README.md

metadata

title: SugarKi
emoji: 🧪
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: true
license: apache-2.0
short_description: Ki prediction for sugar-chemistry enzymes

SugarKi 🧪

A sugar-chemistry-specialized Ki predictor for enzymes (polyol DHs, sugar kinases, glycosidases, sugar phosphatases, aldolases, isomerases, phosphomutases). State-of-the-art Ki models like CatPred fail catastrophically on monosaccharide and polyol inhibitors (R² = −0.95 to −1.5); SugarKi specifically targets that failure mode.

Headline numbers

	Sugar-chemistry Ki	General Ki
CatPred zero-shot	R² = 0.243	R² = 0.578
SELFprot zero-shot	R² = 0.623	R² = 0.314
SugarKi specialist	R² = 0.702	(router fallback to CatPred)

Validation

On MDH-006 WT (mannitol DH) + D-mannitol: predicted Ki = 11.78 mM in substrate mode vs literature 12 mM. 1.8% relative error.

Setup

This Space is the public UI. It calls the private SugarKi backend at Umesh1608/sugarki-backend. The backend hosts the model weights, ESMFold, P2Rank, and the Plan E2 inference pipeline on a ZeroGPU A100.

Required Space secrets

Secret	Purpose
HF_TOKEN	Read access to Umesh1608/sugarki-backend
SUGARKI_BACKEND	(optional) Override default backend Space ID

Architecture

[ user ] → [ MWBC/sugarki frontend (this Space, CPU Basic) ]
              ↓ gradio_client.Client(...)
        [ Umesh1608/sugarki-backend (ZeroGPU A100) ]
              ↓
        ESMFold → P2Rank → GVP-GNN → Plan E2 → log10(Ki/mM)

Citation

Paper in preparation.

License

Apache-2.0.