Synthetic Applicability Domain (SynAD): Navigating Chemical Space for Reliable AI-Driven Reaction Prediction.

Area of Science:

• Organic synthetic chemistry
• Artificial intelligence
• Machine learning

Background:

• Artificial intelligence (AI) is transforming organic synthesis by accelerating hypothesis evaluation and reducing experimental efforts.
• A key limitation is the out-of-distribution (OOD) issue, where AI models fail to predict unseen reactions with novel catalysts, substrates, or conditions.
• Accurate prediction of reaction outcomes is crucial for efficient drug discovery and materials science.

Purpose of the Study:

• To introduce SynAD, a machine learning framework for assessing the predictive capability of AI models in organic synthesis.
• To automatically demarcate reliable and unreliable reactions within the chemical space covered by AI models.
• To provide chemists with a tool to trust AI predictions and strategically allocate resources for de novo discovery.

Main Methods:

• SynAD combines chemical descriptors with model-adaptive distance metrics to define the synthetic applicability domain.
• The framework was validated on the Ullmann Ligand Dataset (ULD) and six additional reaction datasets.
• A SynAD score was developed to quantify the predictability of reaction classes.

Main Results:

• SynAD successfully distinguished predictable chemical space, achieving a prediction accuracy of R² = 0.90 (at 12.3% coverage) on the ULD, a significant improvement from a baseline of R² = -0.21.
• The framework demonstrated consistent performance across multiple datasets, confirming its generalizability.
• The SynAD score effectively quantifies reaction class predictability, guiding experimental focus.

Conclusions:

• SynAD establishes a critical framework for defining the limits of AI predictive capabilities in organic synthesis.
• By identifying reliable reaction spaces, SynAD empowers chemists to trust AI tools and optimize experimental resource allocation.
• This approach accelerates the pace of de novo molecular discovery by providing essential guardrails for AI-driven research.