Beyond the training set: an intuitive method for detecting distribution shift in model-based optimization

Model-based optimization (MBO) is increasingly applied to design problems in science and engineering. A common scenario involves using a fixed training set to train models, with the goal of designing new samples that outperform those present in the training data. A major challenge in this setting is distribution shift, where the distributions of training and design samples are different. While some shift is expected, as the goal is to create better designs, this change can negatively affect model accuracy and subsequently, design quality. Despite the widespread nature of this problem, addressing it demands deep domain knowledge and artful application. To tackle this issue, we propose a straightforward method for design practitioners that detects distribution shifts. This method trains a binary classifier using knowledge of the unlabeled design distribution to separate the training data from the design data. The classifier's logit scores are then used as a proxy measure of distribution shift. We validate our method in a real-world application by running offline MBO and evaluate the effect of distribution shift on design quality. We find that the intensity of the shift in the design distribution varies based on the number of steps taken by the optimization algorithm, and our simple approach can identify these shifts. This enables users to constrain their search to regions where the model's predictions are reliable, thereby increasing the quality of designs.

Forecasting labels under distribution-shift for machine-guided sequence design

The ability to design and optimize biological sequences with specific functionalities would unlock enormous value in technology and healthcare. In recent years, machine learning-guided sequence design has progressed this goal significantly, though validating designed sequences in the lab or clinic takes many months and substantial labor. It is therefore valuable to assess the likelihood that a designed set contains sequences of the desired quality (which often lies outside the label distribution in our training data) before committing resources to an experiment. Forecasting, a prominent concept in many domains where feedback can be delayed (e.g. elections), has not been used or studied in the context of sequence design. Here we propose a method to guide decision-making that forecasts the performance of high-throughput libraries (e.g. containing $10^5$ unique variants) based on estimates provided by models, providing a posterior for the distribution of labels in the library. We show that our method outperforms baselines that naively use model scores to estimate library performance, which are the only tool available today for this purpose.