Using Active Learning to Develop Machine Learning Models for Reaction Yield Prediction

Abstract

Computer aided synthesis planning is a rapidly growing field for suggesting synthetic routes for molecules of interest. The methods used are usually dependent on access to large datasets for training, but with a finite experimental budget there are limitations on how much data can be obtained from experiments. Active learning, which has been used in recent studies with success, is a strategy to identify which data points impact model accuracy the most. However, little has been done to explore the robustness of the methods predicting reaction yield. This study aims to investigate the influence of machine learning algorithms and the number of initial data points on reaction yield prediction for two public high-throughput experimentation datasets. Our results show that active learning based on output margin reached a pre-defined accuracy (AUROC) faster than using passive learning. Feature importance analysis of the trained machine learning models suggested active learning had larger influence on the model accuracy when only a few features were important for the model prediction.

Predicting reaction conditions from limited data through active transfer learning

“…The goal of finding desired reaction conditions is qualitatively different from the goals of other active learning studies for classification, which usually aim to minimize error in the domain of interest. 40,47 A similar formulation was used in the context of drug discovery (up to ∼17 500 molecules), where data points were selectively labeled based on the farthest distance from the support vector machine's classification hyperplane. 70 Another study 71 under a similar setting (with ∼100 million molecules) showed greedy selection to be effective at identifying molecules with the best docking scores.…”

“…38,39 Active learning is therefore well-suited for reaction development, which greatly benefits from efficient exploration and where chemists conduct the next batch of reactions based on previous experimental results. Based on this analogy, reaction optimization 27,28 and reaction condition identification 40 have been demonstrated to benefit from active learning. However, these prior works initiate exploration with randomly selected data points (Fig.…”

Active learning-based exploration of the catalytic pyrolysis of plastic waste

“…Active learning is well-suited for modeling reactions with carefully selected experiments, although its use is not widespread yet in chemical engineering [10]. In chemical engineering, the benefits of active learning have been investigated for drug discovery [10][11][12], catalysis [13][14][15], quantum chemical calculations [8,[16][17][18][19][20], polymer design [21], and retrosynthesis [22,23]. However, active learning has not yet been applied in reaction modeling to the best of our knowledge.…”

“…However, active learning has not yet been applied in reaction modeling to the best of our knowledge. To date, active learning has been predominantly used on simulated databases [8,14,17,18,20] or large experimental datasets (greater than1000 datapoints) [22,23]. This data-intensity makes it impossible for utilization in a real experimental campaign where data is limited due to the high cost and long duration of experiments.…”