Biological Reasoning · wanglab
kegg
KEGG pathway entries paired with genomic variants and reasoning traces to evaluate mechanistic disease prediction.
What kegg is
The KEGG Biological Reasoning dataset pairs pathway information with genomic variant records to support the development and evaluation of multimodal biological reasoning models. Created by researchers at the Wang Lab for the BioReason project, the resource focuses on modelling how genetic alterations affect cellular pathways and result in clinical conditions.
The dataset addresses the challenge of mechanistic variant-to-disease interpretation within DNA language models. By coupling curated biological pathway data with step-by-step reasoning traces, it provides structured evidence for predicting pathological outcomes across human diseases.
How it was made
The dataset contains 1,449 entries extracted from the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database, paired with variant information sourced from ClinVar, dbSNP, OMIM, and COSMIC. Each record includes mechanistic reasoning traces explaining how genetic variants contribute to disease processes across 37 unique diseases.
The corpus is structured as text and provided in Parquet format, making it directly accessible for training and evaluation in standard machine learning workflows.
Key facts
- Total entries
- 1,449
- Disease coverage
- 37 unique diseases
- Source databases
- KEGG, ClinVar, dbSNP, OMIM, COSMIC
- Data format
- Parquet
- Modality
- Text
What you can use it for
- Biological reasoning evaluation
- Benchmarking the multi-step reasoning capabilities of DNA-LLM and multimodal biological architectures.
- Mechanistic variant interpretation
- Training models to explain the functional progression from a specific genomic mutation to a clinical phenotype.
- Pathway-level disease prediction
- Predicting disease outcomes by connecting annotations from ClinVar, dbSNP, OMIM, or COSMIC with KEGG biological pathways.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from wanglab/kegg at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 1,159 rows of default/train
answer
string_label- alzheimer's disease15%
- parkinson's disease14%
- amyotrophic lateral sclerosis13%
question
string_label- Chromosome Number: 14 Network Definition of the pathway: ATXN3* // (BECN1+PIK3C3+ATG14+PIK3R4+AMBRA1) Genes in the pathway: ATXN3; ataxin 3 | BECN1; beclin 1 | PIK3C3; phosphatidylinositol 3-kinase catalytic subunit type 3 | ATG14; autophagy related 14 | PIK3R4; phosphoinositide-3-kinase regulatory subunit 4 | AMBRA1; autophagy and beclin 1 regulator 1 Given this context, what is the biological effect of this ATXN3 allele, specifically what disease does this contribute to?2%
- Chromosome Number: 14 Network Definition of the pathway: "Glutamate -> GRM1 -> GNAQ -> PLCB -> IP3 -> (ITPR1+(ATXN2*,ATXN3*)) -> Ca2+ -- MCU -> Ca2+(mito) -- MPTP -> CYCS" Genes in the pathway: GRM1; glutamate metabotropic receptor 1 | GNAQ; G protein subunit alpha q | PLCB1; phospholipase C beta 1 | PLCB2; phospholipase C beta 2 | PLCB3; phospholipase C beta 3 | PLCB4; phospholipase C beta 4 | ITPR1; inositol 1,4,5-trisphosphate receptor type 1 | ATXN2; ataxin 2 | ATXN3; ataxin 3 | MCU; mitochondrial calcium uniporter | VDAC1; voltage dependent anion channel 1 | VDAC2; voltage dependent anion channel 2 | VDAC3; voltage dependent anion channel 3 | SLC25A4; solute carrier family 25 member 4 | SLC25A5; solute carrier family 25 member 5 | SLC25A6; solute carrier family 25 member 6 | SLC25A31; solute carrier family 25 member 31 | CYCS; cytochrome c, somatic Given this context, what is the biological effect of this ATXN3 allele, specifically what disease does this contribute to?2%
- Chromosome Number: 14 Network Definition of the pathway: PSEN1* -> ITPR -> Ca2+ -- MCU -> Ca2+(mito) -- MPTP -> CYCS Genes in the pathway: PSEN1; presenilin 1 | ITPR1; inositol 1,4,5-trisphosphate receptor type 1 | ITPR2; inositol 1,4,5-trisphosphate receptor type 2 | ITPR3; inositol 1,4,5-trisphosphate receptor type 3 | MCU; mitochondrial calcium uniporter | VDAC1; voltage dependent anion channel 1 | VDAC2; voltage dependent anion channel 2 | VDAC3; voltage dependent anion channel 3 | SLC25A4; solute carrier family 25 member 4 | SLC25A5; solute carrier family 25 member 5 | SLC25A6; solute carrier family 25 member 6 | SLC25A31; solute carrier family 25 member 31 | CYCS; cytochrome c, somatic Given this context, what is the biological effect of this PSEN1 allele, specifically what disease does this contribute to?1%
reasoning
string_text- min
- 1,640
- median
- 2,195
- mean
- 2196.32
- max
- 3,031
reference_sequence
string_text- min
- 4,001
- median
- 4,001
- mean
- 4005.69
- max
- 4,070
variant_sequence
string_text- min
- 4,000
- median
- 4,001
- mean
- 4008.49
- max
- 4,260
Strengths and limits
Good at
- Supplies explicit reasoning traces detailing the mechanistic links between genetic variants and disease states.
- Integrates verified variant annotations across four prominent genomic databases: ClinVar, dbSNP, OMIM, and COSMIC.
- Provides standardised pathway representations spanning 37 distinct disease categories.
Watch out for
- The dataset comprises 1,449 total entries, which is a modest sample size for deep learning applications.
- Phenotypic scope is restricted to the 37 unique diseases represented in the curated set.
Questions
Can I look inside kegg without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is kegg under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
Which databases are integrated into this dataset?
Entries combine biological pathway data from KEGG with genomic variant annotations drawn from ClinVar, dbSNP, OMIM, and COSMIC.
What is the primary objective of the reasoning traces?
The reasoning traces outline mechanistic steps connecting variants to diseases to guide and test biological reasoning in computational models.
How many diseases and records are included?
The dataset contains 1,449 entries covering 37 unique diseases.
How to cite it
Adibvafa Fallahpour, Andrew Magnuson, Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah, Haonan Duan, Omar Ibrahim, Hani Goodarzi, Chris J. Maddison, Bo Wang (2025). BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model. arXiv:2505.23579.
Written from kegg's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.