AI-Assisted Research Pipeline
A four-stage workflow for discovering, validating, and synthesizing scientific literature into actionable intelligence.
This system runs on private infrastructure. Below is a walkthrough of how it works in practice.
See It In Action
Two research sessions showing the pipeline in practice — from literature review to study design.
Mapping a Research Landscape
I want to map the landscape of FGF ligands acting in the brain beyond FGF21. What do we know about FGF1, FGF19, and alpha-Klotho in brain metabolic circuits? Start by checking what’s already in my library, then search PubMed for what’s missing.
vector_search — “FGF19 brain energy homeostasis” — 3 matches
vector_search — “alpha-Klotho brain FGFR1” — 2 matches
vector_search — “FGFR1 hypothalamus mapping expression” — 3 matches
read — Reading full text of 12 papers from library
You already have 12 relevant papers in your library:
FGF1
- Scarlett et al., Nat Med 2016 — Single ICV injection produces sustained diabetes remission (~17 weeks)
FGF19
- Morton et al., 2013 — Central FGF19 improves glucose tolerance
- Marcelin et al., Mol Metab 2014 — Central FGF19 activates ERK, reduces AgRP
FGFR1 / Receptor Mapping
- Hultman et al., J Comp Neurol 2019 — Atlas of beta-Klotho and FGFR expression across the mouse brain
Plus 8 reviews and FGF21-brain papers from your lab. Major gaps: no FGF1 mechanistic follow-ups since Scarlett 2016, no alpha-Klotho papers, no FGF19 hindbrain or behavioral work. Searching PubMed now.
pubmed — “FGF19” AND “brain” AND (“food intake” OR “glucose”) (2018–2026) — 9 results
pubmed — “alpha-Klotho” AND “hypothalamus” AND (“AgRP” OR “food intake”) — 6 results
pubmed — “FGF1” AND “tanycyte” — 4 results
pubmed — “FGFR1” AND “hypothalamus” AND “metabolic” (2020–2026) — 11 results
pubmed — “FGF1” AND “brain” AND “astrocyte” — 7 results
pubmed — “central FGF19” AND “foraging” OR “AgRP” — 3 results
vector_search — Cross-referencing all results against library (57 vault queries)
filter — Removing duplicates, reviews already covered, tangential hits
Found 19 papers you’re missing, across three themes:
FGF1 — Mechanistic Follow-ups
Single-cell transcriptomics showing glial > neuronal response, MAPK/ERK signaling requirements, tanycyte-specific FGFR1 mechanisms, and astrocyte remodeling
FGF15/19 — Behavioral, Transport, and Clinical
Foraging/AgRP behavioral effects, blood-brain barrier transport, and liver/MASH clinical translation (aldafermin, FGF19 mRNA therapy)
Alpha-Klotho — FGFR1 Independence
Alpha-Klotho suppressing AgRP via hypothalamic FGFR1, beta-Klotho–independent mechanisms, and AAV-Klotho overexpression in aging
Want me to add these to your library?
Yes, add all 19. Then build a cross-ligand comparison — what endpoints does each affect, and where do the mechanisms converge or diverge?
pipeline — PMC full text download → markdown conversion → frontmatter enrichment → QMD indexing
read — Reading full text of all 31 papers (12 existing + 19 new)
All 19 in your library. Cross-ligand comparison from the full collection of 31 papers:
Where They Converge
- All four ligands signal through FGFR1 — the common node
- FGF1, FGF19, and alpha-Klotho all suppress AgRP neurons
- All improve glucose tolerance
Where They Diverge
- FGF21 does NOT work through the ARC. Your mapping data puts its receptors in NTS, SCN, VMH, VTA — anatomically dissociated from the others
- FGF1’s sustained remission is unique. Tanycyte/glial remodeling may explain why a single injection corrects diabetes for weeks
- Alpha-Klotho doesn’t need beta-Klotho — accesses FGFR1 through a completely different co-receptor mechanism
The Key Tension
Is FGF1 actually physiological? It’s a paracrine factor, not an endocrine hormone — no signal peptide, not circulating. The diabetes remission is real, but may be a pharmacological effect of flooding hypothalamic FGFRs, not recapitulating normal biology. Alpha-Klotho has the strongest case for physiology: it circulates at measurable levels, declines with age, and the doses used are within a relevant range.
Good. Save this as a working document — established facts first, then the interpretive questions.
index — 19 new papers fully indexed in science vault
Saved. The document covers your existing library inventory, the gap analysis, the cross-ligand endpoint comparison, mechanism convergence and divergence, and the physiology vs. pharmacology assessment. All 19 new papers are indexed for future semantic search.
7 PubMed searches · 57 vault queries · 31 full-text reads · 19 papers imported · 1 session
Designing a Human Study
I have an idea about protein restriction in humans. But I want to know how the famous calorie study was run. That study conducted long-term calorie restriction in humans, used metabolic chambers, etc. How could that work inform a similar design using protein restriction, or perhaps comparing calorie to protein restriction.
vector_search — “CALERIE caloric restriction trial”, collection: science — 1 match
The “famous calorie study” is the CALERIE trial (Comprehensive Assessment of Long-term Effects of Reducing Intake of Energy) — a landmark 2-year, 3-site RCT of 25% caloric restriction in non-obese humans. A protein restriction analog using metabolic chambers would be a natural successor.
I found one CALERIE paper already in your library (Heilbronn et al. 2006, the Phase 1 results). But for a serious study design, you’d want the full set of methods and outcomes papers.
Core CALERIE Publications
| Heilbronn et al., JAMA 2006 | In library | Phase 1 results (6 months) |
| Rochon et al., Contemp Clin Trials 2011 | Missing | Phase 2 study design and methods |
| Ravussin et al., J Gerontol A 2015 | Missing | Phase 2 primary outcomes (2 years) |
| Kraus et al., Lancet Diab Endo 2019 | Missing | Phase 2 cardiometabolic outcomes |
| Das et al., Am J Clin Nutr 2017 | Missing | Body composition & energy expenditure |
Want me to pull the full text for the four missing papers and add them to your library?
Yes, add all four. Then give me an overview of the CALERIE methodology — I need to understand how they actually ran the trial before I can think about a protein restriction version.
pipeline — PMC full text download → markdown conversion → frontmatter enrichment → QMD indexing
read — Reading full text of all 5 CALERIE papers from vault
All four papers are now in your library with full text. Here’s the CALERIE methodology synthesized across all five publications:
Study Design
- Phase 1 (Heilbronn 2006): 6-month pilot, 48 overweight adults, 4 arms (control, 25% CR, 12.5% CR + exercise, very-low-calorie diet)
- Phase 2 (Rochon 2011, Ravussin 2015): 2-year definitive trial, 218 non-obese adults (BMI 22–28) across 3 sites, randomized 2:1 to 25% CR vs. ad-lib control
How They Delivered the Diet
- Phase 1 (first 28 days): All meals provided by the metabolic kitchen — individually calibrated to each participant’s measured energy needs
- Phase 1 (months 2–6) and Phase 2: Participants transitioned to self-selected diets with counseling. No specific macronutrient targets — just required nutritional adequacy while maintaining the caloric deficit
- Behavioral support was extensive: individual counseling, group meetings, weekly weight trajectory graphs, web-based food tracking, and escalating interventions when participants fell off track
How They Confirmed Adherence
This is the critical part for your design:
- Doubly labeled water (DLW) was the gold standard — measured actual total energy expenditure at baseline, 6, 12, and 24 months
- Self-reported food diaries were NOT trusted — participants consistently underestimated intake
- Achieved CR was ~12%, not 25%. The prescribed target was 25%, but DLW confirmed participants actually achieved roughly half that — still enough to produce significant effects, but a critical lesson for study design
OK, so that’s the adherence problem. If they aimed for 25% and got 12%, a protein restriction study will face the same challenge. How would you actually monitor adherence to protein restriction specifically? You can’t use DLW for protein.
vector_search — “protein intake measurement biomarker urine”, collection: science — 2 matches
You’re right — DLW tells you about total energy expenditure, not macronutrient composition. For protein restriction specifically, you need a parallel monitoring strategy:
Gold Standard: 24-Hour Urinary Urea Nitrogen (UUN)
- Most dietary protein is metabolized and excreted as urea — if someone eats more protein than prescribed, it shows up
- Limitation: requires complete 24h urine collection. PABA marker pills can verify completeness
Biological Confirmation Biomarkers
This is where it gets interesting for your work:
- IGF-1 — drops with protein restriction but NOT with caloric restriction alone. A clean discriminator
- FGF21 — increases with protein restriction. Since this is your primary mechanism of interest, it doubles as both an adherence marker and a study endpoint
- T3 (triiodothyronine) — drops with both CR and PR, less specific, but useful in combination
The IGF-1 + FGF21 + T3 triad would let you distinguish true protein restriction adherence from inadvertent caloric restriction — something CALERIE couldn’t do.
This is really useful. Save everything from this session — the CALERIE methods overview and the protein adherence monitoring plan — as a working document.
index — 4 new CALERIE papers fully indexed in science vault
Saved. The document covers CALERIE design overview, diet delivery methods, the adherence measurement challenge (DLW limitations for protein studies), and your protein-specific monitoring strategy with the IGF-1/FGF21/T3 triad. The four new CALERIE papers are fully indexed — they’ll come up in future semantic searches related to dietary intervention design.
How It’s Built
A summary below — full architecture and design decisions are on GitHub.
Discovery & Screening
The AI Assistant (Communications Director) broadly searches new publications and preprints, semantically scores them against an embedding of a curated research profile, and pushes these to the Research Digest.
Curation & Validation
The Research Digest represents a single, aggregated source of new research matching my interests. AI summaries and full text, already held in a backend database, are quickly reviewed. Papers of interest are selected for automated addition to the science vault/library, along with the generation of AI-assisted full text summaries.
Actionable Intelligence
The User and LLM can utilize both the curated knowledge base and PubMed searches to reason over topics, generate summaries, identify gaps, and make unexpected connections. If missing papers are identified, they are directly integrated and embedded to expand the knowledge base.
Integration & Synthesis
Abstracts and full text manuscripts are saved into a literature vault, embedded (QMD) for semantic search, and then classified by thematic area to generate a literature map. This forms the foundation for a local knowledge base.