Relationship Types in the Knowledge Graph and What They Mean

The short answer

A connection in the Research Library says that two records are linked and where that link came from. It does not say that one thing works for the other. The library shows connections and does not rate effectiveness, and a record with many connections is not better studied or more effective than one with few.

This page explains the three kinds of connection the library uses, what each can and cannot tell you, and how to read the wording of a link without reading a result into it. For the wider picture, start with how to read Research Library records.

What a "knowledge graph" is here

A graph is a set of records (herbs, ingredients, genes, conditions, studies) joined by connections. The Knowledge Graph lets you move from one record to the next. Each connection has a type, meaning a short label for the sort of statement being made, such as "this study names this herb" or "this herb contains this compound."

A type matters because different statements carry very different weight. "A trial was registered about this herb" is a different kind of statement from "this herb has been used traditionally for this complaint," and both differ from "this compound binds this protein in a laboratory test."

The three sources of connections

Every connection in the library comes from one of three places.

Why text matching can go wrong

Matching names is harder than it looks. One published evaluation of SemRep, a program that pulls relationships out of PubMed abstracts, found that recognizing and normalizing names was the largest single source of its errors, at 26.9 percent of the errors it analyzed. [1] That is a different tool from the one behind the library's name matches, so it is an illustration and not a measurement of the library. It shows why a name match should be treated as a lead to check against the study itself.

Herb names make this worse. In traditional Chinese medicine, one herb name can refer to several species, and in one documented case only one of the three plants sharing the name contains aristolochic acid, a compound linked to kidney harm. [2] A link made only on a shared name can therefore join two things that are not the same.

Reading link wording without reading a result into it

Plain phrases appear in many databases to describe links. Here is how to read the common ones.

• Contains (herb to compound). A source lists a compound as found in a plant or preparation. In one analysis of 16 commercial herbal tea samples, products labeled with the same species varied widely in their chemistry, so "contains" does not tell you how much. [3]
• Studied for (subject to condition). A study record names both. This tells you research exists or is planned. Check whether it is completed, whether it involved people, animals or cells, and what it found. See registered trials and completed studies and human, animal and laboratory research.
• Traditionally used for. A historical or traditional source records a use. That is a statement about practice, not a clinical finding. NCCIH notes, for example, that Chinese herbal products have been studied for many conditions, yet because many studies were of poor quality no firm conclusions can be made about their effectiveness. [4]
• Interacts with / acts on (compound to gene target). A source links a compound to a gene or protein. This is a proposed mechanism. A mechanism is a possible explanation of how something might act, and it is not the same as an outcome in people. Gene records are covered in genes and biomarkers.

How strong is each kind of evidence

Research types differ in what they can show. According to NCCIH, clinical trials give the clearest information about whether a treatment or lifestyle change is effective and safe in humans. [5] Laboratory and mechanism links sit earlier in that chain. Use the evidence types guide to see where a linked study fits.

Common misunderstandings

• "Many connections means well proven." No. The count of connections reflects how many records name or list the item, not how well it works.
• "A link means a result." No. A link may point to a plan, a laboratory test or a historical note.
• "Imported means checked." No. The default state is "imported." Where a page states that something was editorially checked, it will say so.
• "Name matches are as good as study links." No. They can be wrong and should be confirmed.