Chemical Property Prediction Under Experimental Biases

Predicting the chemical properties of compounds is crucial in discovering novel materials and drugs with specific desired characteristics. Recent significant advances in machine learning technologies have enabled automatic predictive modeling from past experimental data reported in the literature. However, these datasets are often biased because of various reasons, such as experimental plans and publication decisions, and the prediction models trained using such biased datasets often suffer from over-fitting to the biased distributions and perform poorly on subsequent uses. Hence, this study focused on mitigating bias in the experimental datasets. We adopted two techniques from causal inference and domain adaptation combined with graph neural networks that can represent molecular structures. The experimental results in four possible bias scenarios indicated that the inverse propensity scoring-based method made solid improvements, but the domain-invariant representation learning approach failed.