Genentech's Aspen Trains a Small Open Model to Beat Frontier Chemistry AI
Genentech's Prescient Design team built RL environments for drug design tasks, then post-trained a 30B model that matches GPT-5 and Claude Opus.
Takeaways
- Prescient Design at Genentech built a chemistry benchmark that doubles as RL training environments for drug design.
- The paper evaluates GPT-5, Claude Opus 4, and Qwen3-30B across property prediction, translation, and generation tasks.
- Aspen, a post-trained Qwen3-30B-A3B model using DAPO-GRPO, matches frontier closed models on many tasks.
- Training used 256 B200 GPUs, ~900k prompts, and roughly 20 days of wall-clock compute.
- Aspen beats frontier baselines on multiproperty constrained generation and ligand efficiency in 8TTR lead optimization.
- RL sharpens existing skills but fails on SMILES-to-IUPAC and DMPK solubility where base knowledge is absent.
A team from Prescient Design at Genentech has released one of the most thorough looks yet at how well frontier language models can actually do chemistry, along with a recipe for making smaller open models competitive with the closed frontier. Their paper introduces a suite of chemically-grounded tasks framed as reinforcement learning environments, evaluates three model families (GPT-5, Claude Opus 4, and Qwen3-30B-A3B), and post-trains a model called Aspen that closes most of the gap to closed frontier systems on realistic drug design workloads.
The core problem the authors flag is familiar to anyone trying to plug LLMs into a real discovery pipeline. Frontier agents get deployed for target identification, lead optimization, and toxicity prediction, but their ceiling is set by the base model, and current chemistry benchmarks tend toward general knowledge trivia or narrow multiple-choice formats that do not resemble what a medicinal chemist actually does. That leaves teams guessing whether the next model generation will finally handle their workflows, or whether they should invest in specialized training.
Turning the benchmark into a training signal
Every task is written as an RL environment with a computable reward, so the same suite works for evaluation and post-training. The task groups span the practical surface area of small-molecule work:
- RDKit property prediction: Given a SMILES, predict properties like molecular weight, TPSA, LogP, QED, H-bond donor and acceptor counts, or Bertz complexity.
- Experimental prediction: Given 5 to 15 in-context molecules with measured values (proprietary Genentech potency and DMPK data, plus FS-Mol), predict the property of a held-out molecule from the same HDBSCAN-clustered substructure family.
- Multiple choice: SAR ranking questions and a SMILES-equivalence task where distractors are generated by perturbing SELFIES tokens, so they share the query's molecular formula and force reasoning about connectivity.