Quantized local LLMs generate molecules well, but only some follow a reference kinase inhibitor set
- Four open-source LLMs were run locally at original precision and three weight-quantization levels, testing unconstrained small-molecule generation and generation conditioned on in-context reference kinase inhibitors.
- Quantization raised inference speed, most for two of the models, while unconstrained output quality (validity, uniqueness, drug-likeness, diversity, synthesizability) was unaffected.
- In reference-guided generation, only one model stayed close to the reference chemical space at every precision; two largely ignored it, and another lost adherence at the most aggressive quantization.
💡 Points Based on the abstract
Unconstrained metrics like validity and QED are nearly saturated for any decent LLM, so "quantization doesn't hurt" there is a weak test. The informative result is adherence, and it needs a real control: compare against a trivial baseline (random draws from the reference set, or a plain scaffold-decoration script) using Tanimoto/scaffold overlap and novelty. Before swapping a quantized model into a kinase design loop, I'd rerun this on our own series at our target precision.