Machine learning diagnosis by immunoglobulin N-glycan signatures for precision diagnosis of urological diseases

Abstract

Early diagnosis of urological diseases is often difficult due to the lack of specific biomarkers. More powerful and less invasive biomarkers that can be used simultaneously to identify urological diseases could improve patient outcomes. The aim of this study was to evaluate a urological disease-specific scoring system established with a machine learning (ML) approach using Ig N-glycan signatures. Immunoglobulin N-glycan signatures were analyzed by capillary electrophoresis from 1312 serum subjects with hormone-sensitive prostate cancer (n = 234), castration-resistant prostate cancer (n = 94), renal cell carcinoma (n = 100), upper urinary tract urothelial cancer (n = 105), bladder cancer (n = 176), germ cell tumors (n = 73), benign prostatic hyperplasia (n = 95), urosepsis (n = 145), and urinary tract infection (n = 21) as well as healthy volunteers (n = 269). Immunoglobulin N-glycan signature data were used in a supervised-ML model to establish a scoring system that gave the probability of the presence of a urological disease. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC). The supervised-ML urologic disease-specific scores clearly discriminated the urological diseases (AUC 0.78-1.00) and found a distinct N-glycan pattern that contributed to detect each disease. Limitations included the retrospective and limited pathological information regarding urological diseases. The supervised-ML urological disease-specific scoring system based on Ig N-glycan signatures showed excellent diagnostic ability for nine urological diseases using a one-time serum collection and could be a promising approach for the diagnosis of urological diseases.

Keywords: biomarker; glycosylation; immunoglobulin; machine learning; urologic disease.

FIGURE 1

Schematic flow of N‐glycomics of Ig and relative peak area heatmap of 26 different Ig N‐glycans in each disease. (A) A total of 1312 serum samples were subjected to N‐glycomics of Ig. (B) N‐glycan signatures of Ig data. Relative peak area heatmap of 26 different Ig N‐glycans in each disease. Ig N‐glycan concentrations were clustered according to the distinct N‐glycan synthetic pathways and disease groups. BCa, bladder cancer; BPH, benign prostatic hyperplasia; CRPC, castration‐resistant prostate cancer; GCT, germ cell tumor; HSPC, hormone‐sensitive prostate cancer; HV, healthy volunteer; RCC, renal cell carcinoma; US, urosepsis; UTI, urinary tract infection; UTUC, upper urinary tract urothelial cancer

FIGURE 2

N‐glycan signature of Ig. (A) Twenty‐six different Ig N‐glycans were aligned according to the N‐glycan synthetic pathway. N‐glycan structures are indicated by monosaccharide symbols: yellow circles, galactose (Gal); green circles, mannose (Man); blue squares, N‐acetylglucosamine (GlcNAc); red triangles, fucose (Fuc); and magenta diamonds, N‐acetylneuraminic acid (Neu5Ac). BCa, bladder cancer; BPH, benign prostatic hyperplasia; CRPC, castration‐resistant prostate cancer; GCT, germ cell tumor; HSPC, hormone‐sensitive prostate cancer; HV, healthy volunteer; RCC, renal cell carcinoma; US, urosepsis; UTI, urinary tract infection; UTUC, upper urinary tract urothelial cancer

FIGURE 3

Supervised machine learning (ML) diagnostic modeling and evaluation of urological disease‐specific score. (A) ML‐supervised diagnostic modeling by DataRobot. Eighty percent of the dataset (n = 1049) was divided into five mutually exclusive partitions, four of which were used as training and the last used for validation used for modeling of urological disease‐specific scores with the TensorFlow Deep Learning Classifier algorithm. (B) Validation of urological disease‐specific scores by true negative/positive frequencies and receiver operating characteristic curve (ROC) analysis using holdout dataset (20% of whole data, n = 262) and ROC analysis of urological disease‐specific scores using the whole dataset (n = 1312). AUC, area under the ROC curve

FIGURE 4

True and false positive/negative frequency in confusion matrix of supervised machine learning urological disease‐specific score evaluated in holdout dataset. The left column shows each disease‐specific scoring system and the upper row shows the predicted results. True positive/negative and false positive/positive rates for cases determined to have each disease using each disease‐specific scoring system are shown. The size of the green circle represents the true positive/negative frequency. The size of the magenta circle represents the false positive/negative frequency. BCa, bladder cancer; BPH, benign prostatic hyperplasia; CRPC, castration‐resistant prostate cancer; GCT, germ cell tumor; HSPC, hormone‐sensitive prostate cancer; HV, healthy volunteer; RCC, renal cell carcinoma; US, urosepsis; UTI, urinary tract infection; UTUC, upper urinary tract urothelial cancer

Figure 5

Impact of specific N‐glycans for detection of each disease by urological disease‐specific score. The upper graphs represent the impact of N‐glycan structures for the detection of each disease. Relative impact >0.5 is represented as a red bar. A dotted square in the lower Ig N‐glycan synthetic pathway shows the N‐glycan structure with relative impact >0.5 for each disease. N‐glycan structures are indicated by monosaccharide symbols: yellow circles, galactose (Gal); green circles, mannose (Man); blue squares, N‐acetylglucosamine (GlcNAc); red triangles, fucose (Fuc); and magenta diamonds, N‐acetylneuraminic acid (Neu5Ac). BCa, bladder cancer; BPH, benign prostatic hyperplasia; CRPC, castration‐resistant prostate cancer; GCT, germ cell tumor; HSPC, hormone‐sensitive prostate cancer; HV, healthy volunteer; RCC, renal cell carcinoma; US, urosepsis; UTI, urinary tract infection; UTUC, upper urinary tract urothelial cancer

FIGURE 6

Diagnostic accuracy of supervised machine learning urological disease‐specific score for detection of each disease in whole data. (A) Violin plot of urological disease‐specific scores for detecting each disease in the whole dataset. The red line in the violin plots indicates the interquartile range (IQR) and median value. *p < 0.05, **p < 0.005, ***p < 0.001, ****p < 0.0001. ns, not significant. (B) Receiver operating characteristic (ROC) analysis of urological disease‐specific scores for detecting each disease. AUC, area under the ROC curve; BCa, bladder cancer; BPH, benign prostatic hyperplasia; CRPC, castration‐resistant prostate cancer; GCT, germ cell tumor; HSPC, hormone‐sensitive prostate cancer; HV, healthy volunteer; RCC, renal cell carcinoma; US, urosepsis; UTI, urinary tract infection; UTUC, upper urinary tract urothelial cancer

FIGURE 7

Each urological disease‐specific score classified as clinical or pathological parameter in the whole dataset. (A) Violin plot and receiver operating characteristic (ROC) analysis of renal cell carcinoma (RCC) score classified as a pathological stage in whole dataset. (B, C) Violin plots and ROC analyses of bladder cancer (BCa) score and upper urinary tract urothelial cancer (UTUC) score classified as a pathological stage or urine cytology class