Lesson 6 of 7
Module 2 · Biological data and databases
How to find the papers behind every database: PubMed, MeSH and field tags, preprints and reviews, how computers read papers in bulk, and how to use AI search tools without being fooled by invented references.
Where the literature lives
Every database in this module rests on published papers, and sooner or later you need the papers themselves. A good search finds the handful that answer your question; a poor one returns thousands, or misses the key study. Four resources cover most needs.
| Resource | What it holds | Good for |
|---|---|---|
| PubMed | Tens of millions of citations and abstracts in biomedicine and life sciences | The standard first search |
| PubMed Central (PMC) | Free full-text articles | Reading the whole paper, methods and figures |
| Europe PMC | The PubMed abstracts, PMC full text, preprints and more, with text-mined links | Searching inside full text; preprints alongside papers |
| bioRxiv and medRxiv | Preprints in biology and in health sciences | The newest work, before peer review |
How to read this table
Read across each row: the resource, what is inside, and when to use it. The terms mean the following:
- A citation is the record of a paper: title, authors, journal, date and usually the abstract, the short summary at the top of the paper. PubMed, run by the US National Library of Medicine (NLM), holds citations, not the papers. Each record has a PMID, a PubMed identifier, and links out to the full text where it exists.
- Full text is the complete article. PubMed Central is a free archive of it, and its records carry a PMCID, which starts with "PMC". A paper can be in PubMed without being in PMC, if its full text is behind a journal's paywall.
- Europe PMC is a partner archive based at EMBL-EBI. Because it can search the full text, it finds a gene that is mentioned only in a methods section or a figure legend, which an abstract search would miss.
- A preprint is a complete manuscript that the authors post publicly before, or while, it goes through peer review. There is more on how to treat preprints below.
Each of these is a different window onto the same literature, so pick by the question. For a careful, repeatable search, PubMed is the tool, because its records and tags are controlled and its behaviour is documented. To find where a method or a dataset was used, which often appears only in the methods, use full-text search in Europe PMC. Broad web tools such as Google Scholar cover more, including theses and conference papers, but you cannot see or control exactly how they rank results, so they are better for finding one paper you already half know than for a search you must report.
MeSH: the library's own vocabulary
Authors describe the same thing in different words: "heart attack", "myocardial infarction", "MI". To get round this, the NLM tags many PubMed records with MeSH terms (Medical Subject Headings), a controlled vocabulary: a fixed list of agreed terms, each with a definition. Every paper about heart attacks gets the same heading, Myocardial Infarction, whichever words its authors used.
MeSH is arranged as a tree, from broad terms down to narrow ones. When you search a heading, PubMed also searches everything beneath it, which is called exploding the term: searching Respiratory Tract Infections also finds papers tagged with narrower headings such as Pneumonia. Tagging takes time, so the newest records often have no MeSH terms yet. That is why a good search pairs each MeSH term with the plain words authors use.
To find the right heading, use the MeSH database on the NCBI site: type a word and it shows the matching heading, its definition, the terms above and below it in the tree, and the entry terms that map to it. For omics topics, check rather than assume. The heading for cancer is Neoplasms, and many gene-level ideas have no heading of their own, so a gene symbol is usually searched as free text with [tiab]. Search its aliases too: as Identifiers and cross-references showed, older papers may name TP53 as p53.
Boolean operators and field tags
Boolean operators are the three words that combine search terms, written in capitals in PubMed. AND needs both terms, so it narrows; use it to join different concepts. OR accepts either, so it widens; use it to join synonyms. NOT removes papers containing a term; use it sparingly, because it also drops useful papers that mention the term in passing. Whenever you mix operators, group terms in brackets, as in (A OR B) AND C, so the order is clear.
The brackets matter because PubMed reads operators from left to right. Typed without brackets, children AND sepsis OR septicaemia is read as "sepsis in children, or any septicaemia paper", which brings in every septicaemia paper, adult or child. With brackets, children AND (sepsis OR septicaemia) says what was meant. A quick test of any search is to read it aloud and check it means one thing only.
A field tag, written in square brackets after a term, tells PubMed which part of the record to search. Without one, PubMed guesses what you meant (automatic term mapping); a tag makes the search exact and repeatable.
| Tag | Searches | Example |
|---|---|---|
| [tiab] | Title and abstract words | sepsis[tiab] |
| [mh] | MeSH headings, exploded | "Sepsis"[mh] |
| [majr] | MeSH headings marked as a main topic | "Sepsis"[majr] |
| [pt] | Publication type | review[pt] |
How to read this table
Each row is one tag, the part of the record it searches, and an example you could type. Also:
- Quotation marks keep a phrase together, so "septic shock"[tiab] finds the two words side by side. An asterisk truncates a word: infect*[tiab] finds infection, infections and infected.
- [mh] searches the heading and everything beneath it in the tree; [majr] finds only papers where it is a main subject, not a side mention.
- [pt] picks out kinds of article, such as reviews, randomised controlled trials or case reports.
Go deeper: seeing what PubMed actually searched
Automatic term mapping is helpful but invisible. If you type heart attack with no tags, PubMed checks its translation tables and, finding a match, also searches the MeSH heading Myocardial Infarction and related words. You can see exactly what it did: on the Advanced page, open the History and Search Details panel and expand the line for your search. Checking it is the quickest way to explain a surprising result count.
Some habits switch the mapping off. A phrase in quotation marks is searched as written, with no mapping. Truncation with an asterisk does the same: infect* finds words starting with those letters, but will not pull in the MeSH heading for infections. That is why the worked example below always pairs a truncated word with an explicit [mh] term, rather than relying on PubMed to add it.
Filters and saved searches
The filters beside the results do the job of some tags without typing: article type, date, free full text, species, language and age. Note which ones you used, because they quietly change what you see. The Advanced page keeps a numbered search history (#1, #2 and so on) that you can combine, as in #1 AND #2. With a free My NCBI account you can save a search and get email alerts for new matches. A saved search is also a record that anyone can rerun exactly.
One pitfall with filters: several of them, such as species and age, work through MeSH tags. A recent paper that has not been tagged yet will not pass a Humans filter, even if it is about patients. If finding the newest work matters, run the search once with the filter and once without, and look at the recent records the filter dropped. When you report a search, for example in a review or a thesis, give the full query, the filters, the database and the date you ran it, so others can repeat it and see what has been added since.
How to read this figure
Each card is one idea in the question; each grey box is one way of finding it, as a MeSH heading or as title and abstract words. OR collects the boxes in a card, and the vermilion AND markers keep only papers found in all three cards. Too few results: add synonyms to a card. Too many: add a card or a filter.
Worked example · building a PubMed query
The question: does giving vitamin D to children reduce respiratory infections? You want trials, not opinion pieces.
#2 "Respiratory Tract Infections"[mh] OR (respiratory[tiab] AND infection*[tiab])
#3 "Child"[mh] OR child*[tiab] OR pediatric[tiab] OR paediatric[tiab]
#4 #1 AND #2 AND #3
#5 #4 AND randomized controlled trial[pt]
Step by step. First split the question into its concepts: the treatment (vitamin D), the outcome (respiratory infections) and the people (children). For each, look up the MeSH heading in the MeSH database, then add the words authors use; cholecalciferol is the chemical name of vitamin D3. Run each block alone and note its count. Notice the spellings in block 3: PubMed is largely in American English, so include pediatric as well as paediatric. Combine the three blocks with AND in line #4. Finally, limit to trials with the publication type tag, or the matching filter. PubMed spells this type the American way, randomized; if you type it differently, it may not match.
What the answer means. Line #5 is not the answer to the question; it is a list of the studies that could answer it. Read the abstracts, open the most relevant papers in full, and look for a recent systematic review that pools the trials. If a key trial you know of is missing, a block is too narrow: open that trial in PubMed and borrow words from its MeSH terms.
Preprints, reviews and primary research
Peer review is the check a journal arranges before publishing: other experts read the paper and ask for corrections. Preprint servers screen what they post for obvious problems, but do not peer review it. Preprints share results months earlier, which mattered during the COVID-19 pandemic. Treat them as drafts: check for a later or journal version (the server usually links to it), read the methods with extra care, and say "preprint" when you cite one. medRxiv shows a warning on every preprint that it should not guide clinical practice.
The seven questions from Public data and reading a paper work for preprints and journal papers alike. Peer review does not settle them all, and a preprint that answers them well may be sounder than a published paper that does not. Preprints can also change between versions, so note which version you read.
- Reports new data from one study
- Methods and results in full
- Where a claim is first shown
- Summarises many studies
- A quick map of a field
- Reflects the authors' choice of studies
- Searches for every study by a stated method
- Often pools results (a meta-analysis)
- Strongest summary of the evidence
A good habit: start with a recent review to learn the field and its key papers, then go to the primary papers for any fact you plan to rely on. A review can simplify or misquote; the primary paper shows what was actually measured.
Text mining: computers reading papers
No one can read every paper, so computers are used to read them in bulk. Text mining means extracting structured facts from ordinary written text. The first step is usually named-entity recognition (NER): finding the words in a sentence that name a gene, a disease, a chemical, a variant or a species. The second is normalisation: linking each name to an identifier, so that "p53", "TP53" and "tumour protein p53" in human papers all map to NCBI Gene ID 7157. A third step, relation extraction, looks for statements that link two entities, such as "drug X inhibits protein Y".
This is harder than it sounds. Some gene symbols are ordinary words: CAT, SET and WAS are all human genes. One symbol can belong to several species, and one abbreviation can mean different things in different fields. Good tools use the surrounding words to decide. NCBI's PubTator, for example, marks genes, diseases, chemicals, variants and species in PubMed abstracts and many full-text articles, and Europe PMC can highlight such terms on the page as you read.
Text mining also feeds the databases in this module. Expert curators turn findings in papers into entries, such as a protein's function in UniProt. Text mining helps them find the right papers and spot candidate facts, but a person checks each fact against the paper. That is why curated entries cite PMIDs, so you can always go back to the source.
Mined facts are useful at scale, but they carry the tool's errors. A sentence such as "we found no evidence that X binds Y" can be read by a simple tool as a link between X and Y, because negation and uncertainty are hard to detect. Results from text mining are best used to rank papers for reading, or to count how often two entities appear together, and should be checked by eye before they are reported as facts.
Literature-based discovery
Sometimes two groups of papers hold the pieces of an idea that no single paper states. Literature-based discovery looks for these hidden links. If one body of work says A affects B, and a separate body says B is involved in C, but no paper connects A with C, then A and C are a candidate pair worth testing.
How to read this figure
The solid lines are known. One set of papers reported that fish oil makes blood less thick and platelets, the cells that start clots, less sticky. A separate set reported thicker blood and stickier platelets in Raynaud's syndrome, where fingers turn white and numb in the cold. The dashed arc is the new idea that the two sets imply. It is a hypothesis, not a result, and still needs testing. Modern tools run the same search across millions of abstracts, using the entities that text mining has found.
The fish oil example comes from Don Swanson, an information scientist, in "Fish oil, Raynaud's syndrome, and undiscovered public knowledge" (1986, Perspectives in Biology and Medicine). A small controlled trial of fish oil in patients was later reported by DiGiacomo, Kremer and Shah (1989, The American Journal of Medicine).
When you cite this work, keep its limits in view. Swanson's method of linking A to C through B produces many candidate pairs, and most turn out to be dull or wrong. The fish oil link is famous partly because it was one of the few that was tested. A literature-based link is a reason to run an experiment, not evidence in itself.
AI search assistants
Chat tools built on large language models (LLMs), AI systems trained on vast amounts of text to write language, can answer questions about the literature in plain sentences. They help with a starting point, synonyms for a search block, or a summary of a paper you have open. But an LLM writes what sounds likely, and it can produce a reference that does not exist: real-sounding authors, a plausible title, a journal and a year, all invented. Tools that search a real database first and cite what they found are safer, but can still misstate what a paper says.
So treat every AI citation as a lead, not a source. Open each one: find it in PubMed by its PMID, DOI or title, check that it exists, and then check that it actually says what the summary claims. Look for a retraction notice too, as PubMed flags retracted papers. Never put a reference you have not opened into your own work.
Used with this care, assistants can save time. A good pattern is to ask for search terms or synonyms, then run the search yourself in PubMed, so the list of papers comes from a database you can report. When you ask for a summary of a paper, give the tool the paper itself, and check any number it reports against the paper's tables.
Key point
Build searches from concepts: synonyms joined by OR, concepts joined by AND, MeSH plus free text. And never trust a citation, from a person or a machine, until you have opened it.
An example: an AI assistant cites a 2019 trial with a large benefit. PubMed has no paper with that title, and the first author's 2019 paper is on another topic. The reference was invented, and two minutes of checking caught it.
Further reading · Module 2PubMed User Guide: field tags, filters, search history and saved searches. · MeSH at the National Library of Medicine: how the vocabulary is organised. · Europe PMC help: full-text search and preprints. · All reading for this course
Check your understanding
1. A search using "Cystic Fibrosis"[mh] alone misses several papers published last month. What is the most likely reason?
2. You have three names for the same drug and want papers about its use in children. Which search has the right shape?
3. An AI assistant gives you five references supporting a claim. What should you do before using them?
Up next
Case study
One gene, every database