CLEAVR - UNDERSTANDING THE GNRHR / TNBC ANALYSIS A plain-language walkthrough of every concept and every number This document explains, from the ground up, what was measured, what each statistical test actually does, and what every number in the results means. No prior stats background assumed. PART 1: THE BIOLOGY What is TNBC? Breast cancer is often classified by three receptors on the tumor cell surface: estrogen receptor (ER), progesterone receptor (PR), and HER2. Most breast cancers are positive for at least one of these, which matters because there are good targeted drugs for each. "Triple-negative breast cancer" (TNBC) means a tumor is negative for all three - so those targeted drugs don't work, and treatment options are more limited. That's the clinical reason TNBC is a hard problem and why new targeting strategies, like LHRH-conjugated nanoparticle constructs, matter. What are "Lehmann subtypes"? Even though all TNBC tumors share the "triple-negative" label, they are not biologically identical. In 2011, a researcher named Lehmann and colleagues analyzed gene expression across hundreds of TNBC tumors and found they cluster into distinct molecular groups, each with different biology and different drug sensitivities. The subtypes are usually named: BL1 - basal-like 1 (proliferation/cell-cycle genes turned up) BL2 - basal-like 2 (growth factor signaling turned up) M - mesenchymal (genes for cell motility/invasion turned up) MSL - mesenchymal stem-like IM - immunomodulatory (immune-related genes turned up) LAR - luminal androgen receptor (driven by the androgen receptor pathway, more hormone-like biology than the others) This analysis only modeled four of the six (BL1, BL2, M, LAR) - see Part 3 for why. What is GNRHR, and why does it matter here? Every gene has a symbol. GNRHR is the gene that contains the instructions for building the LHRH receptor (also called the GnRH receptor) - the actual protein sitting on the surface of a cell that LHRH (a hormone) or an LHRH-mimicking drug can grab onto. LHRH-conjugated nanoparticle constructs are decorated with LHRH (or triptorelin, a synthetic LHRH-like molecule) specifically so they stick to and get pulled into any cell displaying this receptor. If a tumor cell doesn't have much of this receptor on its surface, the nanoparticle has less to grab onto. Important distinction: GNRH1 vs GNRHR GNRH1 is a different gene - it contains the instructions for the hormone itself (the key). GNRHR contains the instructions for the receptor (the lock). LHRH-nanoparticle constructs supply their own synthetic key (LHRH/triptorelin on the nanoparticle surface) and are looking for the tumor cell's lock. So the receptor gene, GNRHR, is the one that matters for this targeting strategy - this analysis measures GNRHR specifically, not GNRH1. What does "GNRHR expression" mean, numerically? Cells don't just have or not have a gene - genes get "read" and turned into protein at different rates in different cells. RNA-sequencing measures how often a gene's instructions are being actively read in a given tumor sample, producing a number (here, from an algorithm called RSEM) that's roughly proportional to how much of that receptor the cell is likely making. Higher number = more of the gene's message being produced = generally more receptor on the cell surface. This is a proxy, not a direct photograph of receptor protein count on the cell surface, which is worth remembering. PART 2: THE STATISTICS, IN PLAIN TERMS What is a p-value? A p-value answers one narrow question: "if there were truly no real difference between these groups, how likely is it we'd see a difference this big just by random chance in our sample?" A small p-value (conventionally below 0.05) means the pattern is unlikely to be random noise. A large p-value (like the 0.71 and 0.98 in this analysis) means what we observed is entirely consistent with there being no real difference. It does NOT prove there's no difference - it means we don't have evidence of one in this dataset. Why Kruskal-Wallis instead of a simple average comparison? The straightforward approach - compare average GNRHR expression across the four subtypes - assumes the data is shaped like a normal bell curve. Gene expression data is usually skewed (a few very high outlier values pull the average around), so instead this analysis used the Kruskal-Wallis test, which converts every value to a rank (1st lowest, 2nd lowest, etc.) and compares the ranks across groups instead of the raw numbers. This makes it more robust to those outliers and skew. The result was H = 1.38, p = 0.71: essentially, no detectable difference in GNRHR ranks across the four subtypes. What are the Mann-Whitney tests? After the four-way Kruskal-Wallis test, this analysis also ran six follow-up tests, one for each possible pair of subtypes (BL1 vs BL2, BL1 vs M, and so on), using the two-group version of the same rank-based idea (Mann-Whitney U). Every single pair came back non-significant (p > 0.3). What's this "Bonferroni correction" mentioned in the limitations? If you run six statistical tests instead of one, you have six chances to get a false positive by pure luck - so the more tests you run, the more you should raise your bar for what counts as "significant." Bonferroni correction is the simplest way to do that: multiply each p-value by the number of tests (6 here). When that's applied, every pairwise p-value becomes 1.0 - confirming there's no real pairwise signal being missed. This is a good-practice check to run whenever you see multiple p-values reported side by side in any study, including published ones. What is a "median split," and why is it a weaker approach? GNRHR expression is a continuous number - it doesn't come in "high" and "low" boxes naturally. To ask "does expression relate to survival," patients were sorted by GNRHR expression and split into the top half ("high") and bottom half ("low") at the median. This is simple and interpretable but throws away information - two patients right next to each other on either side of the midpoint get treated as categorically different, and the specific cutoff point is arbitrary. It's a common quick-analysis technique, but a more rigorous version would model GNRHR as a continuous variable directly (a Cox proportional-hazards model) rather than splitting it in two. What is a Kaplan-Meier curve? It's a way of plotting "percent of patients still alive" over time, for a group of patients, that correctly handles patients who left the study partway through (called "censoring" - e.g., they were still alive when the data was last collected, so we know they survived at least that long, but not how much longer). The plot in this project shows two such curves - one for GNRHR-high patients, one for GNRHR-low - and they sit almost exactly on top of each other, visually showing there's no separation between the groups. What is the log-rank test? It's the standard statistical test for asking "are these two Kaplan-Meier curves actually different, or could they plausibly be the same underlying survival pattern?" Here it returned p = 0.98 - about as clean a "no detectable difference" as a statistical test can report. Why is the survival result flagged as "underpowered" rather than a strong negative? Of the 115 patients with survival data, only 18 had died by the time TCGA recorded outcomes (a 15.7% event rate). Statistical tests on survival data get their power mostly from the number of events (deaths), not the number of patients - 18 events is a small number to detect anything but a very large effect. So "no significant difference" here means "we didn't find a strong signal with the limited events available," not "we've proven there's no relationship." This distinction matters and is worth stating explicitly if asked. PART 3: WHAT THIS ANALYSIS CAN AND CANNOT CLAIM Can claim: In this specific cohort of 115-116 TCGA TNBC patients, GNRHR (receptor) expression does not show a statistically detectable difference across four marker-gene-defined TNBC subtype groups, and does not show a statistically detectable relationship with overall survival. Cannot claim: That GNRHR expression is definitely uniform across true Lehmann subtypes (the subtyping here is an approximation, not the validated classifier - TCGA's Firehose Legacy dataset doesn't include real Lehmann or PAM50 calls). Cannot claim survival is definitely unrelated to GNRHR (too few events to rule that out confidently). Cannot claim anything about receptor protein levels, receptor accessibility, or nanoparticle uptake specifically - this measures gene transcription, several biological steps removed from "how much LHRH-nanoparticle would actually stick to this cell." Why this is still useful, scientifically: null results with clearly stated limitations are legitimate science, not a failure to find something. The honest, well-caveated version of this finding is more credible in a scientific conversation than an overstated one - and it naturally sets up a better question ("does receptor trafficking or downstream signaling differ by subtype even when raw transcript levels don't?") than a false positive would have. PART 4: EVERY NUMBER, LOCATED Cohort size: 116 TNBC patients total, 115 with usable GNRHR expression data (1 patient has no primary-tumor RNA-seq sample in TCGA at all - not an error). Subtype group sizes: BL1 n=44, M n=42, LAR n=15, BL2 n=14. Kruskal-Wallis: H = 1.38 (the test statistic - larger means more separation between group ranks), p = 0.71 (not significant). Group medians (RSEM units): BL1 = 2.55, BL2 = 2.89, M = 3.08, LAR = 4.55. Pairwise Mann-Whitney p-values: all between 0.31 and 0.81 (all non- significant; all become p = 1.0 after Bonferroni correction). Median GNRHR value used for the high/low survival split: 2.88. Survival comparison: 58 patients "high," 57 "low," 18 total deaths across both groups, log-rank p = 0.98. All of the above were independently recomputed from the raw TCGA files in a second pass, separate from the original run, and matched exactly.