Towards a General-Purpose Model for the Chemistry Discovery Workflow
Laura Mismetti ⋅ Mariana Alves ⋅ Carlo Baldassari ⋅ Rhyan Barrett ⋅ Sarah de Ruiter ⋅ Serafina Di Gioia ⋅ Mara Graziani ⋅ Flavia Iovane ⋅ Anna Kelmanson ⋅ Rémi Schlama ⋅ Federico Zipoli ⋅ Uros Zivanovic ⋅ Andreas Krause ⋅ Rui Moreira ⋅ Cecilia M Rodrigues ⋅ Philippe Schwaller ⋅ Tiago Rodrigues ⋅ Teodoro Laino ⋅ Jannis Born ⋅ Marvin Alberts
Abstract
We present GraniteChem, a generalist chemistry assistant covering a wide task range across the chemistry discovery cycle including general chemistry reasoning, conditioned molecular generation, forward reaction prediction and retrosynthesis, and spectroscopic structure elucidation from Nuclear Magnetic Resonance (NMR) data. Built on Granite-3.3-8B, GraniteChem is trained on over 85B tokens via supervised finetuning (SFT) followed by group relative policy optimisation (GRPO) with shaped chemistry rewards. GraniteChem matches or outperforms specialist models on their own tasks, while sustaining strong performance across all domains. In an experimental case study, GraniteChem designed two compounds against metabolic dysfunction associated steatotic liver disease under selectivity and synthesizability constraints. One compound exhibited strong dose-dependent anti-necroptotic activity in both mouse ($EC_{50}$ = 11.6 µM) and human ($EC_{50}$ = 12.7 µM) cells, with follow-up biochemical profiling consistent with allosteric RIPK3 modulation.
Chat is not available.
Successful Page Load