Chemical property prediction under experimental biases

Area of Science:

• Computational chemistry
• Machine learning
• Data science

Background:

• Predicting chemical properties is vital for drug and material discovery.
• Machine learning models trained on literature data often exhibit bias, leading to poor performance.
• Dataset bias arises from experimental design and publication practices.

Purpose of the Study:

• To develop methods for mitigating bias in experimental chemical datasets.
• To improve the reliability and generalizability of predictive models for chemical properties.
• To address the over-fitting issues caused by biased data distributions.

Main Methods:

• Utilized graph neural networks to represent molecular structures.
• Applied two causal inference techniques: inverse propensity scoring and counter-factual regression.
• Evaluated methods across four distinct bias scenarios.

Main Results:

• Both inverse propensity scoring and counter-factual regression demonstrated significant improvements in prediction accuracy.
• The proposed methods effectively mitigated the negative impacts of dataset bias.
• Enhanced model robustness and performance in diverse bias conditions.

Conclusions:

• Causal inference techniques combined with graph neural networks offer a powerful solution for bias mitigation in chemical data.
• These approaches enhance the utility of machine learning for reliable chemical property prediction.
• The findings pave the way for more accurate and dependable computational discovery processes.