Q (sonnet-5.5): Is KRAS G12C more aggressive than G12D? ▶ read_guide { "uri": "cbioportal://statistical-tests-guide" } ▶ read_guide { "uri": "cbioportal://mutation-frequency-guide" } ◀ result Statistical Test Selection Guide ================================ Purpose ------- This guide ensures the correct statistical test is selected before performing any group comparison, matching cBioPortal's own Group Comparison defaults. HARD RULES — NEVER FABRICATE A STATISTIC ---------------------------------------- ClickHouse cannot run statistical tests. The agent therefore must NEVER produce a derived statistic that is not a literal column value from a SQL result. Specifically: 1. **Never invent a p-value.** Not "p < 0.001", not "p ≈ 0.05", not any p-value. If the user asks "what is the p-value?", the answer is *"I can't compute that — here is the 2x2 contingency table (or group statistics). Run it in cBioPortal's Group Comparison tab, in R with `fisher.test(...)` / `wilcox.test(...)`, or in Python with `scipy.stats.fisher_exact(...)` / `mannwhitneyu(...)`."* 2. **Never claim mutual exclusivity (or co-occurrence) from a contingency table alone.** A 2x2 table is not a test. The shape "altered/not altered × group A/group B" needs Fisher's exact + a defined direction (odds ratio < 1 with significant p). Without that test, the agent presents the table and stops. Descriptive phrasing is also a claim: "largely/mostly mutually exclusive", "rarely co-occur", "tend to co-occur" are forbidden without the test — point to cBioPortal's Mutual Exclusivity tab instead. 3. **Never report a "median" that came from `AVG(...)` or any non-median aggregate.** "Median" and "mean" are different statistics; for skewed clinical distributions (especially survival) they differ substantially. Use ClickHouse's `quantile(0.5)(...)` for actual median, and label arithmetic averages as "mean", never "median". 4. **Never report a hazard ratio, odds ratio, risk ratio, or relative risk** that wasn't computed by an external tool. These require regression / model fitting that ClickHouse does not do. 5. **Never report median overall survival from `AVG(OS_MONTHS)` or even `quantile(0.5)(OS_MONTHS)`.** Median OS requires Kaplan-Meier estimation, which handles censoring (`OS_STATUS = 0:LIVING` means the event hasn't happened yet). Naive medians/means over `OS_MONTHS` ignore censoring and are systematically wrong. The correct handoff: return the raw `(OS_MONTHS, OS_STATUS)` pairs (or descriptive counts: N events, N censored, follow-up range) and tell the user to run KM in R (`survival::survfit`) or Python (`lifelines.KaplanMeierFitter`), or use cBioPortal's Survival comparison. If asked to do any of (1)–(5), respond with the appropriate handoff template from the "Approved Response Templates" section below — do not produce the number. Ambiguous-Term Clarification (before any analysis) -------------------------------------------------- If the user's question uses an ambiguous term, ask before computing anything. Examples: - **"aggressive"** → could mean shorter overall survival, higher metastasis rate, higher grade/stage, higher TMB, or specific molecular features. Ask which definit … (9984 more chars) ◀ result # Mutation Frequency Analysis Guide ## IMPORTANT: Reporting Mutation Frequencies - **ALWAYS report frequencies as percentages**, not raw counts: `frequency = (altered_samples / total_profiled_samples) × 100` - For quick frequency lookups, **prefer the TCGA Pan-Cancer Atlas study first**, then offer to expand to other studies - When reporting across multiple studies, show **ranges** (e.g., "TP53 is mutated in 30–60% of samples") rather than a single average - **NEVER** sum mutation events across studies to compute an aggregate frequency — this can exceed 100% due to double-counting - Warn users that samples may overlap across cohorts (e.g., MSK studies may share patients) - **Choose and state the counting unit**: use patient-level frequencies for prevalence/rate questions unless the user explicitly asks for samples; use sample-level frequencies when the user asks about samples. - **For "across cancer types" questions**, jump to the [Cross-Cancer-Type Mutation Frequency](#cross-cancer-type-mutation-frequency) section below — there is one correct recipe and several common wrong ones. ## Counting Unit: Samples vs Patients Before answering any mutation count or frequency question, decide whether the unit is samples or patients and state that choice in the answer. | User wording | Counting unit | |--------------|---------------| | "prevalence", "rate", "fraction of patients", "patients with", "how common is" | Patient-level: `COUNT(DISTINCT patient_unique_id)` | | "samples", "specimens", "biopsies", sample-level cohort composition | Sample-level: `COUNT(DISTINCT sample_unique_id)` | | Ambiguous | Ask, or default to patient-level for prevalence/rate language and say so | ### Cross-study sample-count caveat When an answer touches more than one study and reports a sample count, prepend a one-line caveat: > Sample IDs are unique within cBioPortal study prefixes, not guaranteed biological-sample identifiers across studies; overlapping cohorts can count the same patient/sample more than once. Prefer one of these safer approaches: - Use a shipped `cancer_study_query_preferences` cohort such as `pan_cancer_tcga` or `all_studies_non_redundant`. - Restrict to one named study. - Aggregate by `patient_unique_id` when the biological question is patient prevalence. ## STOP rule: a frequency above 100% means your query is wrong If your query returns a frequency over 100%, **do not try to debug or explain the data inconsistency to the user**. The cause is always one of these query bugs: - Summing mutation events instead of `COUNT(DISTINCT sample_unique_id)` for the numerator - Using a study-wide sample count as the denominator instead of the gene-specific profiled count - Cross-study aggregation where the same biological sample appears under multiple `sample_unique_id` values (e.g., MSK-IMPACT and MSK-CHORD share patients) - **Joining the profiled CTE through `gene_panel` / `gene_panel_list` without a WES branch.** `gene_panel_id = 'WES'` is *not* a row in … (36467 more chars) ═ answer (success) "Aggressive" can mean several different things, so I need to know which one you mean before I compare G12C and G12D: - **Shorter survival** (OS, PFS or DFS). This needs Kaplan–Meier curves and a log-rank test, which I can set up for you. - **More metastatic disease**, meaning a higher rate of metastatic vs. primary samples or a higher rate of metastasis at presentation. - **Higher grade or stage** at diagnosis. - **Higher tumor mutation burden** or other molecular features. I also need to know which cohort to use. KRAS G12C and G12D mean different things in different tumor types. G12C is enriched in lung adenocarcinoma, and G12D is common in pancreatic and colorectal cancer. A pooled comparison would mostly reflect tumor-type mix. Pick one: - **Lung adenocarcinoma**, e.g. TCGA PanCancer Atlas LUAD, or a larger cohort such as MSK-CHORD. - **Pancreatic or colorectal cancer**, where G12D is common and G12C is rare, so the comparison is underpowered. - **A pan-cancer cohort** such as MSK-CHORD, with cancer type as a covariate. Once you tell me the outcome and the cohort, I'll pull the group sizes and event or censoring counts for each variant. I can't compute p-values or hazard ratios myself, but I can link you to a Group Comparison in cBioPortal that does.