It should be emphasized, however, that in the vast majority of FOVs, the ERAs recognize true positivity from true negativity. It should be remembered that the purpose of the ERAs and FOVs is to divide a slide into grids and select the areas with the highest epithelial percentage to guide TMA construction. Even if some grids show some degree of false negativity or false positivity, this would not change the overall effectiveness of the ERAs in guiding TMA construction because most of the FOVs would be interpreted accurately by the ERAs.Discussion
The main benefits of both algorithmic and robotic TMA in guiding TMA construction are increased efficiency and the removal of human error. Since manual TMA reflects the current standard, the advantages of both algorithmic TMA and robotic TMA should be determined based on a comparison with this standard.
Manual TMA presently involves retrospective manual selection of cases which require archival slide retrieval, re-review, manual circling of regions of interest, and alignment with the original tissue block from which the TMA is constructed. This process is quite time consuming, taking weeks to months.
We wish to point out that the main issue herein was not to test whether the algorithmic TMA or robotic TMA is superior to manual TMA in producing a higher density of epithelial cells. Our focus was whether TMA construction could be automated and whether a robot could replace a human being. Indeed, even if both automated TMA and robotic TMA yielded equivalent results to manual TMA in terms of epithelial cell density, the automated approach would still be highly advantageous.
Using this algorithm-driven technology, glass slides of potential cases which are candidates for future TMA construction can be prospectively scanned and made into a library of virtual slides. These virtual slides can then be analyzed with our ERAs and FOV algorithms and the regions mapped for TMA production. Slide scanning and algorithmic analysis each take approximately 15 min per slide, but if it is done at the time of routine pathological slide review, it can create a library of virtual slide images and identify prospective cancer regions during the course of pathologists’ everyday work load. This information then can be used to guide construction of either algorithmic TMA, or robotic TMA with TMAker. Overall, scanning and algorithmic analysis leading to either algorithmic TMA or robotic TMA is presently much faster and efficient than for manual TMA, reducing the overall time from weeks or months to hours (11,12,13).
The present study demonstrated that both algorithmic TMA and robotic TMA are superior to manual TMA in identifying areas of the slide with a higher epithelial percentage and, therefore, a higher cancer cell density. This observation held across different cancer types (breast, colon, and lung). However, the overall significance of this observation is unclear. Although it cannot be presumed that more is necessarily always better, the presence of a greater epithelial percentage may improve signal detection, and that this is where the automated technology platform particularly excels compared with manual methods.
Furthermore, if one remembers that automated TMA construction is only the first step in ultimate TMA analysis where ERAs and subsequent specific-recognition algorithms (SRAs) will be applied to analyze immunoreactivity in tumor cells but not stroma, certainly cores richer in tumor cells will be more easily analyzed and either false negatives or false positives from stromal contamination will be minimized. This is the reason that the increase in cancer cell density by the automated approach is desired. However, in order to generally show that molecular data obtained on epithelial cell–enriched tissue arrays are more representative or clinically relevant than manual arrays, we first must produce and then analyze these arrays algorithmically.
Although we have not demonstrated that increased epithelial percentage adds value from the standpoint of added clinical information, from another perspective increased epithelial percentage does add value. In general, grids with the very highest epithelial percentages are likely to contain invasive cancer because normal epithelium—especially in the breast, colon, and lung—exhibits a much lower epithelial percentage. These higher epithelial percentages are the very reason why the algorithmic approach is desired. Our imaging algorithms will not directly distinguish carcinoma in situ from invasive carcinoma; however, increased epithelial percentages can indirectly help distinguish these areas. Those grids with the highest epithelial percentages are more likely to contain invasive carcinoma than carcinoma in situ because usually carcinoma in situ is surrounded by a higher percentage of non-epithelial stroma.
TMAs are made and interpreted manually by a time-consuming process (14,15,16). In contrast, analogous cDNA microarrays and genomic libraries are made and interpreted rapidly, because both construction and interpretation are computer-driven (17,18). cDNA microarray technology is highly automated and high-throughput with relatively manageable biological material. The largest cDNA arrays now spot over 2 million features per slide, and it is possible that DNA microarrays may eventually be replaced with high-throughput sequencing in the near future. TMAs, on the other hand, have not yet advanced to this stage. Unless fully automated in both construction as well as interpretation (as in robotic TMA), TMA technology will never approach the high-throughput profiling capabilities of cDNA microarrays.
Manual TMAs have traditionally been limited to 1–3 cores per sample (7), although the optimal number of cores to reduce limited sample bias is still controversial. The optimal number of cores and the optimal epithelial percentage directly address an optimal data representation issue. This issue notwithstanding, one of the benefits of the robotic TMA approach is that the technology platform presented is flexible in allowing a variable number of cores, depending on the heterogeneity of the marker being studied. As more and more tumor biomarkers and tumor-associated signaling pathways are investigated, the varying degree of heterogeneity within tumors for various antigens and signaling pathways may require a range in the number of cores per patient sample to overcome limited sample bias. For example, Her-2/neu immunoreactivity in breast cancer is usually fairly homogeneous and one core may be sufficient; on the other hand, estrogen receptor (ER) immunoreactivity in breast cancer is more heterogeneous and three cores may be necessary. In studying selected phosphorylated transcription factors like p-ETS, even more heterogeneity may be present, requiring a significantly greater number than three cores per case. Certainly any robotic TMA construction system that can produce as many cores as desired from a given sample—for so-called “multiplex” TMAs—would be poised to determine what value overcomes limited sample bias for any given antigen. There are no real tissue limitations of multiplex TMAs. Multiplex TMA manufacturing will not completely consume the tumor. Most tumors removed from patients are made into multiple whole slides with corresponding paraffin blocks, and there are usually several blocks available. The typical slide obtained from the block contains an area of tumor that is 2 cm × 2 cm (4 cm2, or 400 mm2). A typical TMA core is 1 mm in diameter. Therefore, ~400 or more cores can be derived from one paraffin block. Even if 10 cores are produced from each block, this consumesonly 2.5% of the tumor. If the algorithm ranks the areas of cancer cell density, it would be fairly easy to select the 10 areas with the highest density without compromising or exhausting the block.
In conclusion, the use of ERAs and FOV algorithms on virtual slides to guide TMA selection and ultimately TMA construction will result in TMAs becoming truly part of biomarker high throughput screening.