group).

Figure 7

Discriminant regions. Note that the LD1 values and the LD2 values are non-linear combinations of the variables (see Equations (1) and (2)), therefore they can be calculated for any individual. Each subject in the data appears as a point, whose group depends on a colour: control (black), glaucoma (green), and diabetes (magenta).