AI maps a century of aging research, surfaces trends-and the gaps holding back translation
BUFFALO, NY - December 23, 2025 - Researchers analyzed 460,000+ PubMed abstracts (1925-2023) to build a thematic map of aging science. The result: clear shifts in focus, growing separation between basic and clinical work, and specific pairings that deserve more attention.
"This study outlines shifting priorities and translational gaps in aging research and offers a scalable, data-driven alternative to conventional reviews."
Key findings at a glance
- Shift in focus: early work centered on cellular biology and animal models; recent decades lean harder into clinical geriatrics and neurodegeneration (especially Alzheimer's and dementia).
- Basic vs. clinical split: both areas grew, but often in parallel with limited cross-talk. Clinical studies emphasize care and disease management; basic science leans into oxidative stress, telomeres, mitochondria, and senescence.
- Emerging topics still siloed: autophagy, RNA biology, and nutrient sensing are expanding but remain poorly tied to clinical endpoints.
- Strong connections persist: cancer-aging links remain robust. Meanwhile, promising but underexplored pairs include mitochondrial dysfunction ↔ senescence and epigenetics ↔ autophagy.
Why this matters for your work
If you lead a lab, run a clinical program, or allocate funding, this map shows where effort compounds-and where it stalls. The field is producing more data than any one expert can track. An AI-first overview exposes missed bridges that could accelerate bench-to-bedside progress.
Methods-built for scale and repeatability
- Corpus: 460,000+ abstracts on aging (1925-2023).
- Pipeline: topic modeling (LDA), TF-IDF term weighting, dimensionality reduction, and clustering to define themes and their trajectories over time.
- Semantic overlap: measured how top terms from one cluster score across others to flag well-studied and neglected relationships.
Figure highlight: finding underexplored connections
The reported heatmaps show how the top terms from each cluster "light up" (or don't) when tested against other clusters. This surfaces the top three most-studied and least-studied relationships, both across the full corpus and within biology-of-aging (BoA) clusters. Practical takeaway: if your work sits in mitochondria or epigenetics, you have clear opportunities to connect with senescence and autophagy research lines that remain underserviced.
Actionable steps for researchers and teams
- Design studies around the gaps: pair mitochondrial readouts with senescence markers; add epigenetic profiling to autophagy interventions.
- Co-author across silos: match basic labs with geriatrics and neurology groups to predefine translational endpoints.
- Rebalance portfolios: maintain Alzheimer's focus, but reserve budget for cross-domain projects that test mechanistic insights in clinical contexts.
- Adopt literature-mapping in-house: run periodic NLP scans of your niche to track saturation, novelty, and overlap.
- Pre-register "bridge" trials: explicitly link mechanistic biomarkers to clinical outcomes in protocol design.
For data and AI teams supporting research
- Operationalize the pipeline: automate updates with quarterly PubMed pulls; version the topic model to monitor drift.
- Add time-series: quantify rise/fall of topics to forecast where evidence will be abundant or scarce next year.
- Overlay funding signals: correlate topic momentum with grants and policy priorities to anticipate review preferences.
- Deliver dashboards: expose overlap scores and cluster proximity to PIs for study ideation and portfolio reviews.
Context and caveats
- Abstract-only data can miss nuance; full-text analyses could refine connections.
- Indexing and keyword drift over decades introduce bias; periodic recalibration and expert review help.
- Despite these limits, the method is transparent, scalable, and repeatable across subfields.
Open access, citation, and resources
Paper: A natural language processing-driven map of aging research (Aging-US, Vol. 17, Issue 11, Nov 25, 2025). DOI: https://doi.org/10.18632/aging.206340
Abstract video: Watch on YouTube
Correspondence: Jorge Sanz-Ros - jsanzros@stanford.edu
COI: The authors report no conflicts of interest. Large language models assisted coding in Google Colab.
Copyright: © 2025 Perez-Maletzki and Sanz-Ros. Open access under CC BY 4.0, permitting use and redistribution with attribution.
If you're building similar analyses
- Start small: replicate the pipeline on your lab's publications plus a focused PubMed query.
- Use the overlap matrix to flag "missing conversations" and propose cross-domain grants.
- Upskill your team on practical NLP for literature mining: curated AI courses.
Your membership also unlocks:
