LSMA

LONDON SCHOOL OF MULTIOMICS & AI

Free course

Free course

0 of 73 lessons

Lesson 5 of 7

Module 2 · Biological data and databases

Public data and reading a paper

Public data and reading a paper

Public data and reading a paper

You do not need your own samples to start. Large public datasets are free to explore, and knowing how they were made is also how you judge any multi-omics paper.

⏱ 10 min readModule 2 · Biological data and databasesFree
Lesson at a glance4 steps · about 10 min
  1. The main resources
  2. How to read this table
  3. Seven questions for any omics paper
  4. What to look for under each question

In the first four lessons of this module, from File formats you will meet: FASTA, GenBank, GFF, VCF, PDB to Reading a gene record and a protein record, you met the formats, databases and identifiers that hold single genes and proteins. This lesson zooms out to whole studies: the large public datasets behind most omics papers, and how to judge a paper built on them.

The main resources

A handful of public resources hold most of the cancer omics data you are likely to use. Each was built for a different purpose, so it helps to know what each one is for before you start looking.

The difference between them is mostly one of design. TCGA and CPTAC are consortium projects: one large team collected samples to a shared plan and processed them with the same pipelines, so samples can be compared with each other. GEO and ArrayExpress collect what individual labs deposit, so each dataset follows its own design, and two datasets on the same disease may not be comparable at all. cBioPortal sits on top, making existing studies easy to browse rather than adding new data of its own.

ResourceWhat it holdsGood for
TCGA, via the GDC portalMore than 11,000 patients across 33 cancer types: DNA, RNA, methylation and clinical dataPan-cancer and tumour-type analyses
CPTACProteomics for a range of tumour types, with matched genomic dataAdding the protein layer
cBioPortalA browser view of many cancer studiesExploring genes without coding
GEO and ArrayExpressDatasets deposited with published papersFinding data for a specific question

How to read this table

Read across each row: the name, what is inside, and the kind of question it suits. The abbreviations stand for the following:

  • TCGA is The Cancer Genome Atlas, a large US project that measured several layers in the same tumours, usually with matched normal samples. Its data is distributed through the GDC (Genomic Data Commons), a portal run by the US National Cancer Institute. Pan-cancer means across many cancer types at once, for example asking whether a gene is mutated more often in some tumour types than others.
  • CPTAC is the Clinical Proteomic Tumor Analysis Consortium. It measured proteins, and often phosphorylated proteins, by mass spectrometry (Proteomics and metabolomics) in tumours that were also sequenced, so you can compare the RNA and protein layers in the same samples.
  • cBioPortal is a website that loads many cancer studies, TCGA among them, and shows mutations, copy number and expression for any gene you type in. It is the quickest way to ask "how often is this gene altered, and in which cancers?"
  • GEO (Gene Expression Omnibus) and ArrayExpress are archives where researchers deposit the data behind a published paper. They hold many thousands of smaller studies, often with a single, focused design.

Each resource has a known blind spot. TCGA collected mostly primary tumours taken at surgery, largely before treatment, so it says less about metastases or tumours that have stopped responding to a drug. In cBioPortal, the frequency of an alteration depends on which studies you select; when two selected studies include some of the same patients, those patients can be counted twice, so check which studies are in a query before quoting a percentage. In GEO, each dataset is a series with an accession starting GSE, each sample has one starting GSM, and the measuring platform one starting GPL. The processed values were made by the authors with their own choices, so read the sample descriptions and the processing notes before combining a GEO series with anything else. For sequencing studies, the raw reads behind a GEO series usually sit in the SRA, the read archive from The main databases.

Most processed data is open access. Raw reads and inherited (germline) variants can identify a person, so they are controlled access: researchers apply for permission and agree how the data will be used.

Processed data is data after the pipeline of How sequencing works has turned reads into summaries, such as a count matrix or a table of methylation levels. Why can raw data identify someone? A person's germline variants are close to unique, and relatives share many of them, so a full set of variants works like a fingerprint. In practice, an application goes to a data access committee, which checks who you are, what you plan to do and how you will store the data securely. In the United States this goes through dbGaP; in Europe many datasets are held in the European Genome-phenome Archive (EGA). For learning, the open-access processed data is usually all you need.

Open access does not mean free of conditions. Most resources ask you to cite them, some ask you to wait for the original consortium's main papers before publishing on the same question, and none allow attempts to identify the people who gave samples. Read the data use terms on the portal before downloading, and record the release or date of what you used, because portals reprocess their data from time to time and the numbers can change.

Public data and reading a paper · LSMA free course Two panels. Open access holds processed summaries such as gene counts, methylation levels and clinical summaries, which anyone can download. Controlled access holds raw reads and germline variants, which can identify a person, behind a lock that requires an application. 1 · OPEN ACCESS 2 · CONTROLLED ACCESS Gene counts per sample Methylation levels per site Clinical summaries Raw reads (FASTQ, BAM) Germline variant calls Processed: download freely Can identify a person Apply first, agree terms of use
The usual split in large cancer resources such as TCGA: summaries that cannot identify anyone are open, while raw reads and inherited variants sit behind an application. The exact lists differ from resource to resource. Schematic.
Go deeper: reading a TCGA barcode

Every TCGA sample has a barcode, and its parts tell you what you are holding. A barcode starts like TCGA-XX-YYYY-01A. TCGA is the project. XX is the tissue source site, the centre that collected the sample. YYYY identifies the patient, so the first three parts together name one person.

The next two digits are the sample type, and they matter most. 01 is a primary solid tumour, 06 a metastasis, 10 normal blood and 11 normal tissue taken near the tumour. The letter after them marks the vial. Longer barcodes go on to name the portion, the analyte (DNA or RNA) and the sequencing plate.

So two files whose barcodes agree up to the patient but end in -01 and -11 are a tumour and its matched normal from the same person. Filter on the sample type before an analysis, or you may mix tumours with normal tissue without noticing.

Seven questions for any omics paper

The same questions you would ask of your own data apply to anyone else's. Work through them in order; each one maps onto an earlier lesson.

A practical order for reading: the abstract to learn the claim, the figures to see the evidence, then the methods to see how the evidence was made, and only then the discussion. Also find the data availability statement, usually near the end. It should give accessions in archives such as GEO, SRA or EGA, and with them you can open the samples yourself and check, for example, how many there really are and how they were labelled. A paper whose data cannot be found is harder to trust, however clear its figures.

  1. What is the question? Discovery, classification or prediction?
  2. Which samples? How many patients, which tumour types, how were they chosen?
  3. Which layer and assay? Does it match the question?
  4. How deep and how good? Coverage, quality control, purity.
  5. How normalised and tested? Were p-values adjusted for multiple testing?
  6. Batches? Were groups processed together or separately?
  7. Validated? In an independent cohort, or by another method?

What to look for under each question

  • The question. Discovery looks for something new, such as genes that differ between groups. Classification sorts samples into known or new groups, such as tumour subtypes. Prediction estimates something for a future patient, such as response to a drug. Prediction needs the strongest evidence, because it will be applied to people who were not in the study.
  • The samples. Small numbers make chance findings likely. How patients were chosen matters as much: patients from one specialist hospital may be sicker, or treated differently, from patients elsewhere.
  • The layer. A claim about which genes a tumour is using needs RNA; a claim about activity of a signalling pathway is stronger with protein or phosphorylation data (The central dogma, as data and Proteomics and metabolomics).
  • Depth and quality. Look for coverage (The central dogma, as data), quality control (How sequencing works) and tumour purity (Genomics). A low-purity sample can hide real mutations.
  • Normalisation and testing. Check which normalisation was used and that the reported p-values were adjusted (Bulk transcriptomics).
  • Batches. If all of one group was processed separately from the other, the difference may be the batch (Bulk transcriptomics).
  • Validation. An independent cohort is a separate group of patients, ideally from another centre, that played no part in finding the result. Confirming a finding there, or with a different method such as IHC instead of RNA-seq, is the strongest test of all.

One more trap runs through several of these questions: using the same samples both to choose and to test. If the genes in a signature were selected because they separated two groups in a set of patients, then showing that the signature separates those same patients proves very little; it was built to do that. The same holds for machine-learning models tuned on the data they are then scored on. Look for a clear statement that the samples used to build the result were kept apart from those used to test it. When that split is missing or unclear, read any reported accuracy as an upper bound.

Worked example · a claim, checked

A paper reports a 12-gene signature that separates two subtypes, found in 40 patients from one hospital, with no adjustment for multiple testing and no external validation.

question ✓ assay ✓ samples: 40, single site
testing: unadjusted validation: none
→ a promising hypothesis, not yet a finding

Step by step. A gene signature is a small set of genes whose combined expression is used to sort samples. Go through the questions one at a time. The question (classification) and the assay are reasonable, so they get a tick. 40 patients from one site is small, and any quirk of that hospital's patients or lab is built into the result. If the 12 genes were picked from about 20,000 without adjustment, then, as in Bulk transcriptomics, about 20,000 × 0.05 = 1,000 genes could pass p < 0.05 by chance, so some or all of the 12 may be noise. With no validation, nothing has tested whether the signature works on anyone else.

What the answer means. The paper is not wrong; it is early. The fair summary is "worth testing", and the next step would be to apply the same 12 genes, unchanged, to an independent cohort. A signature that keeps separating the subtypes there has earned the word "finding".

Checklists like these exist in published form. For studies of tumour markers that predict outcome, the REMARK guidelines (McShane and colleagues, 2005, published in several journals including the Journal of the National Cancer Institute) list what a paper should report, including how samples were chosen and whether results were validated. The danger in the second example, picking the cut-off that gives the biggest difference, was shown by Altman and colleagues (1994, Journal of the National Cancer Institute).

These guidelines describe good reporting, not proof that a result is true. A paper can tick every box and still fail to replicate, so independent validation remains the final test.

Key point

A paper's strength lives in its methods section. Read it before you trust the abstract's conclusions.

A second example: an abstract states that a methylation marker "predicts survival". In the methods you find that the patients were split into high and low groups at whichever cut-off gave the most striking difference, and that the result was not tested in a second cohort. Choosing the cut-off after seeing the outcome makes a striking difference much more likely by chance. The abstract is accurate, but only the methods tell you how much weight it can bear.

Further reading · Module 2Sources for this module and the rest of the course are gathered on one page. · All reading for this course

Check your understanding

1. Why are raw sequencing reads usually controlled access?

2. Which resource lets you explore a gene across many cancer studies without coding?

3. A signature was found in 40 patients and never tested elsewhere. How should you treat it?

Up next

Lesson 6

Searching and mining the literature

Ask BioTutor

Create a free website with Framer, the website builder loved by startups, designers and agencies.