Overview

The Concept Association Score is a statistical measure that evaluates the relationship between two biological concepts by analyzing their indirect connections within a knowledge graph. This guide explains what the score represents, how it is calculated (with visual examples), and how it can help you interpret associations.  

What Does the Score Mean?

The Concept Association Score shows how strongly two biological concepts are connected, even if they lack a direct link. It compares their shared connections against similar concept pairs in a reference set and expresses the result as a percentile rank:

• High Percentile Rank: Indicates a stronger connection based on shared indirect relationships.
• Low Percentile Rank: Suggests weaker or no meaningful association.

The score is particularly useful for uncovering hidden or indirect relationships in complex datasets

How Is the Score Calculated?

The Concept Association Score is derived through three key steps: building concept profiles, calculating their shared overlap using the inner product, and comparing the result to a reference set. Below, we explain these steps in detail and include an example calculation. 

Step 1: Creating the Concept Profile. 

A concept profile represents all the connections a biological concept has in the knowledge graph, weighted by the similarity of the connected concepts. 

See the Image above, where Concept A connects to other concepts (C1, C2, C3, etc.). Each connection has a weight (e.g., WA,C1). The concept profile (CP) is the summation of these weights for all connections.  

  
Example:
Suppose Concept A has connections to: C1 with a weight of 0.8, C2 with a weight of 0.6, C3 with a weight of 0.7, C4 with a weight of 0.3

The concept profile for A is: CP(A) = [0.8, 0.6, 0.7, 0.3]

Similarly, Concept B connects to related concepts (including some that overlap with A): C1 with a weight of 0.7, C2 with a weight of 0.5, C5 with a weight of 0.6, C3 with a weight of 0.4

The concept profile for B is:CP(B) = [0.7, 0.5, 0.6, 0.4]
  

Step 2: Creating the Concept Profile.

The inner product measures the overlap between the two concept profiles by multiplying the weights of shared connections and summing the results.

See the Image above, where Concept A and Concept B share connections with some common concepts (e.g., C1, C2, C3). Multiply the corresponding weights of shared concepts and sum them to calculate the inner product.. The concept profile (CP) is the summation of these weights for all connections.  

  
Example:
From CP(A) and CP(B), the shared connections are: C1: WA,C1 × WB,C1 = 0.8 × 0.7 = 0.56, C2: WA,C2 × WB,C2 = 0.6 × 0.5 = 0.30, C3: WA,C3 × WB,C3 = 0.7 × 0.4 = 0.28

The inner product is:0.56 + 0.30 + 0.28 = 1.14
  

Step 3: Comparing Against the Reference Set.

The inner product score (1.14 in this example) is compared to a reference set of similar concept pairs that are known to be directly related. The reference set consists of other pairs of concepts within the same semantic category as A and B. 

The scores from the reference set form a frequency distribution. The percentile rank of the calculated inner product is determined relative to this distribution.

  
Example:
If the reference set’s distribution of inner product scores is:
   •    25th percentile = 0.8
   •    50th percentile = 1.0
   •    75th percentile = 1.2
  

The scores from the reference set form a frequency distribution. The score of 1.14 falls between the 50th and 75th percentiles. Thus, the Concept Association Score for A and B is approximately in the 70th percentile, indicating a moderately strong association.

How Can You Use the Score?

The score helps you identify and rank associations in your research. A high score suggests that two concepts are likely to have a meaningful connection worth exploring further, while a low score indicates weaker or less significant associations. 

By visualizing the relationships and understanding how the score is derived, as shown in Images 1 and 2, you can better interpret and validate the results in your biological research.