Does Chronic Stress Really Lower Hippocampal BDNF? What an LLM Read of 9,585 Studies Actually Says

This is an accessible companion to our preprint. Preprint (Research Square, v1, not yet peer-reviewed): https://www.researchsquare.com/article/rs-9884522/v1 · DOI 10.21203/rs.3.rs-9884522/v1. This is a scoping review of what the literature reports — it measures reporting patterns across abstracts, not pooled biological effect sizes. Nothing here is clinical advice.

TL;DR (Quick Answer)

We used a large language model to read and structure 9,585 PubMed abstracts on chronic stress and the brain (2008–2026), then asked a simple question: do the textbook one-liners actually hold across the whole literature? Several don't — at least not as stated.

• "Stress lowers hippocampal BDNF" is not what the literature reports. Across 1,434 BDNF–hippocampus entries, decrease (46%) and increase (46%) are almost exactly tied — and the balance has flipped over time, from decrease-dominant (56% in 2010–2014) to increase-dominant (57% in 2020–2026).
• Neuroinflammation has overtaken the HPA axis. Inflammation genes (IL1B, TNF, IL6, NFKB1, NLRP3; 1,143 mentions) passed cortisol/HPA genes (NR3C1, CRH, FKBP5; 597 mentions) around 2018 (inflammation/HPA mention ratio: 0.32 → 1.03 → 2.15 across the last three half-decades).
• The "amygdala grows under stress" story conflates two different measurements. Separating structural from functional reports resolves it: the hippocampus and amygdala both trend structural decrease but functional increase.
• The biggest blind spot is the insula × inflammation axis — combinations like NR3C1/IL1B/TNF/IL6 × insula have zero abstracts each despite >10 expected.

The honest caveat that frames everything below: these are reporting patterns — the aggregate direction of what abstracts say — not a meta-analysis of independent effect sizes.

Background: a field big enough that nobody has read all of it

The chronic-stress literature is enormous, and most of our "settled" claims — stress shrinks the hippocampus, enlarges the amygdala, lowers BDNF, runs through cortisol — come from narrative reviews or modest meta-analyses that each cover a slice of it. Nobody has read every paper. So a reasonable question is whether those narratives survive contact with the full body of literature, read end to end.

That's now feasible. Instead of hand-curating a few hundred studies, you can have a language model extract structured fields from tens of thousands of abstracts and look at the distribution.

How we read 9,585 studies (and what that does and doesn't buy you)

We pulled 10,001 unique PubMed articles (2008–2026) across six query themes, then had Qwen3.5-397B (a 397B-parameter open model, 3-bit quantized) extract structured data from each abstract — study type, stress type, gene mentions (HUGO symbols), brain regions with an effect direction (decrease/increase/no change), and the measurement metric (structural/functional/cellular). 9,585 abstracts (95.8%) parsed cleanly. The model ran on a GPU-less CPU cluster with no paid APIs — the engineering backstory (running a ~400B model across cheap CPU nodes with llama.cpp) is its own write-up here.

The single most important thing to understand about this method:

It measures discourse, not biology. An abstract saying "hippocampal volume decreased" might be reporting its own result, citing someone else's, or stating a hypothesis. So what we count is the aggregate direction of what the literature reports — useful for seeing narratives, trends, and gaps, but not a substitute for pooled effect sizes. We use "mentions" and "reporting patterns" deliberately throughout. (We also re-ran everything on primary studies only, N = 8,487, excluding 902 reviews and 196 meta-analyses, to check that citation echo wasn't driving the trends.)

Finding 1 — BDNF in the hippocampus: a coin flip, and the coin is turning

The canonical claim is that stress lowers BDNF (brain-derived neurotrophic factor) in the hippocampus. Across 1,434 BDNF–hippocampus entries, the reports split 46% decrease / 46% increase / 8% no change — essentially even. And it's time-dependent: decrease-dominant in 2010–2014 (56%), increase-dominant in 2020–2026 (57%).

It's also stress-type-dependent, which is probably the practical lesson: early-life stress skews toward BDNF decrease (55%), while acute stress skews toward increase (46%). Pooling all "stress" together is what manufactures the tidy-but-misleading single direction.

Finding 2 — Neuroinflammation overtook the HPA axis around 2018

For decades the stress story was a cortisol story: the HPA axis, glucocorticoid receptor (NR3C1), CRH, FKBP5. The literature has quietly moved. Counting gene mentions, the neuroinflammation set (IL1B, TNF, IL6, NFKB1, NLRP3; 1,143 mentions) overtook the HPA set (NR3C1, CRH, FKBP5; 597 mentions), with the crossover around 2018:

Period	Inflammation / HPA mention ratio
2010–2014	0.32 (HPA dominant)
2015–2019	1.03 (crossover)
2020–2026	2.15 (inflammation dominant)

If you design stress studies, the implication is concrete: measure inflammatory markers (IL-6, TNF-α, IL-1β) alongside cortisol, because that's where the field's attention — and likely its next mechanisms — now sits.

Finding 3 — The amygdala "enlarges" only if you mix up your rulers

"Stress shrinks the hippocampus and enlarges the amygdala" is a staple. Splitting reports by measurement metric shows the contradiction is an artifact of mixing structural (volume, gray matter) and functional (activity, connectivity) measures:

Region	Metric	N	% Decrease	% Increase	Dominant
Hippocampus	Structural	975	67%	17%	↓ decrease
Hippocampus	Functional	661	36%	41%	↑ increase
Amygdala	Structural	504	50%	33%	↓ decrease
Amygdala	Functional	1,070	27%	45%	↑ increase

Both regions trend structural decrease but functional increase. The popular "amygdala gets bigger" line is really a functional hyperactivity story; structurally, the literature more often reports the amygdala decreasing. This structural-decrease / functional-increase split is consistent with the allostatic load framework (McEwen, 1999) — early compensatory hyperfunction giving way to structural wear.

Finding 4 — The largest gap: the insula and inflammation

Scanning co-occurrence across 2,157 genes and 1,930 brain regions for combinations that appear far less than expected surfaced 15 significantly under-studied pairings. The biggest: the insula × inflammation axis — NR3C1, IL1B, TNF, and IL6 each paired with the insula have zero abstracts despite >10 expected. The insula is central to interoception and emotional awareness; its near-absence from the molecular stress literature is a concrete, fundable gap.

Why this matters

• For reading the field: a top-line consensus ("stress lowers BDNF") can hide an even split once you read everything — and the split is structured (by stress type, by era, by measurement). Aggregate narratives are where nuance goes to die.
• As a method: LLM-powered extraction turns "too much literature to read" into a queryable map of what the field reports — good for spotting trend reversals and gaps that no single reviewer would catch — provided you treat the output as discourse, not as effect sizes.

Honest Limitations

1. Reporting patterns, not effect sizes. Mention counts are not a meta-analysis. Full-text extraction with effect sizes, CIs, and sample sizes is the proper next step (in progress).
2. Abstracts only, binary direction. Continuous effects are flattened to decrease/increase, and abstracts over-report positive findings (null results are likely under-counted — note the 8% "no change" for BDNF).
3. Mentioned ≠ studied. A gene named in an abstract may be discussed, not measured, inflating co-occurrence.
4. Keyword-based metric tags (structural/functional/cellular) can be misclassified.
5. No causal inference. The BDNF temporal reversal could reflect methodological/citation trends rather than biology.
6. Preprint, not peer-reviewed, and validated against only 50 manually-checked articles so far. Treat all of this as hypothesis-generating.

FAQ

Q: So does chronic stress lower BDNF or not?

The literature doesn't speak with one voice: across 1,434 hippocampal-BDNF reports it's ~46% decrease vs ~46% increase, and which way it leans depends on stress type (early-life → decrease) and publication era (recent years → increase). The safe statement is the specific one — name the stress type, region, and measurement — not the generic "stress lowers BDNF."

Q: Does "neuroinflammation overtook the HPA axis" mean cortisol doesn't matter?

No. It means the literature's attention shifted — inflammation genes are now mentioned about twice as often as HPA-axis genes. Cortisol biology is still real; the point is that inflammatory pathways are where most new work is pointing.

Q: Is this a meta-analysis?

No — and that distinction is the whole point. It's a scoping review of reporting patterns extracted by an LLM. It maps what the literature says and where the gaps are; it does not pool effect sizes. A quantitative full-text meta-analysis is the follow-up.

Q: Can I see the underlying studies?

Yes — the complete list of all 9,585 included studies (PMID, title, journal, year, PMC ID) is Supplementary Table S1 on the preprint page, so you can pull any source abstract/full text from PubMed/PMC. Structured extraction data and analysis code are available from the author on request.

Resources

• Preprint (Research Square, v1): https://www.researchsquare.com/article/rs-9884522/v1 · DOI 10.21203/rs.3.rs-9884522/v1
• All 9,585 studies: Supplementary Table S1 (on the preprint page)
• The engineering backstory (running a ~400B model on a GPU-less CPU cluster): Mining 10,000 papers on a GPU-less cluster
• Related on this site: How stress damages your brain · a sibling large-scale genetics companion (MDD drug targets)
• Framework: allostatic load (McEwen, 1999, NEJM).