cansim-terms

An ablation removes one component or modality at a time and re-measures performance, which is the only way to know what each part contributes; multi-task heads let one shared backbone serve several outputs.

Accuracy is the share of predictions that were right; macro-F1 averages the F1 score of each class equally, so a model cannot look good by getting only the common cancer types right.

An actionable alteration is a change in a tumour's DNA that maps directly to an approved or investigational drug, so finding it changes what the patient is offered.

AJCC stage is the stage group (I to IV) assigned from the TNM manual published by the American Joint Committee on Cancer; it is the staging most clinical tables record.

DuckDB is an in-process analytical database that queries Parquet files with SQL; PostgreSQL is the standard client-server database; either can hold the metadata spine that joins patients, specimens and assays.

Analytical validation shows a test measures what it claims, reliably; clinical validation shows the measurement actually predicts the patient outcome it is meant to.

AnnData is the Python object, saved as an .h5ad file, that holds an expression matrix together with its per-cell and per-gene annotations; it is the lingua franca of single-cell analysis.

HDF5 and Zarr store large numerical arrays in chunks (Zarr is the cloud-friendly one, and OME-Zarr its bio-imaging profile for slide tiles); Parquet stores tables column by column for fast analytical queries.

Attention-based multiple-instance learning gives each tile of a slide a learned weight and sums the weighted tile vectors into one slide vector, so a slide-level label can train the model and the weights show which regions mattered.

An autoregressive model predicts the next element of a sequence from the ones before it; trained on DNA it learns to continue a genome, which is how Evo 2 was built.

A batch effect is a difference in the data caused by when, where or how samples were processed rather than by biology; if it lines up with the outcome, a model learns the batch instead of the disease.

A benchmark is a fixed dataset and task on which models are compared, and a leaderboard ranks them; both mislead once the test data has been seen in training or the metric no longer tracks the goal.

Bootstrapping refits a model or recomputes a statistic on many resamples of the data drawn with replacement, and the spread of the results is a confidence interval that needs no formula.

Bulk RNA sequencing reads all the messenger RNA in a piece of tumour at once, giving one averaged expression value per gene for the whole sample.

The C-index is the fraction of patient pairs in which the model ranked the patient who had the event sooner as higher risk; 0.5 is a coin toss and 1.0 is perfect ranking.

A calibrated model's predicted 30 percent risk really happens about 30 percent of the time; a reliability diagram plots predicted against observed, and the Brier score measures the squared gap.

A hub for the vocabulary of cancer AI: the assays and cohorts models train on, the machine-learning and statistics terms in their papers, the standards their data must follow and the licences that govern reuse.

Because a tumour sample is a mix of cancer, stromal and immune cells, a difference between two samples may only reflect a different mix of cells.

Cancer cell lines are tumour cells grown indefinitely in the laboratory; they are convenient for drug testing but only imperfectly stand in for a patient's tumour.

A patient is censored when the study ends or they leave before the event (death or progression) happens, so we only know they survived at least that long.

CITATION.cff is a small file in a code repository that says, in a machine-readable way, how to cite the software; GitHub and Zenodo read it to generate citations.

Clinical covariates are the ordinary facts about a patient, such as age, stage, node count and whether they had chemotherapy or radiotherapy, that any prognostic model must beat or build on.

Much of what is known about a patient sits in free text, in clinic notes and pathology reports; language models are now used to read it and to pair it with slides.

When a cancer copies the HER2 gene many times over, its neighbours on chromosome 17 are copied with it; those co-amplified genes are passengers, not the driver.

Conformal prediction wraps any model so that, instead of one answer, it returns a set or interval guaranteed to contain the truth a chosen fraction of the time (say 90 percent), assuming new patients resemble the calibration patients.

Contrastive learning trains a model to pull matching pairs (two views of one slide, or one patient's RNA and protein) close together in embedding space and push non-matching pairs apart.

Controlled-access data can identify a person (raw sequence reads, germline variants), so it is held in archives like dbGaP and EGA and released only to approved researchers under an agreement.

A copy number alteration is a stretch of DNA that a tumour has gained extra copies of or lost, from a single gene to a whole chromosome arm.

Cross-validation splits the data into k folds, trains on k minus one and tests on the last, rotating so every sample is tested once; stratified folds keep the class balance the same in each fold.

Leakage is when information from the test data reaches the model during training, through a scaler fitted on all samples, a patient split across folds or a benchmark that appeared in the pretraining set, making the model look better than it is.

A data use agreement is the contract a researcher signs to receive controlled data, promising to protect it and use it only as the participants consented.

DCLK1 is a kinase gene from neuronal development that also marks a rare stem-like cell population in the gut, studied as a colorectal and pancreatic cancer stem-cell marker.

Mass spectrometry weighs fragments of every protein in a sample to identify and quantify thousands of proteins and their phosphorylation sites, the closest measurement to what a cell is actually doing.

Digital pathology scans glass slides into gigapixel images that can be viewed, shared and analysed by software, which is what makes AI on pathology possible.

Methylation arrays measure, at hundreds of thousands of CpG sites, the fraction of DNA copies carrying a methyl mark; that fraction, between 0 and 1, is the beta value.

DNAJC12 is a co-chaperone gene that is strongly expressed in oestrogen receptor-positive (luminal) breast cancers and is used as a luminal marker.

Domain shift is the mismatch between the data a model was trained on and the data it meets later (cell lines versus tumours, one sequencing platform versus another); domain adaptation is the family of methods that try to bridge it.

Drug sensitivity is how strongly a tumour or cell line is held back by a compound; it is summarised by the concentration that halves growth (IC50) or by the area under the whole dose-response curve.

The floor for any drug-response model is predicting each drug's average effect across cell lines; tuned gradient-boosted trees such as LightGBM often tie or beat deep models under honest splits, which is why evaluation frameworks now exist.

How a drug-response dataset is split decides what a model's accuracy means: hold out cell lines to test personalised prediction, hold out drugs to test drug design, hold out tissues to test repurposing; holding out random pairs only tests imputation.

E2F7 is an unusual member of the E2F family that represses genes rather than activating them, helping cells pause after DNA damage.

An embedding is a list of numbers a model produces to stand for an input (a tile, a gene, a patient), placed so that similar inputs land near each other.

Endocrine therapy blocks the hormones that feed some breast and prostate cancers; endocrine resistance is when the cancer learns to grow without them.

An Ensembl gene ID such as ENSG00000141736 is a stable machine identifier for a gene that survives symbol changes, so pipelines join on it rather than on names.

External validation tests a model on patients from a different hospital, country or platform than it was trained on; it is the only evidence that a result travels.

Fine-tuning updates a pretrained model's weights on the new task; using it frozen keeps the weights fixed and trains only a small head on its features; LoRA is a cheap middle way that trains small low-rank updates.

A foundation model is a large model trained on a vast amount of data without a specific task in mind, then adapted to many downstream uses.

Genes that rise and fall together across samples form co-expression modules; this correlation structure is what expression models mostly learn.

Gene set enrichment analysis asks whether the genes that changed in an experiment cluster in a known pathway or signature, turning a list of genes into a biological story.

A genome build is the version of the human reference sequence coordinates are measured against; mixing GRCh38 and the older hg19 puts variants at the wrong positions.

Sequence models read DNA or protein letters the way language models read text: ESM predicts protein structure and function, Enformer predicts gene expression from DNA sequence, and Evo 2 models whole genomes with a million-base context.

GISTIC scans copy number data across many tumours to find the regions that are amplified or deleted far more often than chance, the likely targets of selection.

Foundation models are trained on graphics processors, and mixed precision stores most numbers in 16-bit floats to halve memory and double speed; a laptop-class chip can fine-tune small models but pretraining at scale waits for a data-centre GPU.

Grade describes how abnormal the cancer cells look under the microscope; stage describes how far the cancer has grown and spread.

GRB7 is an adaptor protein gene that sits next to HER2 on chromosome 17 and is copied along with it in HER2-positive cancers.

H&E is the purple and pink stain on almost every pathology slide: haematoxylin colours cell nuclei blue-purple and eosin colours the cytoplasm and connective tissue pink.

The HGNC symbol is the one official short name for each human gene (ERBB2, not HER2 or NEU), so that datasets can be joined without guessing.

HGVS is the standard way to write a DNA or protein change, such as EGFR c.2573T>G or p.Leu858Arg, so that the same variant is named the same way everywhere.

The Hugging Face Hub is the website where most machine-learning models and many datasets are published, with a model card, version history and, for gated models, an access request form.

The Human Phenotype Ontology is a standard vocabulary of clinical features (signs, symptoms, findings) with codes, used to describe patients in a way computers can compare.