T-5224 br Gene ontology analyses of the
Gene ontology analyses of the 8 genes included in the classifier revealed highly significant enrichment of diseases associated with T-5224 malignances (Fig 3, A). The generated network by GeneMA-NIA shows that there are no direct interactions between the 8 genes of the model although 4 of them show coexpression (Fig 3, B).
Classifier validation. The 8-gene classifier was validated in an inde-pendent multicentric and international cohort of 249 voided urine samples from PFBC. The 8-gene classifier performed well in this real clinical scenario (AUC = 0.823), achieving an overall SN of 96%. More importantly, the SN of the classifier in non high-risk NMIBC group was 94%. Notably, the SN of the classifier in the subset of recurrent tumors (SN = 92%; n = 72) was comparable to that of pri-mary tumors (SN = 100%; n = 62).
On the other hand, the SP of the classifier decreased to 18% in this validation set, as evidenced by AUC value (Table II). Interestingly, we found that the performance of the 8-gene classifier was not affected by number of tumors (P = 0.43) nor by tumor size (P = 0.34). Fig 4 depicts the risk probabilities derived from the 8-gene classifier in NR-PFBC, non high-risk NMIBC and high-risk NMIBC+MIBC, both in training and testing sets.
In this phase, we oversampled patients with BC. We therefore cannot directly calculate predictive values from the test results. In daily practice, it is known that recurrence is detected in approxi-mately 10% of all FU cystoscopies performed (90% of patients previ-ously diagnosed with NMIBC are without recurrence at time of FU cystoscopy). In order to calculate predictive values that reflect values in real clinical practice, we assumed the distribution of recurrent vs nonrecurrent to be 10% vs 90%. For this, we multiplied the NR-
Translational Research Montalbo et al 79
Non High-risk NMIBC
Non High-risk NMIBC
C, NR-PFBC, High-risk NMIBC and MIBC
Fig 2. (A) Heat map representation of the top 50 differentially expressed genes between low-grade NMIBC and the other groups from the discovery set. Red pixels correspond to an increased abundance of mRNA in the urine samples, whereas green pixels indicate decreased mRNA levels. (B) PCA representation of differences between all samples.
Abbreviations: C, controls without neoplasic urological disease; NMIBC, Non-muscle invasive bladder cancer; NR-PFBC, non-recurrent control patients in follow-up for BC. (Color version of figure is available online.)
PFBC samples by 10. Thus, the positive and negative predictive value (NPV) in the validation phase become 66% and 97%, respec-tively. If NMIBC patients with a negative classifier do not undergo cystoscopy, this means that »17% of all cystoscopies can be pre-vented at the cost of 2% of recurrences remaining undiagnosed.
Performance comparison of 8-gene classifier and urine cytology. Overall cytology results were available for 198 patients in training and 209 in testing set. Performance of cytology in detecting BC is shown in Table II. In both training and testing set, SN of the 8-gene classifier (80% and 96%, respectively) was higher than that of urine cytology (35% and 43%, respectively) in our cohort of samples. NPV is also higher in both training and testing sets for the 8-gene classifier than for urine cytology (Table II). This means, in the case of the testing set, that 53% of the tumors (51 out of 96) that were missed by urine cytology were detected by the 8-gene classifier. Notably, the classi-fier only misses four tumors (all were non high-risk NMIBC) which also were not detected by cytology. On the contrary, positive predic-tive value and SP are higher for urine cytology than for the 8-gene classifier (Table II).