Q: Why did you focus on neuroscience-specific journals?
A: We narrowed down our data collection to journals dedicated to neuroscience to keep the content highly relevant. Including broad, multi-disciplinary journals, such as Nature or Science, or the entirety of PubMed would have introduced many papers outside neuroscience, which could skew the results and dilute clear trends. By using only specialty neuroscience journals, we ensure that the topics and trends reflect neuroscience research specifically.
Q: Why didn’t you use GPT-4o mini to identify neuroscience papers in non-neuroscience journals?
A: Many papers sit on the boundary between neuroscience and other fields. To validate GPT-4o mini’s classification accuracy, we’d need hundreds of manually labeled papers as a benchmark, which we don’t have. Even if we did, those labels would reflect someone’s subjective judgment about where to draw the line. Keyword extraction, however, works well because it’s open ended: More than one set of tags can describe the same paper. For consistency, we chose to rely on the established SCImago Journal & Country Rank list of neuroscience journals.
Q: Why are the results so clinically focused?
A: Many neuroscience papers, especially in neurology and neuroimaging, mention diseases or patient-related terms in titles and abstracts to signal clinical relevance. As a result, those terms naturally became prominent in our analysis.
Our dataset also includes several clinically oriented neuroscience journals, including Biological Psychiatry, Neurobiology of Disease, Experimental Neurology, Neural Regeneration Research, Neurotrauma Reports, Translational Neuroscience and eNeurologicalSci, which further amplifies clinical terminology. It may also be that a substantial share of neuroscience research is clinically focused, something we may later quantify by comparing our dataset with the broader literature.
Q: How did you define the categories? Why isn’t my subfield represented?
A: We manually designed categories based on discussions with the team and collaborators. Some subfields are more difficult to categorize than others and were beyond the scope of this release. The current list reflects scope, clarity and user interface space constraints for this release—the selected categories do not indicate the importance of the subfield. We might iterate; if something should be promoted or renamed, please share example papers/terms and we’ll consider incorporating that feedback in a future update.
Q: Why didn’t you use an algorithm to define the categories?
A: Algorithmic clustering produces categories that are difficult to interpret and explain to users. Clustering abstracts based on embeddings generates statistically valid but obscure (less interpretable) groupings that don’t always align with how researchers actually think about the field. These clusters also lack stability. If the algorithm is rerun with updated data, categories could reorganize entirely, undermining our need for a consistent, clear taxonomy.
Instead, we prioritized categories that neuroscientists would instantly recognize and find useful. Furthermore, the editorial approach allowed us to move quickly within timeline constraints while delivering intuitive, meaningful organization.
Based on community feedback and evolving needs, we may explore algorithmic approaches in future iterations.
Q: Why can’t I find my favorite term?
A: Our interactive app highlights a curated subset of terms, so it’s possible that a specific term you care about isn’t immediately visible. In the current interface, we focused on the most frequent and relevant terms within each category. If your favorite term does not appear on the map, it still may be available via the search bar.
We did extract a very large number of keywords (approximately 10 per abstract) from neuroscience papers, and your term is most likely in our raw data. However, not every term made it into the app’s display because of limits on what we can show at once. If an important term appears to be missing, it might simply not yet be available in the user interface. We’re looking into ways to enable searching for any term, and we welcome suggestions of terms that you feel should be prioritized.
Q: Why are common systems neuroscience terms missing from the analysis?
A: Some well-known terms from systems neuroscience might not appear in the current categories, and this stems from the way we automated tagging in our analysis. Our algorithm had to handle more than 15,000 candidate terms, and if a term was extremely broad, very niche or ambiguous, it may not have been assigned cleanly to a category.
We also had to make judgment calls due to time and resource constraints about which terms to highlight. We may later refine the tag list, so if you notice an important concept missing, please let us know—we appreciate the feedback and can include it in a future update.
Q: Why are some terms grouped together? What are “parent” terms?
A: Some entries in our dataset represent parent terms for entire neurotransmitter systems, such as the dopaminergic system. These systems include many molecular components and derivatives, such as transmitters, pharmacological agents, metabolites, receptors and transporters that are part of the same signaling pathway. Rather than accommodate each one separately on the map, we group them at the neurotransmission system level.
In contrast, when terms are simply alternative names or refer to variants of the same method or construct, such as “functional brain imaging,” “functional MRI” and “fMRI,” we normalize or merge them.
Q: My search term produces results with many fewer noted publications than I would expect. Why?
A: Our publication counts per term are intentionally conservative. Rather than counting every occurrence of a word in every paper, we tag a paper with a term only if that term is identified as one of the paper’s key topics. In other words, we used a large language model-based keyword extraction process to find the main topics of each article, instead of doing a raw text search. This tactic avoids inflating counts with passing mentions. (For example, an article might name-drop many techniques or diseases in the introduction without focusing on them.) It also prevents errors from short or common words that appear inside other terms. Thus, the totals you see for a given term may be lower than a naive full-text count, but the trends over time should be more meaningful and reflective of genuine research focus than if we used a raw search.
Q: The trends over time for certain words are very cyclic (up one year, down the next, up again the year after that). Why?
A: For terms with low publication counts, the trend lines can appear erratic or “noisy” from year to year. If a term is associated with only a handful of papers annually, a small change (say, 5 papers one year and 10 the next) translates into a large percentage swing, which can look like a cyclic rise-and-fall pattern. In contrast, high-frequency terms (hundreds of papers per year) produce less “cyclic” curves. We’ve tried to reduce this noise by setting a minimum frequency threshold so that extremely rare terms are filtered out or aggregated. If you still see a jagged pattern for a particular term, it likely means that term’s usage is near the lower limit of detection. Proportional variability is expected when numbers are small, and we’re exploring methods to smooth these trends further.
Q: In your description of your methods, you say “The model was specifically instructed to preserve the exact terminology from the source text to ensure fidelity.” How did you validate this?
A: We took specific steps to make sure the keywords our model extracted were exactly the same as the terms used in the source papers. The large language model of choice was given explicit instructions (system message) to preserve original phrasing, meaning that if a paper mentions “hippocampal formation,” the model should output “hippocampal formation” as a tag rather than a synonym such as “hippocampus.” We also provided two example extractions to reinforce this rule.