Compare genes
Each sample becomes 64 binary indicators: one when Gene A is expressed above Gene B, otherwise zero.
VENETOCLAX + AZACITIDINE RESPONSE
Cross-domain response prediction from gene expression
A research portal for the frozen ComboKR-GP model: robust gene-pair features, explicit cohort-direction removal and RBF prediction.
WITHIN-SAMPLE PAIRING
Gene A 8.4
>
Gene B 5.1
PAIR FEATURE 1
Relative ordering is preserved across many expression scales.
64gene-pair features
63unique genes
Localprivacy-first inference
PREDICTION WORKSPACE
Upload an expression matrix
CSV or TSV; samples can be rows or columns. All 63 required genes must be present. Pair features make the calculation insensitive to monotonic expression scaling.
- Runs entirely in your browser
- No patient data is uploaded
- Returns an uncalibrated response score
Drop your expression matrix hereCSV or TSV, maximum 20 MB
SCORE GUIDE
How to read the response score
The model returns a continuous, uncalibrated score. Higher values are more response-like, but crossing zero does not establish response and no clinical cutoff has been validated.
> 0.585Very highApproximately the top 10%
0.321–0.585High
-0.527–0.321Middle rangeIndeterminate
-0.827–-0.527Low
< -0.827Very lowApproximately the bottom 10%
Important interpretation
- 0 is not a response threshold. Different cohorts produced different optimal cutoffs.
- The score is not a probability. A score of 0.6 does not mean a 60% chance of response.
- Use scores comparatively. They are most useful for ranking samples processed in a consistent way.
- The bands are descriptive. They come from pooled score quantiles across 415 available reference samples.
DATA ACCESS
Original dataset access
Use the official repositories below to obtain each source dataset with its original documentation and access conditions. The site does not redistribute combined training or test matrices.
ZENODO
FPMTB
↗GEOExpression, ex vivo drug response and clinical files from the original Zenodo record.
Open accessRJAML Cohorts 1 & 2
↗GITHUBBoth clinical cohorts are provided through the same GEO series, GSE289786.
GSE289786BeatAML2
↗Author-maintained repository with normalized expression, inhibitor response and clinical files.
Open processed dataMODEL RESOURCES
Examples and reproducibility files
The prediction example, blank input template, frozen gene pairs and summary metrics remain available directly from this portal.
METHOD
From expression to response score
The published ComboKR concept is adapted here into a compact cross-domain transcriptomic predictor.
Remove cohort direction
The strongest linear FPMTB-versus-clinical direction is learned once and its component is subtracted from every sample.
Measure similarity
An RBF kernel compares the corrected pair pattern with the 176 pooled training profiles.
Return a score
Weighted similarities are combined into a continuous, uncalibrated response score.
Expression matrix→64 binary pairs→Domain correction→RBF response score
EVIDENCE
Retained model performance
AUROC point estimates from the final retained model. These results are exploratory because evaluation cohorts informed model retention.
Development readoutn = 98
FPMTB
0.916AUROC
Exploratory evaluationn = 32
C1 held-out
0.734AUROC
Selection targetn = 32
Cohort 2
0.700AUROC
Exploratory evaluationn = 253
BeatAML
0.810AUROC
Research use only.This prototype is not a medical device, does not provide a calibrated probability or clinical cutoff, and requires independent prospective validation.
PROJECT
ComboKR-GP Research Portal
Designed to make the frozen research model inspectable and reproducible: inputs remain local, downloadable resources are versioned with the model, and limitations are visible beside the results.