How to write a literature review in biology

A biology literature review has to do something the writing guides rarely mention: it has to establish that a mechanism is understood well enough for your experiment to be the obvious next one. Coverage is not the goal. The goal is a chain of evidence that ends with a question your bench work can answer.

The field-specific difficulties are practical rather than conceptual. Gene and protein names are unstable, results depend heavily on the model system and the assay, preprints circulate ahead of peer review and change between versions, and retractions in the life sciences are common enough that inherited citations are a real hazard.

This guide works through those. The general AI-assisted workflow is in how to write a literature review with AI, so what follows is the layer that keeps a life sciences review from quietly going wrong.

What a biology literature review has to establish

Read your own draft against four questions. What is known about the system, and with what confidence. What is disputed, and why the disputes exist. What has never been tested, as opposed to tested and found null. And why your organism, assay and design are the right instruments for the untested part.

The fourth question is where most reviews are thin. Students describe the biology thoroughly and then introduce their methods as though the choice were obvious. Examiners want the review to make the choice inevitable: if the literature shows a mechanism has only been demonstrated in immortalised cell lines, your use of primary cells stops being a preference and becomes an argument.

It helps to write the last paragraph of the chapter first, in draft. If you cannot yet state what remains unknown in one specific sentence, you are not finished reading.

Confidence is worth stating explicitly rather than implying. There is a real difference between a mechanism demonstrated by loss and gain of function in two independent systems and one inferred from a correlation in a single transcriptomic dataset, and a review that writes both in the same register is not telling the reader what it actually knows.

Search terms: gene, protein and organism synonyms

Nomenclature is the biggest silent source of missed literature in biology. A gene may have been renamed, its protein product usually has a different name, older papers use the historical symbol, and the same symbol can refer to different genes in different organisms. Search only the current official symbol and you will miss the foundational work.

Build a synonym block before you search, and keep it in your notes so the search is reproducible.

Run the blocks separately before you combine them, and look at the counts. If your gene block returns almost nothing while the pathway block returns thousands, the symbol you searched is probably not the one the field uses, and it is far better to discover that in the first hour than in the third month.

• Current approved gene symbol plus the full approved name.
• Historical symbols and aliases, including the name used in the original discovery paper.
• Protein product names and common abbreviations, which often differ from the gene symbol.
• Organism terms in several forms: species name, common name, and strain or cultivar where it matters.
• Process and pathway terms, so you catch papers that study the mechanism without naming your gene in the title.
• Assay and technique terms, when your question is really about what a particular method can show.

Where to search and how wide to cast

Run the synonym-aware string wide first. Literature Finder covers an internal corpus alongside live OpenAlex, PubMed, Semantic Scholar and arXiv, which reaches both the published life sciences literature and the physics-adjacent preprint literature that quantitative biology often ends up in. If your programme expects specific additional databases, run those separately and record them in your methods with their own strings and dates.

Then do the part search cannot do. Follow the reference lists of the two or three best recent reviews in your area, follow citations forwards to see who has challenged them, and check the supplementary material of the papers closest to your question, because negative results and method details are frequently exiled there.

Set a stopping rule before you begin, or the searching never ends. A workable one is saturation by citation: when three consecutive papers you open cite nothing you have not already seen, the core of that sub-literature is in your library and further searching is producing diminishing returns.

Keep everything in one project as you go. Uploading PDFs and pulling paper details from a DOI puts the whole corpus in one searchable place, with reading statuses so you can tell the papers you have actually read from the ones you have merely collected.

Preprints and version drift

Preprint servers have become normal infrastructure in the life sciences, and for a fast-moving question the preprint literature may be where the current state of the field actually is. Using it is fine. Using it carelessly is not.

Three rules keep you safe. Label preprints as preprints in your text, so the reader knows the work has not been peer reviewed. Record the version and the date you read, since figures and conclusions can change between versions. And check again before submission whether a peer-reviewed version has appeared, because if the published numbers differ from the ones you cited, your sentence is now wrong.

The same discipline applies to any claim you carry from a preprint into your own reasoning. If a whole paragraph of your justification rests on unreviewed work, say so plainly rather than letting the citation format imply more authority than the source has.

Check for a published version at least twice: once when you finish the chapter and once in the final fortnight before submission. In fast-moving areas a preprint can be peer reviewed, revised and published inside a single academic term.

Read the methods: model system, assay and sample size

In biology the finding cannot be separated from how it was produced. The same question asked in a knockout mouse, an immortalised cell line, an organoid and a field population can give different answers, and all four can be correct within their systems.

So extract methods deliberately, not incidentally. For each study record the organism or system, strain or line, the assay and its readout, the number of biological versus technical replicates, the controls, and the statistical test used. That set of fields becomes the columns of your evidence matrix and does most of the explanatory work later when studies disagree.

This is also why grounding matters here. Methods detail lives in the full text and the supplement, never in the abstract. CiteDash verifies claims against full text you actually hold, whether that is an open access paper or a PDF you uploaded, which is exactly the material an abstract-level summary would skip. Where a paper is paywalled and you cannot hold the text, treat any claim about its methods as unverified until you have read it.

Record reagent and resource identifiers wherever papers supply them. Antibody validation is a known weak point in the literature, and two studies that appear to disagree about where a protein sits sometimes differ only in the antibody used. Noting the reagent takes one line in your matrix and occasionally explains an entire contradiction.

Synthesise mechanism, not abstracts

A review that walks through papers one at a time reads as a reading list. A review that walks through a mechanism, recruiting papers as evidence for each step, reads as understanding.

Build the chapter as a chain: this is the pathway, this is the evidence for each step, these are the steps where the evidence is indirect or comes from a distant model system, and this is the step nobody has tested directly. The weak links in that chain are where your project should sit, and stating them explicitly is what makes the transition into your aims feel earned rather than announced.

An evidence matrix across your saved papers makes the chain visible. When you can read down a column and see that six studies used the same cell line and the two that used primary tissue disagree, you have located a boundary condition, which is a far better paragraph than a summary of six agreeing abstracts.

Say plainly when a step in the chain rests on a distant model. Evidence from yeast, from an immortalised line, from a mouse and from a human cohort do not carry equal weight for a claim about human physiology, and a review that grades its own evidence in the text reads as confident rather than hedging.

Contradictory results are the chapter, not a problem

New researchers often panic when two papers disagree and try to decide which one to believe. Experienced ones look for the variable that differs. In biology that variable is usually concrete: species or strain, developmental stage, dose and timing, growth conditions, antibody or reagent, detection method, or how the outcome was quantified.

Write the disagreement out. State both findings, list the methodological differences, and say which explanation you find most plausible and why. If the field has not resolved it, say that too. Examiners read this as maturity, and it is frequently the paragraph that produces a viva question you are already prepared for.

Also distinguish an untested question from a tested one that produced nothing. Null results are underpublished, so absence of a finding in the literature is not evidence of absence. If you can, check conference abstracts, theses and supplementary material before claiming nobody has looked.

When you cannot resolve a disagreement, make it productive. A paragraph that closes by naming the experiment which would settle the question is a strong ending, particularly when that experiment is the one you are proposing to run.

Retractions and image integrity in the life sciences

Retraction is a live issue in biology, and image duplication is one of the more common triggers. The risk to you is not that you will cite a paper you know to be retracted; it is that you will inherit a citation from another paper's reference list and never check it.

Make screening routine rather than heroic. CiteDash flags retracted papers with a badge and blocks them from being cited, and the free retraction check will screen a DOI directly if you are working outside a project. Rerun the screen close to submission, because a paper can be retracted after you cite it.

Watch for the softer signals too: expressions of concern, corrections and errata. These do not remove a paper, but they change how much weight a claim can carry, and mentioning one in your text shows you read past the abstract.

Citation style and the last checks

Biology programmes commonly use Nature style, Harvard or ACS, all of which are among the twelve styles Reference Manager supports. Numbered styles in particular are easier to manage when citations behave as objects rather than typed strings, so renumbering after you move a paragraph is automatic instead of manual.

Then run the chapter against this list before you send it on.

• Every mechanistic claim names the system it came from, not just the finding.
• Preprints are labelled as preprints, with version and date recorded.
• Gene and protein names follow your organism's nomenclature convention consistently, with historical synonyms noted at first use.
• Contradictory findings are addressed rather than omitted, with a proposed explanation.
• Every reference has been screened for retraction, including the ones inherited from other papers.
• The final section states one specific untested question, and it is the one your experiments address.

Common mistakes in a biology literature review

The errors that cost marks in this chapter are mostly about precision rather than about how much you read.

• Reporting a result without its system, so a finding from one cell line is written as though it were general.
• Missing an entire literature because the search used only the current approved gene symbol.
• Citing a preprint as though it were peer reviewed, or citing a version that has since changed.
• Describing the pathway thoroughly and then introducing your methods with nothing connecting the two.
• Treating absence of published evidence as evidence of absence, when null results are simply underreported.
• Inheriting references from other papers' reference lists without opening the sources, which is how retracted work keeps circulating.