Sera were previously collected from male volunteer blood donors without cancer (n=15, controls), or patients with prostate cancer (n=85). Sera from patients were grouped according to stage of disease: newly diagnosed localized prostate cancer (new Dx, n=15), castration-sensitive non-metastatic prostate cancer (nmCSPC, n=40), castration-resistant non-metastatic prostate cancer (nmCRPC, n=15) and castration-resistant metastatic disease (mCRPC, n=15). Sera were also collected serially from the individuals with nmCSPC, who were enrolled on clinical trials in which 20 patients were treated with standard ADT (gonadotropin-releasing hormone analog given every 3 months)20 and the other 20 were treated with an investigational antitumor DNA vaccine encoding prostatic acid phosphatase (PAP; pTVG-HP, with granulocyte-macrophage colony-stimulating factor co-delivered as a vaccine adjuvant, given every 14 days for 6 administrations).21 Sera were collected at baseline, and at 3 months and 6 months following initiation of treatment for these patients. All samples were stored between −20°C and −80°C until use for analysis.

Gene products from 1463 of the most highly expressed transcripts in prostate cancer22 23 and 148 predicted ORFs in prostate cancer were selected for inclusion on the array (online supplemental table 1). Gene products included 125 antigens previously identified as recognized by IgG from patients with prostate cancer.24 The potential ORFs were selected from a list by Iyer et al of long RNAs with in silico evidence of coding potential.25 There were 74 transcripts designated as having a ‘Cancer Association’, ‘prostate’ tissue association and category of ‘tucp’ (transcript of unknown coding potential).26 ORFs were then predicted using EMBOSS: getorf, with the top two longest ORFs for each long RNA included on the microarray.

Peptide synthesis was performed as previously described, using a light-directed array synthesis in a Roche Nimblegen (Madison, Wisconsin, USA) maskless array synthesizer.27 Cycles of amino acid coupling were repeated until 16-mer peptides were synthesized on arrays containing 12 replicates of 177,604 peptides per subarray. Sera were diluted 1:100 with binding buffer (0.1 M Tris, 1% alkali-soluble casein, 0.05% Tween-20), incubated overnight at 4°C and washed. IgG was detected using an Alexa Fluor 647-labeled antihuman IgG secondary antibody (Jackson ImmunoResearch Labs, West Grove, Pennsylvania, USA). After final washing, arrays were dried and read using a Roche MS 200 microarray scanner, and signals were extracted using Roche internally developed software. Fluorescent signals were converted into arbitrary units (AU) with intensity plots ranging from 0 to 65 000 AU. Spatial correction, background correction and quantile normalization were performed on raw array signal intensities by Roche as previously described.28 All samples were evaluated in triplicate on separate arrays. Samples were considered positive for an antibody response at a given probe if the signal crossed 2¹² fluorescence units, with a sliding scale p value <0.05 in at least two of three technical replicates.28 A binding buffer only control was also run to confirm the absence of signal above the 2¹² threshold.

Data analyses were performed in R V.3.6.229 and RStudio30 using many available extension packages and visualization tools as well as custom scripts. To support reproducibility, workflow details are supplied in an R markdown document and the rendered online supplemental statistical. These materials are also available at: https://github.com/wiscstatman/immunostat-prostate

To characterize antibody responses to a wide variety of prostate cancer-associated proteins in patients with prostate cancer, we designed a peptide microarray able to be screened with patient sera. This array included peptides spanning the amino acid sequences of 1463 of the most abundantly expressed gene products in metastatic prostate cancer,22 23 including 125 proteins identified in previous studies examining serum antibody responses in patients with prostate cancer.24 39 We also included peptides spanning the predicted amino acid sequences of 148 potential open reading frames (ORFs) from lncRNAs that have been shown to be highly expressed in prostate cancer. We included these given their strong association with prostate cancer. While most would likely serve as negative controls as they would not be expected to encode gene products, other groups have shown that some lncRNAs may be translated into unstable peptides or even functional proteins, especially with the dysregulation induced by cancer.40–42 Hence, we reasoned that a few might serve as antibody targets in patients with prostate cancer.

The 16-mer peptides spanning the amino acid sequences of these 1611 gene products, and overlapping by 12 amino acids, were used to generate a microarray comprising 177 604 peptides. The complete list of probes and corresponding proteins is available in online supplemental table 1. The manufacture of the array and synthesis of peptides was performed as previously described.43 The characteristics of the proteins included in the array are summarized in figure 1, using data retrieved from UniProt.38 Sixty-nine per cent of proteins included were those typically localized within the cytoplasm or nucleus, or that traffic between the two compartments (figure 1A). Approximately 6% of the proteins were localized to the ribosomes. The median protein length was 483 amino acids (figure 1B).

We next assayed serum samples collected from controls and patients with different stages of prostate cancer for peptide-specific IgG responses using the microarray. Examples of the primary data are shown in online supplemental figure 1A,B. To assess the reproducibility of the assay, we calculated Pearson’s correlation coefficents between each pair of technical replicates and found high correlation on average among replicates (figure 1C). To determine the degree of variability among serum samples, we calculated the mean correlation coefficient across all pairs of distinct serum samples. We observed low correlation between the average pair of serum samples (figure 1C). In a complementary approach, we fit a linear mixed-effects model to estimate the amount of biological variation and technical variation across our triplicate data for each peptide. We found that the average ratio of biological variation to total variation was 0.74, indicating low technical variation (figure 1D).

Included in this study were six patients who had serum collected at two different time points, when they had an early stage of disease and again when they had a later stage of disease. Notably, these serum samples from the same patients had especially high correlation coefficients (figure 1E). This suggests that while there is high variation among individuals, each particular individual had smaller variation in his antibody repertoire over time. These six patients had their first serum collection removed from further analysis to prevent inflating their impact on our results.

To determine whether the array could detect IgG to common prostate antigens, we first defined a ‘positive’ antibody response to individual peptides using previously described criteria.28 Using binding buffer as a negative control, no peptides met these criteria (not shown). Two examples of positive responses are shown in figure 2A. We specifically evaluated responses to peptides derived from well-defined prostate target antigens PSA, PAP and the androgen receptor (AR). Overall, 7.1% of patients with prostate cancer (13.3% of patients with mCRPC) assayed on the array displayed antibody responses against peptides derived from PSA, while 6.7% of controls had PSA responses; 8.2% of patients with prostate cancer (13.3% of patients with mCRPC) and 0% of controls had responses to PAP. Finally, 5.9% of patients with prostate cancer (13.3% of patients with mCRPC) and 20.0% of controls recognized peptides derived from the ligand-binding domain of AR. Given the small sample sizes, none of the antibody responses to these proteins was significantly different in frequency in patients with cancer compared with controls.

We next tested the hypothesis that patients with higher disease burden would recognize more peptides, potentially due to increased presentation of cancer-associated proteins.44 We found no correlation between stage of disease and the number of probes recognized at either the peptide level or the protein level. The median numbers of proteins recognized were 321 for controls, 303 for new Dx, 353 for nmCSPC, 249 for nmCRPC and 320 for mCRPC (figure 2B,C). The median numbers of peptides were 919 for controls, 832 for new Dx, 712 for nmCSPC, 708 for nmCRPC and 754 for mCRPC. We noted a substantial amount of heterogeneity in antibody responses among patients. For instance, the number of proteins recognized by controls ranged from 188 to 922. Similarly, we did not observe an association between subject age and number of proteins recognized (data not shown).

Having established that there was a large diversity in antibody repertoires among patients, we next examined whether there were any broad trends in the types of proteins that were recognized. While only 0.4% of calls were positive overall, 20% of peptides were recognized by at least one subject. Nearly all proteins (1570 of 1611, 97%) had one or more peptides recognized by at least one subject. Conversely, there were no proteins that were recognized by all patients. Most proteins (1326 of 1611, 82%) were recognized by both controls and patients with cancer (figure 3A). As expected, one of the largest categories of proteins that were not recognized were ORFs from lncRNAs (figure 3B, online supplemental table 2); however, contrary to our expectations, the majority of lncRNAs (141 of 148, 95%) were recognized by at least one patient (figure 3A).

We hypothesized that while the overall number of proteins recognized may not increase with burden of disease, the composition of proteins recognized may be different. We employed a KW test to identify peptides that had significantly different fluorescence intensities across clinical stages and controls. This test identified 13 279 significant peptides (online supplemental table 3). We used principal component analysis (PCA) to visualize the residual fluorescence levels after subtracting the grand mean fluorescence level for each peptide and observed that patients tended to group with other patients with the same clinical stage of disease (figure 4A). Patients with castration-resistant tumors, and mCRPC in particular, tended to cluster especially closely to one another. Notably, controls did not exhibit this clustering. We were particularly interested in the subset of peptides that had significantly different fluorescence signals in patients with cancer compared with controls. We identified these peptides by using a Wilcoxon rank-sum test and specifically focused on those that had differences in median fluorescence of at least twofold in patients with cancer compared with controls (figure 4B, left; online supplemental table 4). To discover which peptides were driving the especially strong clustering of patients with mCRPC, we repeated this procedure to find peptides with significantly different fluorescence in patients with mCRPC compared with all other patients (figure 5B, right; online supplemental table 5). Unexpectedly, we detected only 110 peptides associated with the cancer versus control comparison, but found 4246 peptides in the mCRPC versus all other comparisons.

We applied this same approach to identify the number of peptides that had significantly higher or lower signals in patients in one clinical stage of disease compared with patients in the previous clinical stage. The largest change in number of recognized peptides occurred between the castration-sensitive (nmCSPC) and castration-resistant (nmCRPC) populations (figure 4C; online supplemental tables 6–9). Examples of the fluorescence signals of peptides that are detected by this strategy are shown in figure 4D,E.

From our initial list of 13,279 peptides, we identified 6708 of these peptides that were significant in one of the six comparisons made in figure 4B,C. We visualized the residual fluorescence levels of these peptides after removing the grand mean for each peptide in figure 5A. As in figure 4A, we observed high similarity in antibody profiles between patients with the same stage of disease. We next more closely examined the sets of proteins we had identified earlier for common features and associations with cellular processes. GO analysis revealed that the genes corresponding to the 68 peptides that were recognized more robustly in patients with cancer compared with controls were associated with mRNA export from the nucleus and the cell-cell contact zone (figure 5B). GO analysis of the 3123 peptides that had particularly strong antibody responses in patients with mCRPC showed an enrichment for proteins associated with nucleic acid binding, RNA metabolism, gene regulation and downregulation of metabolism (figure 5C). One of the significant terms within the ‘non-membrane-bounded organelle’ term was the cytosolic large ribosomal subunit. To investigate the large difference in antibody repertoires between patients with nmCSPC and nmCRPC, we performed GO analysis on the 2612 peptides with significantly higher signal in nmCRPC than nmCSPC. We identified differences in recognition of proteins associated with nucleic acid binding, chromatin structure, amide metabolism and protein localization to the membrane (figure 5D).

Based on our finding that individual patients tended to have relatively small variation in their antibody responses over time, we hypothesized that this could make the microarray particularly sensitive for detecting changes induced by treatment in a longitudinal analysis. To test the potential of this platform for studying treatment effects, we used the serial serum samples available from the 40 patients with nmCSPC who were treated with either ADT or an investigational DNA vaccine. Consistent with our observations in figure 1E, we found high correlation between samples from an individual patient over time (figure 6A).

We next fit a linear mixed-effects model to determine if there were any peptides against which there was increased signal over time. In the vaccine-treated patients, we found 5680 significant peptides that had a coefficient of time fixed-effect of at least 0.3333, indicating a twofold increase in signal every 3 months (online supplemental table 10). We were unable to detect any peptides against which ADT-treated patients developed increasing antibody signal over time using this procedure (figure 6B). Examples of the fluorescence levels of 3 peptides over time in ADT-treated and vaccine-treated patients are shown in figure 6C.

We visualized the changes in peptide recognition over time in vaccine-treated patients by plotting the residuals of the null model in the heatmap in figure 7A. This further demonstrated that vaccine-treated patients had robust increases in antibody responses to these 5680 peptides. To characterize these peptides, we performed GO analysis. We found that a significantly enriched set of these antibodies were specific to nucleic acid binding proteins. There were also more antibodies against proteins associated with RNA metabolism, ion binding and ribosomal or nucleolar cellular components than would be expected by chance (figure 7B).