Apply for a GPU community grant: Academic project
Hello Hugging Face team,
I am Md Delower Hossain, a doctoral researcher and PhD candidate in AI at the University of Alabama at Birmingham (UAB), working with the SPARC team on computational drug discovery. I am applying for a ZeroGPU community grant for this Space, Delower/dpp4-molgemma-generator.
Description of the app. This Space is an interactive molecule generator for DPP-4 (dipeptidyl peptidase-4) drug discovery. It serves our fine-tuned MolGemma model — a compact SMILES language model fine-tuned on 9,338 curated DPP-4 potent/moderate compounds plus the 12 FDA-approved gliptins (best validation loss 0.1800). Visitors pick a sample count (100/250/500/1000), a sampling temperature (0.8, 1.0, or both), and optional advanced generation arguments, and the app generates candidate molecules live, then validates every structure with RDKit, checks novelty against the fine-tuning set, and reports a predicted pIC50 / IC50 and potency class for each molecule from our published DPP-4 LTN model. It is intended for researchers, students, and the open drug-discovery community to explore the model's generations without installing anything. A companion DPP-4 Kd (binding affinity) predictor app in the same open series is in preparation.
Justification. The project is fully open and non-commercial: the model repo (Delower/dpp4-molgemma-finetuned) already publishes the full fine-tuned model, tokenizer, config, generation script and RDKit validation pipeline, and this Space will likewise be public with its complete app code. My account's two free ZeroGPU Spaces are already used by the earlier apps in this series — Delower/dpp4-pic50-webapp (DPP-4 pIC50 predictor) and Delower/egfr-pic50-webapp (EGFR pIC50 predictor), both live on ZeroGPU — and creating a third Gradio Space now requires a paid plan, which is why I am applying through the community grant programme. ZeroGPU is the right fit technically: the generator is a small (30.5M-parameter) Keras model, well within ZeroGPU limits, but it samples autoregressively token by token, so a 100-molecule batch takes roughly 7 s per molecule (~12 minutes) on CPU, which makes interactive use impractical; on GPU the same batch is a short, bursty call — exactly the usage pattern ZeroGPU is designed for, and each visitor's usage counts against their own ZeroGPU quota.
The Space currently exists as a static placeholder; the complete Gradio app is built and tested locally, and I will push it (flipping the SDK to Gradio) as soon as the grant is attached. I am happy to provide any further context, and I will credit Hugging Face ZeroGPU in the Space and in the project documentation. Thank you very much for supporting open research demos.
— Md Delower Hossain (Delower), UAB SPARC