GOLLuM trains a language model to select the next experiment using uncertainty estimates

GOLLuM learns from experimental results and selects the next test using uncertainty estimates

On August 28, Nature Machine Intelligence published a paper on GOLLuM, a method for selecting the next experiment. It reads textual descriptions of candidate conditions, learns from existing results, and works with a probabilistic model to choose what to test next.

Laboratories spend time and reagents selecting the next experiment from thousands of valid candidates. In chemistry and materials science, the number of possible reactions, molecules, and compositions grows rapidly.

“From reaction optimization to molecular design, experimental discovery poses the same costly question: which candidate should be tested next when time and resources are limited?”

Bayesian optimization selects the next experiment using a prediction and an estimate of how uncertain that prediction is. It usually requires predefined numerical features that describe molecules, reactions, or materials. GOLLuM instead receives a textual description of the conditions, which a language model converts into a numerical representation. A Gaussian process, which is a statistical model, uses this representation to predict the result and estimate its uncertainty. The selection algorithm uses both values.

Results from completed experiments modify this representation, bringing descriptions associated with similar results closer together. In the Buchwald–Hartwig reaction dataset, the model grouped conditions involving iodine-containing reagents in a region associated with high yields, meaning large proportions of the target product, while placing conditions involving chlorine-containing reagents in a region associated with low yields.

The authors also asked several language models to propose reaction conditions directly. The models invented chemical structures, suggested candidates outside the defined search space, repeated conditions that had already been tested, or returned answers that the software could not parse. In GOLLuM, language represents the candidates, while a probabilistic algorithm selects the next test using the expected result and its estimated uncertainty.

The paper compared the methods across 23 tasks in organic, analytical, and process chemistry, materials science, and molecular design. For each task, the search began with ten conditions whose results were below the median. With a budget of 50 experiments, GOLLuM identified an average of 36.3% of the conditions in the top 5% of results. Conventional Bayesian optimization using predefined features identified 29.7%. In the Buchwald–Hartwig reaction benchmark, the proportion of identified conditions from the top 5% increased from 24% with a language model that had not been fine-tuned to 43% after joint training with the Gaussian process.

The authors released the code, data, and benchmark configurations so that these results can be reproduced.