What my Paperpile library says about me ↗
For something a little different: I exported my Paperpile library (2,052 papers, ~12 years of saving things I wanted to read) and pointed Claude Code at it. Asked it what my reading habits say about me. Took an afternoon. Some of what fell out was predictable. One thing wasn't.
I was particularly interested in how my scientific interests shifted since starting my own lab: It's always good to have your priors checked.
A UMAP of the recent 1,264 papers (TF-IDF on titles and abstracts → KMeans → UMAP) shows the structure clearly. Microglia, neurodegeneration, and glia cluster at one end. Computational methods and LLM papers at the other. Sharper than I expected from a personal library.
(Caveats: bookmark ≠ read)
But one thing actually surprised me. The report flagged an unusual tight cluster: 49 papers on cardiolipin and phospholipid biochemistry, with Mike Schlame as senior on seven of them, strong content on tafazzin, Barth syndrome, and membrane remodeling enzymes. Claude noted it was a narrow niche for an aging-biology library, sitting tightly enough that it looked clearly project-driven rather than background reading.
It was. This was a now-iceboxed project from my PhD days in Cologne. We had seen that cardiolipin in the liver of caloric-restricted mice, a highly specialized membrane lipid unique to mitochondria, shifted to a very specific combination of carbon atoms and double bonds (72 carbons, 8 double bonds). These molecules are huge and can take on a variety of compositions, so it was genuinely odd that the metabolically active liver would enact this big shift.
Cardiolipin anchors the respiratory chain complexes in the mitochondrial membrane. At that time, supercomplexes were all the craze. The natural follow-up was whether dietary restriction (DR) vs. ad libitum (AL) remodels mitochondrial supercomplexes through cardiolipin composition. Steady-state supercomplex levels didn't budge across days or tissues. But in-vitro import suggested DR reduced supercomplex assembly - possibly a turnover effect. Real signal, never finished. We ran out of runway and the project got iceboxed.
I'd genuinely forgotten how much I cared about it until the analysis surfaced the cluster. Made me dig through old lab meeting decks tonight.
The bigger point: AI literature analysis isn't just about generating new ideas. Sometimes it surfaces old ones you forgot you cared about.
Would love to see the trends you find in your own literature list.
