No abstract

Clinical care increasingly relies on individualized, highly multidimensional data for patient phenotyping. This review focusses on two technologies which are contributing to this trend and, in their mutual overlap, may provide a novel and powerful contribution to future clinical care. Artificial intelligence (AI), in particular the technique of deep learning, is steadily making inroads to clinical care and has demonstrated the ability to extract clinically relevant features from complex datasets. Likewise, mass spectrometry (MS), in particular high-resolution MS, is gaining the ability to provide highly multidimensional datasets from patient samples with the quality necessary for clinical decision making. The convergence of these trends has resulted in the nascent application of deep learning analytics to high resolution mass spectra. Due to the large volumes of content provided by high resolution mass spectra and the ability of multilayer neural networks to extract subtle features from complex datasets, future integration of their mutual application is identified as a research area of high priority for clinical care. Factors are identified that have contributed to historical progress for each technology and recommendations are provided to accelerate the mutual application of deep learning and mass spectrometry techniques to patient care.

249 Background: A 22-gene GC has been proposed to refine risk stratification of localized PCa by conventional NCCN criteria, and this may potentially influence treatment recommendations. Nonetheless, majority of studies looking at the utility of GC were conducted in White and non-White men from Western cohorts. We therefore investigated the association of GC with NCCN risk groups (RG) in an Asian PCa cohort. Additionally, we examined for inter-racial differences in molecular subtyping between Asian and White/non-White PCa. Methods: GC (Decipher Biosciences Inc., CA) was performed on diagnostic biopsies of men who were treated with radiotherapy +/- hormonal therapy at a single institution (N = 75). ISUP Gleason’s grade (GG) and tumor cellularity were reviewed by an expert GU pathologist. RNA was extracted from 2 x 2.0-mm tumor cores using Qiagen AllPrep DNA/RNA FFPE Kit (Qiagen, Germany) and gene expression was performed on Affymetrix Human Exon 1.0 ST Array (ThermoFischer, CA). PAM50 molecular subtyping was derived using the DecipherGRID database. Results: We profiled 80 tumors from 75 patients, comprising of 18 (24.0%), 9 (12.0%), 21 (28.0%), and 19 (25.3%) NCCN low-/favorable intermediate-, unfavorable intermediate-, high- and very high-RG, respectively; of note, 8 (10.7%) patients had regional/metastatic disease at diagnosis. Using the GC, 27 (33.8%), 14 (17.5%) and 39 (48.8%) were classified as low- (<0.45), intermediate- (0.45-0.6) and high-RG, respectively (>0.6). When stratified using a three-tier clinico-genomic (CG) classification system (Spratt et al. 2017), 6 of 21 (28.6%) NCCN-defined high-risk and 4 of 19 (21.1%) very high-risk patients were downgraded to CG-defined intermediate-/low-risk, while 2 of 27 (7.4%) NCCN low-/intermediate-risk patients were in fact upgraded to CG high-risk. Next, we interrogated the PAM50 basal-luminal signature in our cohort. Interestingly, when matched to White (N = 5762) and non-White (N = 155) for NCCN RG, ISUP GG and age, we observed a high proportion of basal subtype (62.7%) in Asians, which contrasted the prevalence observed in White (16.7%) and non-White (15.9%) North American patients (Chi-sq P <0.001). Conclusions: Here, we demonstrated the utility of the 22-gene GC for refining the NCCN risk stratification in a largest Asian PCa dataset to-date. An unexpectedly high proportion of PAM50 basal-subtype was observed, suggesting race-specific differences of the tumor transcriptome.