Severity modeling of propionic acidemia using clinical and laboratory biomarkers

Abstract

Purpose: To conduct a proof-of-principle study to identify subtypes of propionic acidemia (PA) and associated biomarkers.

Methods: Data from a clinically diverse PA patient population ( https://ichgcp.net/clinical-trials-registry/NCT02890342 ) were used to train and test machine learning models, identify PA-relevant biomarkers, and perform validation analysis using data from liver-transplanted participants. k-Means clustering was used to test for the existence of PA subtypes. Expert knowledge was used to define PA subtypes (mild and severe). Given expert classification, supervised machine learning (support vector machine with a polynomial kernel, svmPoly) performed dimensional reduction to define relevant features of each PA subtype.

Results: Forty participants enrolled in the study; five underwent liver transplant. Analysis with k-means clustering indicated that several PA subtypes may exist on the biochemical continuum. The conventional PA biomarkers, plasma total 2-methylctirate and propionylcarnitine, were not statistically significantly different between nontransplanted and transplanted participants motivating us to search for other biomarkers. Unbiased dimensional reduction using svmPoly revealed that plasma transthyretin, alanine:serine ratio, GDF15, FGF21, and in vivo 1-13C-propionate oxidation, play roles in defining PA subtypes.

Conclusion: Support vector machine prioritized biomarkers that helped classify propionic acidemia patients according to severity subtypes, with important ramifications for future clinical trials and management of PA.

Conflict of interest statement

The authors declare no competing interests.

© 2021. This is a U.S. government work and not under copyright protection in the U.S.; foreign copyright protection may apply.

Figures

Fig. 1. Propionic acidemia (PA) biomarkers, plasma propionylcarnitine and plasma 2-methylcitrate, are associated with levocarnitine dose and eGFR.

(a,b) There was no difference in the levels of total plasma 2-methylcitrate (2-MC) and propionylcarnitine (C3) between unpaired transplanted and nontransplanted cohorts. (c) There was no difference in the C3 levels between paired pre- and post–liver transplant samples. (d) Enteral carnitine dosage was not different between transplanted and nontransplanted participants. (e) Plasma C3 and a levocarnitine dosage showed a significant, but modest, association. (f–g) Creatinine-based eGFR was lower in liver-transplanted participants. (h–i) Plasma 2MC and C3 showed significant association with cystatin C–based eGFR. Significant p values are denoted by *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. 2-MC 2-methylcitrate, C3 propionylcarnitine, Crt-based eGFR creatinine-based estimated glomerular filtration rate, Cys C-based eGFR cystatin C–based estimated glomerular filtration rate, LT liver transplant, n.s. not significant.

Fig. 2. k-Means clustering suggested the existence of two or more classes of propionic acidemia (PA).

(a) The elbow method of clustering analysis of 12 clinical parameters (denoted by an orange brace) suggested the existence of several classes of PA. Since the low number of participants in each group adversely affected our ability to perform supervised machine learning, in all subsequent analysis we posited two classes of PA (mild and severe). (b) k-Means clustering of 12 clinical parameters using two centroids (shown as a large filled blue triangle and red circle) did not unequivocally separate the classes of PA. An inverted triangle with white color fill denotes a patient homozygous for two Amish alleles (PCCB: c.1606A>G; p.Asn536Asp). (c) The elbow analysis on the two known biochemical parameters (plasma propionylcarnitine and total plasma 2-methylcitrate) suggested the existence several classes of propionic acidemia. (d) k-Means clustering analysis of the two biochemical parameters using three centers suggested the existence of three classes of PA. An inverted triangle with white color fill denotes a patient homozygous for two Amish PCCB alleles (PCCB: c.1606A>G; p.Asn536Asp). Cluster 3 was composed of two participants whose clinical course was characterized by progressive chronic kidney disease and a low left ventricular ejection fraction. Better separation of clusters using biochemical parameters motivated our search for a novel biochemical biomarker, which would be less affected by the renal status.

Fig. 3. Support vector machine modeling defines the role of propionylcarnitine and total 2-methylcitrate and identified novel biomarkers delineating classes of propionic acidemia (PA).

(a) Comparison of mean accuracies generated by support vector machine models using 12 clinical parameters versus 12 clinical parameters plus 1 biochemical variable (2-MC, C3, or 1-13C-propionate oxidation). The support vector machine model using 12 clinical parameters plus 1-13C-propionate oxidation had the highest mean accuracy in predicting judges’ scores than any other model. (b) The number of variables in each combination affects the mean accuracy of model. Each dot represents the mean accuracy of 1,000 resamplings of the order in which participants could have joined the study. This figure demonstrates that the overall performance of models including 12 clinical parameters plus 1-13C-propionate oxidation peaks around 6–10 variables per combination. (c) Averaged variable importance of parameters in the final support vector machine model using 12 clinical variable plus 1-13C-propionate oxidation generated from 1,000 random resamplings. This statistical experiment revealed that the model prioritized 1-13C-propionate oxidation, FSIQ, height z-score as the top three variables. (d) The time-ordered series composed of the order in which PA participants joined the study suggests that model accuracy plateaued after the first ten participants. Degradation of the mean accuracy coincided with when a participant was added, whose PA class did not reach the consensus among judges (e.g., after subject 17). 13C Ox 1-13C-propionate oxidation, 2-MC total 2-methylcitrate, C3 propionylcarnitine, FSIQ full scale IQ, LT liver transplant, n.s. not significant. ALT alanine aminotransferase, SNHL sensorineural hearing loss, RBC red blood cells, WBC white blood cells, CysC eGFR cystatin C-based eGFR, LVEF%, percent ejection fraction of left ventricle, %RDA recommended dietary allowance, ONA optic nerve abnormality. Significant p values are denoted by *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 4. 1-13C-Propionate oxidation is associated with plasma propionylcarnitine, total 2-methylcitrate, and select clinical parameters.

In this cohort, 1-13C-propionate oxidation was not associated with a cystatin C–based estimated glomerular filtration rate and there was a trend toward statistical significance for LVEF% (p value = 0.07). Logistic regression was performed to evaluate the association between sensorineural hearing loss, optic nerve abnormality, and the 1-13C-propionate oxidation. ALT alanine aminotransferase, eGFR estimated glomerular filtration rate, LVEF% left ventricular ejection fraction percent, RBC red blood cell count, RDA recommended dietary allowance, WBC white blood cell count, n.s. not significant. Red dotted lines denote a 95% prediction band.

Fig. 5. External validation analysis of novel biomarkers using data from liver-transplanted participants supported the role of select biomarkers in defining mild and severe propionic acidemia (PA).

(a,b) Two known diagnostic PA biomarkers, plasma total 2-methylcitrate and propionylcarnitine, were not different between the mild PA and liver-transplanted participants. (c) Plasma transthyretin was ranked 42/301 by the svmPoly model. The difference between severe and mild PA as well as severe PA and PA after LT was statistically significant, but there was an overlap in values between severe PA and PA after LT. (d) Plasma GDF15 was significantly lower in the mild PA compared to severe PA and there was an overlap in values between severe PA and PA after LT. (e). Plasma FGF21 was significantly lower in mild PA and PA after LT when compared to the severe PA. There was a minimal overlap in the values between the severe PA and PA after LT classes. (f) In this cohort, plasma alanine:serine ratio was ranked high by the svmPoly model (8/301). However, there was a significant overlap between severe PA and PA after LT. (g) The mean 1-13C-propionate oxidation in transplanted PA participants was significantly higher that the severe PA class. Importantly, there was no overlap in values between severe PA participants and PA after LT. 2-MC total 2-methylcitrate, C3 propionylcarnitine, LT liver transplant, n.s. not significant, SVM support vector machines. Gray shading denotes the range of parameters found in transplanted PA patients. Significant p values are denoted by *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.