Diabetome logoDiabetomeRun Prediction

One SMILES. 305 endpoints.

Multitask D-MPNN across 205 diabetes targets — kinases, GPCRs, enzymes, nuclear receptors — with conformal intervals and applicability-domain scores.

Draw structure
3D structure
DPP-4 · P27487
205
UniProt targets
< 1 s
Inference time
91.6%
Conformal coverage
0.58
Median scaffold R²
how it works

One molecule in. 305 predictions out.

01

Draw or import your molecule

Paste a SMILES string, draw interactively with JSME, or import SDF/MOL files. Any drug-like molecule is accepted.

step 01
02

Encode as ECFP4 fingerprints

Extended-connectivity fingerprints at radius 4 (2048 bits) capture substructural patterns. Counts preserve frequency — no information loss.

ECFP4 · 2048 bits
67 active bits of 2048
step 02
03

Score across 305 endpoints in one pass

A single multitask D-MPNN predicts all 305 regression endpoints simultaneously — scaffold-split BindingDB training with conformal calibration.

pIC50 predictions · DPP-4
XGBoost
12.6
LightGBM
12.1
RF
11.8
Ridge
11.1
ChemProp
12.9
Ensemble
13.2
step 03
04

Ranked with conformal intervals

90% conformal prediction intervals and Tanimoto-based applicability domain scores quantify reliability. Sort by activity, AD score or CI width.

Ranked predictions · 305 endpoints
Dipeptidyl peptidase 4
8.41In-domain
Glycogen synthase kinase-3β
7.85In-domain
SGLT2
7.62In-domain
PPARγ
6.93Borderline
Insulin receptor
6.44Borderline
step 04
Validation

Rigorously tested.

0.759
Temporal R²
93 tasks · 17k future compounds
38%
FDR vs Ridge
tasks sig. better (BH q=0.05)
0.89
Median clf AUC
197 classification tasks
6
Published benchmarks
clinical models compared
start predicting

Paste a SMILES. Get ranked, calibrated predictions.

205 diabetes targets · 305 endpoints · conformal intervals · applicability-domain scores.

Open molecule editor Browse targets