A quantitative uncertainty metric controls error in neural network-driven chemical discovery

Area of Science:

• Computational chemistry
• Materials science
• Artificial intelligence in drug discovery

Background:

• Machine learning (ML) models, including artificial neural networks, offer rapid compound characterization, complementing traditional high-throughput screening.
• Accurate identification of out-of-domain predictions is crucial for reliable large-scale chemical space exploration using ML.
• Existing uncertainty metrics for neural networks are often computationally expensive or require complex feature engineering.

Purpose of the Study:

• To introduce a novel, low-cost uncertainty metric for neural network ML models.
• To enable reliable chemical space exploration by quantifying model applicability domain.
• To provide a method for predictive error control in chemical discovery.

Main Methods:

• Developed a new uncertainty metric based on the distance to available data in the latent space of a neural network.
• Applied the metric to both inorganic and organic chemistry datasets.
• Compared the performance of the new metric against established uncertainty quantification methods.

Main Results:

• The proposed latent distance metric is a low-cost and quantitative measure of uncertainty.
• This approach demonstrates calibrated performance exceeding widely used uncertainty metrics.
• The metric is applicable to models of increasing complexity without additional computational cost.

Conclusions:

• The distance to available data in latent space serves as an effective uncertainty metric for ML models in chemistry.
• This method facilitates predictive error control and aids in identifying valuable data points for active learning strategies.
• The approach enhances the reliability and scalability of ML-driven chemical discovery.