Your decision to spend one hour on setting up a knowledge graph shapes the next six months of your literature review.

Here's the guide for starting one.

axy-app.com/blog/set-up-...

#KnowledgeGraphs

Same cell, two names: one paper calls it "CD16+ monocyte," another "non-classical monocyte." Neither tells you exactly where in the kidney or lung it sits. That's the real tax on merging single-cell/spatial datasets before analysis even starts.

ChatGPT nails a 10-second paper summary. That's not scientific discovery. Studies on GPT-4 found that it invents citations at a real rate as it generates tokens, rather than retrieving sources.
We summarized why that gap matters for research: axy-app.com/blog/why-cha...

Papers with Code went dark on July 24, 2025.
No announcement, No export window, No blog post
Just 502 errors where the SOTA leaderboards used to be. Seven years of ML's shared reference point, gone over a weekend.

PubMed Knowledge Graph 2.0 just linked 36M+ papers, 1.3M patents, and 480K+ clinical trials into one graph. A gene-disease finding from 2012 → a 2015 patent → a 2019 trial can be traced as a single path. That kind of infrastructure didn't really exist before.

We made a list of research tools that can actually save your time in academics, and no, that's not another 50-app roundup. Research Rabbit for literature discovery (free, visual, syncs with Zotero). Overleaf for LaTeX co-authoring. Get an ORCID if you still don't have one. 🧵

Before 2007: a fish list from the North Atlantic and a mollusc catalogue from the Mediterranean could disagree on the name for the same organism, with no shared identifier to reconcile them. WoRMS closed that gap. 🧵

You hit Figure 3 in a paper: a tangle of coloured dots joined by arrows pointing every direction. Caption just says "knowledge graph." You read the rest fine but this one figure stops you. It's not you. Nobody taught you the grammar. 🧵

DGL and PyTorch Geometric were both born in 2019. DGL's bet: treat the graph itself as the core abstraction , framework-neutral across PyTorch, TensorFlow, and MXNet rather than bolting sparse tensors onto an existing dense-tensor framework. 🧵

GBIF has 2.8 billion occurrence records across 10 million species from 1,600+ institutions in 197 countries. That still doesn't make it a random sample of life on Earth, and treating it like one has gotten papers published that didn't survive scrutiny. 🧵

Every knowledge graph whether GBIF for biodiversity, Monarch for gene-disease links, PubMed KG for the literature gets rebuilt in isolation. Connecting two of them by hand still takes ~2 weeks. That's the exact seam AI is now stitching. 🧵

You spend weeks on Google Scholar, Connected Papers, bioRxiv, conference abstracts - then hit a 15-year-old paper (sometimes in another language) that already tested your exact idea. That's not a research productivity problem. It's an infrastructure gap in science. 🧵