Can human experts predict solubility better than computers?

Area of Science:

• Computational chemistry
• Drug discovery
• Cheminformatics

Background:

• Accurate prediction of aqueous solubility is crucial for drug development.
• Human expert intuition and machine learning algorithms are both used for such predictions.

Purpose of the Study:

• To compare the predictive performance of human experts against machine learning algorithms for aqueous solubility.
• To evaluate the effectiveness of consensus predictions from both humans and algorithms.

Main Methods:

• A survey was conducted with human experts from pharmaceutical and academic fields to predict aqueous solubility.
• Ten machine learning algorithms, including a multi-layer perceptron, were applied to the same dataset.
• Consensus predictions were generated by taking the median of individual predictions for both groups.

Main Results:

• The best machine learning algorithm (multi-layer perceptron) achieved an RMSE of 0.985 and R² of 0.706.
• The best individual human predictor achieved an RMSE of 0.942 and R² of 0.723.
• Consensus predictions from both humans and algorithms demonstrated excellent predictivity, with statistically insignificant differences between the two consensus groups.

Conclusions:

• Human experts perform essentially equally well as machine learning algorithms in predicting aqueous solubility.
• The 'wisdom of crowds' effect, achieved through median consensus, significantly enhances prediction accuracy for both human and algorithmic approaches.