Varmovinator

The world's best (and first) variant learning model validator (outside CAGI challenges).

For a given variant, this will pull basic information from standard bioinformatic databases such as

• disease association ^1,2
• gene ontology ²
• uniprot information ²
• environmental factors that influence penetrance
• putative molecular consequences ¹
• putatative clinical consequences ¹
• population frequency ¹
• protein effects ²
• gene ²

1: Opencravat 2: Uniprot 3: LitVar

The data will be pulled using SPARQL. An example SPARQL query is included.

We have selected the following three models to test:

1. BioGPT
2. StabilityLM (Stable Vicuna 13-b)
3. GPT4 3a. GPT4 prompted with a KG

Questions we are asking the model:

• What diseases is rsX associated with?
• Expand the text rsX
• What environmental factors affect the penetrance of rsX?
• What transcription factor pathways are affected by rsX?

X's tested in this case:

• rs6003
• rs80357914
• rs563410947

We will come up with a model to

• tag discrepant assertions
• calculate a consistency score

Note: some answers are non-sensical, or at least decontextualized; indicating that some models likely need to be at least prompted