Methods overview
We analyzed approximately 346,000 neuroscience papers published in neuroscience-specific journals from 1975 through 2024 to create an interactive map showing how the field has evolved over time. Our goal was to identify prominent research trends in neuroscience and track their popularity across decades.
1. Building our dataset
We first downloaded PubMed’s baseline, which contains data about tens of millions of biomedical research papers. To focus exclusively on neuroscience, we used SCImagoJR, a journal ranking database, to identify the subset of journals dedicated solely to neuroscience, excluding those that also cover other fields. We analyzed papers only from these neuroscience-specific journals. (See our answers to frequently asked questions for further explanation.)
2. Extracting key terms
For each paper, we used GPT-4o mini to analyze the title and abstract and extract up to 10 key terms, such as “synaptic plasticity,” “fMRI” and “Alzheimer’s disease.” The model was specifically instructed to preserve the exact terminology from the source text to ensure fidelity. This process generated a comprehensive dataset of keywords representing the primary topics of neuroscience research over approximately the past 50 years.
3. Organizing key terms into categories
To make sense of thousands of different terms, we organized them into the following categories:
- Methods, metrics and technology
- Biological and cognitive processes
- Brain regions
- Molecules
- Diseases and conditions
To automate the process, we used a deep-learning model specialized in capturing the meaning of natural language to generate numerical fingerprints for each key term and each category. By comparing each term’s fingerprint against all category fingerprints, we assigned terms to whichever category showed the highest similarity. We manually reviewed these categorizations to ensure accuracy, removing or re-categorizing terms that did not fit their assigned category.
4. Cleaning and standardizing terms
The raw data contained numerous terminological variants, such as synonyms, acronyms and plural forms. For example, “fMRI,” “functional MRI” and “functional magnetic resonance imaging” all refer to the same brain-scanning technique. To consolidate these terms into a consistent and usable structure, our automated system recognizes equivalent terms, including spelling and plurals; expands acronyms to full names; and links common and scientific species names, such as “zebrafish” and “Danio rerio.”
The map displays umbrella terms, such as calcium imaging, to keep things readable. Specific subterms, such as “Ca(2+) imaging,” “Ca2+ imaging” and “two-photon calcium imaging,” remain searchable and appear in detail panels.
5. Analysis
Terms on the map are arranged so that items that tend to occur together are placed closer to each other. We used a dimensionality-reduction method (UMAP) to position related terms closer together. The height of the “hills” on the map, as well as the type size, indicate the number of publications in a topic area. The color overlay reveals regions in which activity has increased recently, calculated by comparing how much more or less often a term appeared in the past five years versus the five years before that.
On the Trends page, terms are ranked according to total publication count, change in publication share (past five years versus previous five) and year of first occurrence. The Term Shifts visualization reveals the relative change in publication frequency for the 50 most-used terms. The thickness of each stream corresponds to the share of the term in a given year within this term subset. The color and stacking indicates the recency of the term: Those that are more frequent in recent years are stacked on top and shown in shades of red, whereas terms decreasing in frequency are shown in shades of blue on the bottom of the stack.