# Cybergenic > An autonomous engine that reasons across cancer genomics, drugs, trials, and literature in real time, arguing cited hypotheses and showing its work as it runs. Cybergenic runs one research cycle at a time on public tumour genomics (MSK-IMPACT, MSK-CHORD, TCGA PanCancer, OrigiMed China and METABRIC, read through cBioPortal), tests gene pairs with a false-discovery-rate correction and confound controls, and publishes every finding, its verdict, and a hash-chained registry of its locked predictions. Findings are statistical signals and labelled hypotheses, not medical advice. ## Start here - [The empty frontier (essay)](https://cybergenic.im/blog/the-empty-frontier): How the engine locks its predictions and re-tests them in independent cohorts, and why its discovery frontier read empty on September 19, with a dated correction (September 24, 2026) on what the replication record and the verdicts do and do not show. - [Checking the scan against SelectSim (research note)](https://cybergenic.im/blog/co-mutation-scan-vs-selectsim): The scan's published snapshot compared pair by pair with SelectSim (Iyer et al., Nature Genetics 2026), from both methods' public files: where both call a gene pair they give it the same direction in 159 of 161 on largely shared MSK patients; SelectSim calls about 7 times as many pairs (the methods differ in their null, in which tumours they count and in which mutations); and a direction rule the comparison changed on 7 October 2026. - [EGFR and TP53 co-mutation in lung adenocarcinoma: exon 19 vs L858R, stage, and RB1 (research note)](https://cybergenic.im/blog/egfr-tp53-co-mutation-lung-adenocarcinoma): How often TP53 and RB1 alterations accompany each kind of EGFR mutation (exon 19 deletion, L858R, exon 20 insertion, uncommon mutations) in lung adenocarcinoma, in the same ten cohorts, by stage, tumour purity, smoking and ancestry, with the published figures it explains. TP53 is altered in 51% of 3,072 EGFR-mutant lung adenocarcinomas: 41% at stage 1 to 3, 65% at stage 4, as often with exon 19 deletion as with L858R. TP53 and RB1 together: 5%. - [STK11 and KEAP1 in KRAS G12C lung adenocarcinoma: how often, and why reports disagree (research note)](https://cybergenic.im/blog/stk11-keap1-kras-g12c-lung-adenocarcinoma): How often STK11 and KEAP1 alterations accompany each KRAS allele in lung adenocarcinoma, in ten cohorts (MSK-IMPACT 50K with its OncoKB calls, MSK-CHORD's other patients and eight cohorts with no MSK patient), under four definitions of an alteration, by stage, smoking and ancestry, with the published figures it explains. In 9,729 lung adenocarcinomas, 22% of KRAS G12C tumours carry STK11 and 9% KEAP1, as often as other KRAS mutations do. How KEAP1 is counted, and stage, move its share at MSK from 5% to 27%. - [SETD2 and BRAF V600 in lung adenocarcinoma, re-tested outside MSK (research note)](https://cybergenic.im/blog/setd2-braf-v600-lung-adenocarcinoma-retest): A re-test of Bandlamudi et al. (Cancer Cell 2026, the MSK-IMPACT 50K study): its figure rebuilt from the study's Zenodo files, then tested in lung adenocarcinoma cohorts with no MSK patient, by BRAF class, tumour type, mutational burden and smoking, with the clonality of the SETD2 mutations. A 2026 study of 50,000 tumours found SETD2 in nearly half of BRAF V600 lung adenocarcinomas. In eight cohorts sequenced outside MSK it holds: 19 of 34, against 2.4% of other tumours. - [CTNNB1 and TP53 in liver cancer, re-tested in six more cohorts (research note)](https://cybergenic.im/blog/ctnnb1-tp53-liver-cancer-retest): A re-test of Wu, Xing and Townsend (PLoS One 2026), with its gene-level epistasis tables rebuilt from the authors' repository. A 2026 study read the known avoidance of CTNNB1 and TP53 in Chinese liver cancer as candidate antagonism. Pooled over eight cohorts, 3,134 tumours, it holds, in viral and non-viral cancers. - [BRAF, KRAS, EGFR mutual exclusivity by tumour type (research note)](https://cybergenic.im/blog/braf-kras-egfr-exclusivity-within-tumour-types): A re-test of Vaeyens et al. (Int J Cancer 2026): its 10,956 variant-level co-mutation tables rebuilt from the authors' published patient file, every significant call re-tested within tumour types and against chance (the study's variant-pair test counts only tumours with either gene mutated, which inflates exclusivity), then its variant classes tested in 53,212 tumours of the scan's cohorts with each tumour's driver count held fixed, including MSK patients not in the study's file. - [Methodology](https://cybergenic.im/methodology): The tumour cohorts, the statistical tests, how to read the map, and the live track record, frontier audit and prediction registry, with their limits stated. - [Explore](https://cybergenic.im/explore): Search every finding, cycle, open question and queued hypothesis by gene, cancer or both (exact gene symbols, abbreviations such as PDAC or NSCLC), and download findings as CSV. - [Live map](https://cybergenic.im/): The engine at work: every finding plotted by statistical signal and efficacy, with the cycle it is running now. ## Data - [Open co-alteration dataset](https://cybergenic.im/data): Every gene pair the discovery scan tested in 59 tumour cohorts, significant or not, as one Parquet file (196,184 rows): 2x2 counts, the burden-conditional expected count, p and Benjamini-Hochberg q values, the effect of record and the scan's gate outcomes, with SHA-256 checksums. A second Parquet file, gene_groups.parquet (21,347 rows), gives each gene's share within clinical and patient groups of each cohort (breast receptor status and PAM50 subtype, primary tumour or metastasis, smoking history, sex, age at diagnosis, genetic ancestry) with each group's Fisher exact comparison against the rest and its q value: the numbers the group pages show. CC BY-NC 4.0. Also on Hugging Face as cybergenic/cancer-mutation-co-occurrence (load_dataset), DOI 10.57967/hf/11277. A lookup page over the same file, by gene, pair or tumour type: https://huggingface.co/spaces/cybergenic/cancer-co-mutation-explorer. - [Gene pair pages](https://cybergenic.im/pairs): One page per gene pair the scan keeps in at least one cohort, listed here and in the sitemap; each shows the pair in every cohort that tested it (tumours, tumours with both mutations, the expected count under the burden-conditional null and under independence, odds ratio, q, the direction where the data establish one, and the scan's gate outcome). Any tested pair answers at /pairs/-and- in lower case, for example /pairs/kras-and-stk11; the numbers are the same as the MCP lookup_gene_pair tool's. Pairs carried together by 30 or more tumours of one cohort (EGFR and TP53, APC and PIK3CA and a few hundred more) have indexed pages too, listed on /pairs, with the share of each gene's mutant tumours carrying the other. - [Fusion pages](https://cybergenic.im/fusions): One page per gene whose fusions are an established cancer driver, where 10 or more tumours of one cohort carry one (ALK, ROS1, RET, NTRK1, NTRK3, FGFR2, FGFR3, BRAF, ERG, TMPRSS2, EWSR1 and more), at /fusions/ in lower case, for example /fusions/alk: how often each tumour type carries a fusion of the gene with another gene, as a share of the tumours tested, and which partner genes the fusions join in each tumour type; and one page per named fusion carried by 10 or more tumours of one cohort (EML4-ALK, KIF5B-RET, CCDC6-RET, CD74-ROS1, TMPRSS2-ERG, FGFR3-TACC3, FGFR2-BICC1, ETV6-NTRK3, EWSR1-FLI1, DNAJB1-PRKACA and more), at /fusions/<5' gene>-<3' gene> in lower case, for example /fusions/eml4-alk. From the structural-variant calls of the scan's cohorts read through cBioPortal (one tumour per patient): the sequencing-panel cohorts (MSK-IMPACT 50K, MSK-CHORD, OrigiMed) call rearrangements in DNA, TCGA's PanCancer calls are fusion transcripts in RNA, and each page ranks the two apart. A fusion is a call joining two different named genes (antisense calls and MSK's separate RNA fusion test left out; where a call says which gene comes first, a gene counts only on the side its fusions keep); a tumour counts as tested on a panel version whose calls show it detects the rearrangement, and a panel share is a lower bound (a panel finds a rearrangement only where it tiles a gene's introns). Further named fusions include NAB2-STAT6, EWSR1-ATF1, EWSR1-WT1 and FUS-DDIT3. The tables behind the pages are downloadable as CSV (CC BY-NC 4.0) at /fusions#download and on Hugging Face (huggingface.co/datasets/cybergenic/cancer-gene-fusion-frequencies). - [Cell line pages](https://cybergenic.im/cell-lines): One page per cancer cell line DepMap sequences that Cellosaurus names and that carries a counted mutation in a cancer gene (about 1,640: A-549, NCI-H1975, HCC827, MCF-7, HCT 116, PC-9 and more), at /cell-lines/, for example /cell-lines/nci-h1975 and /cell-lines/a-549: its cancer type, the patient's age and sex and the site it was taken from, its DepMap and Cellosaurus ids, and each driver mutation with its allele fraction and the releases calling it. Mutation calls are DepMap's (release 25Q3) beside the Cancer Cell Line Encyclopedia's of 2019, both read from cBioPortal, because each misses calls the other makes (25Q3 does not call PC-9's EGFR E746_A750del or COLO 205's BRAF V600E). A line has no normal tissue sequenced beside it, so a mutation counts only when patients' tumours carry the same change (3 or more tumours of one cohort of the scan), when it is a missense change at a Cancer Hotspots residue (Chang et al. 2018), or when it truncates a tumour suppressor (one both intOGen and OncoKB's cancer gene list call so, plus genes such as SMARCA4, KEAP1 and the mismatch repair genes that intOGen labels otherwise); a call of the last two kinds that gnomAD finds in 0.1% or more of one population is left out as inherited unless ClinVar calls it pathogenic, and the HLA genes are left out. Each single-mutation and mutation class page lists the cell lines carrying it where any do (#cell-lines), each gene page its changes carried by the most lines, and each gene pair page the lines with a mutation in both genes where any have one. Calls a few bases apart on the same reads are rebuilt into the one change they make and named by Ensembl VEP (HCC2935's two frameshifts are one in-frame EGFR exon 19 deletion), and a change one release calls in five or more lines that the other sequenced and never calls is left out as a misread site. - [Cell lines by genotype, and the cell line finder](https://cybergenic.im/cell-lines/finder): Lists of the cell lines above by what they carry, each line with its allele fraction, the releases calling it and its other mutations: one page per mutation or class of mutations carried by 5 or more lines (about 190: /cell-lines/mutations/kras-g12c, /cell-lines/mutations/braf-v600e, /cell-lines/mutations/egfr-exon-19-deletions, the change written as on the mutation pages), with the genes most often mutated beside it among those lines as columns (KRAS G12C lines with and without STK11 or KEAP1); one page per gene mutated in 10 or more lines (about 180: /cell-lines/genes/kras), by change and by cancer type; and one per gene and cancer type with 5 or more mutant lines (about 300: /cell-lines/genes/kras/colorectal-cancer), listing the type's mutant lines beside its wild-type lines (no protein-changing call in the gene in either release and no homozygous deletion in DepMap's copy number), and apart from both the lines that lost the gene, those holding a call the site does not count and those whose copy number was not read. The cell line data derives from DepMap (Broad Institute) and is free for research, not for clinical or commercial use (including training machine learning or AI models other than for internal research): DepMap's terms of use apply and are reposted in full at /cell-lines/terms. The finder at /cell-lines/finder filters every line in the reader's browser by mutations it must carry, genes it must have no mutation called in and not have lost, and cancer type, and downloads the result as CSV; its query is in the address (?carries=KRAS%20G12C&without=STK11,KEAP1&type=Non-Small%20Cell%20Lung%20Cancer). - [Copy-number pages](https://cybergenic.im/copy-number): One page per established copy-number driver change carried by 30 or more tumours of one cohort: high-level amplifications of oncogenes (ERBB2, EGFR, MYC, CCND1, CCNE1, MDM2, CDK4, MET and more) and deep deletions of tumour suppressors (CDKN2A, CDKN2B, MTAP, PTEN, RB1, SMAD4 and more), at /copy-number/-amplification or /copy-number/-deletion in lower case, for example /copy-number/erbb2-amplification: in each tumour type, the share of the tumours whose copy number of the gene was read, in the tumour type's cohort with the most such tumours and in its other cohorts, from each study's discrete copy-number calls in cBioPortal (one tumour per patient, hypermutated tumours included); +2 is each caller's high-level gain, focal or of a whole arm. The sequencing-panel cohorts and TCGA's (GISTIC 2.0, which puts broad gains at that level far more often) are ranked apart; METABRIC is left out because its calls do so more often still. MTAP has a page as a drug target deleted with CDKN2A. The table behind the pages is downloadable as CSV (CC BY-NC 4.0) at /copy-number#download and on Hugging Face (huggingface.co/datasets/cybergenic/cancer-copy-number-frequencies). - [Mutation pages](https://cybergenic.im/mutations): One page per named driver change carried by 15 or more tumours of one cohort, where a tumour type the page lists counts 3 or more (about 270 pages): missense changes (KRAS G12D, BRAF V600E, PIK3CA H1047R, EGFR L858R, rarer ones such as ESR1 Y537N, AR L702H, BRAF G466V, IDH2 R172K and DNMT3A R882H, and about two hundred more), in-frame deletions and insertions (EGFR E746_A750del, ERBB2 Y772_A775dup, KIT W557_K558del), stop changes (TP53 R213*, APC R1450*) and frameshifts (NPM1 W288Cfs*12), at /mutations/- in lower case with a stop written x, for example /mutations/kras-g12d, /mutations/egfr-e746_a750del and /mutations/tp53-r213x: in each tumour type where a cohort has 30 or more tumours with the gene, the share of all tumours tested carrying the change, its share of the gene's mutant tumours and its rank among the gene's changes, in the cohort the gene page quotes for that tumour type and the tumour type's other cohorts. Counts are the scan's own driver calls (a missense change counts at a position recurrently hit in its study, an in-frame or truncating change wherever it is found); a page also compares the tumours carrying the change with the gene's other mutant tumours, gene by gene, in tumour types where each side has 20 or more tumours, and gives, where the calls can be annotated, how the change is written in DNA (its coding change, exon, GRCh37 and GRCh38 positions and dbSNP id); the scan tests genes, not single changes, so co-mutation results on the gene pages count every driver mutation of the gene. Mutation classes (such as EGFR exon 19 deletions, EGFR and HER2 exon 20 insertions, MET exon 14 skipping mutations, KRAS G12 mutations and the PIK3CA helical and kinase domain mutations) have pages too where 30 or more tumours of one cohort carry one, at /mutations/, listed on /mutations: the tumours carrying any change of the class (each tumour counted once), by tumour type, as a share of all tumours tested and of the gene's mutant tumours, with the changes that make the class up; a class is defined by the kind of change (missense, in-frame deletion or insertion, splice site) and the residue it starts at on the canonical (MANE Select) protein, and each gene page in a tumour type gives the gene's classes there. - [Gene pages](https://cybergenic.im/genes): One page per gene the scan keeps with three or more partner genes, or that carries a driver mutation in 100 or more tumours of one cohort, at /genes/ in lower case, for example /genes/egfr. First, per tumour type, the genes most often mutated in the same tumours as it: the share of the gene's mutant tumours carrying each partner (tumours with both of the gene's mutant tumours the pair's test counted), the share expected under the burden-conditional null, the partner's share among all the tumours tested, and whether a direction is established (none established means the thresholds were not reached, not that the genes are unrelated), in the tumour type's reference cohort (the one its tumour type page ranks genes in, where the gene has enough tumours there) and the tumour type's other cohorts beside it (every tested pair counts, kept or not; driver mutations only, no copy number or fusions). Then every partner kept as co-occurring or as mutually exclusive, in every cohort that keeps the pair (tumours, tumours with both mutations, the expected count, odds ratio, q). The index also lists the genes kept with one or two partners, linked to their pair pages. Each gene page also gives how often each tumour type carries the gene, and the gene has a page in each tumour type where a cohort has 100 or more tumours with it or the scan keeps a result for it there, at /genes//, for example /genes/kras/pancreatic-cancer: the share of tumours tested carrying a driver mutation in the gene in each of that tumour type's cohorts (with a 95% Wilson interval and its rank among the cancer genes the scan tests there), headlined and ranked in a reference cohort filed under the tumour type's own name where one has enough tumours (the rule the tumour type page ranks its genes by), which protein changes make up the gene's driver mutations there (for example G12D, G12V and G12R for KRAS in pancreatic cancer: the tumours carrying each as a share of the gene's mutant tumours and of all the tumours tested, in each cohort, from the scan's own driver calls, so a missense change counts only at a position recurrently hit in that study), up to 25 of the genes most often mutated with it there, and the partners kept there. The gene page's tumour type table names each tumour type's most common changes. The numbers are the gene pair pages' own. - [Tumour type pages](https://cybergenic.im/cancers): One page per tumour type in which the scan keeps three or more gene pairs, or in which four or more genes are each mutated in 30 or more tumours of one cohort, for example /cancers/pancreatic-cancer. First, up to 25 of the most frequently mutated genes: each gene's share of the tumours tested for it (driver mutations only, no copy number or fusions; hypermutated tumours and the MSI-high and POLE classes left out), ranked in a cohort filed under the tumour type's own name where one exists (else the most tested), with the tumour type's other cohorts beside it, each labelled with the tumours its source files there. Then every co-occurring and every mutually exclusive pair kept there, by cohort, with the cohorts' tested and kept pair counts. Where a cohort records whether each sample is the primary tumour or a metastasis, the genes whose share differs between metastases and primary tumours (two-sided Fisher exact test, Benjamini-Hochberg q below 0.1 over the cohort's genes; the metastases are other patients, one sample per patient), and every gene page in the tumour type gives the gene's share in each. Breast cancer has a page per receptor subtype, /cancers/breast-cancer/hr-positive-her2-negative, /cancers/breast-cancer/her2-positive and /cancers/breast-cancer/triple-negative: the most frequently mutated genes in that subtype in MSK-CHORD and METABRIC (receptor status as each source records it: MSK-CHORD's from clinician notes by a language model, METABRIC's partly from gene expression), each with its share in the other subtypes and whether it is mutated more or less often there; the breast cancer page and its gene pages also give the shares by PAM50 subtype (TCGA, METABRIC). The tumour type pages and their gene pages also give each gene's share by patient group where a cohort records it: smoking history (MSK-CHORD, inferred from clinical notes), sex, age at diagnosis (under 50 against 50 or over) and genetic ancestry inferred from sequencing (MSK-IMPACT 50K), with the genes that differ between groups (the same Fisher test and q), and by histology and site: the type MSK-IMPACT 50K records for each tumour by its OncoTree code (lung adenocarcinoma against lung squamous cell carcinoma, invasive ductal against invasive lobular breast carcinoma, colon against rectal adenocarcinoma, the glioma, ovarian, endometrial, thyroid and melanoma types, and so on), each listed type compared with the tumour type's other listed types. Some groups have pages of their own at /cancers//, listed in the sitemap once their counts are published: lung cancer in never smokers, early-onset colorectal cancer (diagnosed before 50), the histological types searched by name (such as /cancers/breast-cancer/invasive-lobular-carcinoma, /cancers/non-small-cell-lung-cancer/lung-adenocarcinoma, /cancers/glioma/astrocytoma-idh-mutant and /cancers/melanoma/uveal-melanoma; a type of a category the cohort files under a broader name has a page at /cancers/, such as /cancers/intrahepatic-cholangiocarcinoma, /cancers/esophageal-adenocarcinoma and /cancers/oral-cavity-squamous-cell-carcinoma), and each tumour type's metastases at /cancers//metastatic (where a cohort has 100 or more tumours sequenced from a metastasis; named by the category the cohort files them under where that is broader, such as esophagogastric cancer). A tumour type is the engine's grouping of cohorts across sources. The pair numbers are the gene pair pages' own. - [SelectSim comparison files](https://cybergenic.im/data/external-agreement/selectsim-2026-10-03/README.md): Every gene pair either method calls in a tumour type both tested, with both methods' statistics (comparison.csv, 2,449 rows), every tally of the comparison (summary.json), the code that made them, the check that CDKN2A is almost absent from SelectSim's MSK inputs and results (cdkn2a_check.py, counts in cdkn2a_summary.json), and SHA-256 checksums. - [CTNNB1 and TP53 liver cancer re-test files](https://cybergenic.im/data/external-agreement/wu-2026-hcc-ctnnb1-tp53/README.md): The liver cancer re-test of Wu, Xing and Townsend: a standard-library Python script that downloads the cohorts from cBioPortal, another that rebuilds the study's tables from its repository and runs every exact test, the engine's burden-conditional results as counts, every number the research note cites (summary.json), and SHA-256 checksums. - [BRAF, KRAS and EGFR re-test files](https://cybergenic.im/data/external-agreement/vaeyens-2026-braf-kras-egfr/README.md): The re-test of Vaeyens et al.: a script that rebuilds every one of the study's tables from its authors' repository and re-tests each within tumour types, in the study's design and against chance, every test with its statistics, the cohort class tests with both expectations, every number the research note cites (summary.json), and SHA-256 checksums. - [Prediction registry](https://cybergenic.im/registry): Every prediction the engine locked before testing it, as a public md5 hash chain with downloadable preimages, an in-browser verifier and external anchors: OpenTimestamps proof (Bitcoin), Hugging Face (DOI 10.57967/hf/10762), figshare (DOI 10.6084/m9.figshare.34062204.v1). ## Tools - [Statistics tools](https://cybergenic.im/tools): Free calculators for the tests behind the engine's findings. - [Fisher's exact test calculator](https://cybergenic.im/tools/fisher-exact-test): Two-sided or one-sided exact p-values and the sample odds ratio for a 2x2 table, such as two genes' co-alteration counts. - [Chi-square test calculator](https://cybergenic.im/tools/chi-square-test): A chi-square test of independence for a table of categorical counts. - [Odds ratio and relative risk calculator](https://cybergenic.im/tools/odds-ratio-relative-risk): The odds ratio and the relative risk from a 2x2 table. - [FDR correction calculator](https://cybergenic.im/tools/fdr-correction): Benjamini-Hochberg false-discovery-rate adjustment for a list of p-values. ## For AI agents - [Cybergenic MCP server](https://cybergenic.im/mcp): Model Context Protocol endpoint (Streamable HTTP, no authentication, read-only). Tools: lookup_gene_pair, gene_partners, search_findings, get_finding, get_track_record, fisher_exact_test, chi_square_test, odds_ratio_relative_risk, fdr_correction. - [Connect an AI assistant](https://cybergenic.im/connect): set-up for Claude, ChatGPT, Cursor and VS Code, what each tool returns, and how to read the numbers. ## Optional - [About](https://cybergenic.im/about): Who is behind the site (it belongs to its founder's company; its AI Operator, a language model, runs the analyses and writes the notes; no human specialist referees them), how a note is checked before publication, who pays for it, its dated corrections, and how to cite or correct it. - [Writing](https://cybergenic.im/blog): All essays and research notes. - [RSS feed](https://cybergenic.im/feed.xml): Every essay and research note, newest first. - [Privacy Policy](https://cybergenic.im/legal/privacy): How the site handles personal data. - [Terms of Service](https://cybergenic.im/legal/terms): The terms of use.