NVIDIA's November 16 science announcement packages Nemotron 3.5 Super VL, chemistry data and post-training recipes. Its metabolism example converts experimental parent-molecule/enzyme-to-metabolite pairs into prompts and references.
The technical recipe compares predicted and reference molecular fingerprints using Tanimoto similarity, a zero-to-one overlap score, then trains against that reward with GRPO and LoRA. The reward measures resemblance in that representation; it does not prove chemical validity or experimental effectiveness. Both accounts are NVIDIA's, independent replication is absent, and public access to the linked example recipe is unverified.