Abstract
Machine learning sits at the centre of how new science gets done, and its success depends on how scientific objects are represented, training data are sourced and curated, and models are evaluated. Chemical reaction prediction brings all three together. Across five single-step tasks, models predict products, starting materials for a target, reaction conditions, yield, and reaction type. These tools address an important synthesis bottleneck in drug discovery. Despite more than fifty years of work, the literature remains fragmented, and prior surveys typically treat individual tasks or the field's history separately. We survey representative work through a unified, cross-task lens, tracing the field from Corey's rule-based system of 1969 to foundation models trained on broad chemistry data, and comparing model design, reaction representation, data quality, and evaluation across tasks. Three patterns emerge. Hand-coded reaction templates never disappeared; modern models often learn or reuse their structure. Self-supervised pretraining, where models first learn from broad collections without task-specific labels, appears across all five task literatures and is common in recent neural methods, although simpler approaches remain competitive in some settings. Yield models that perform well on narrow experimental datasets often lose reliability on diverse laboratory and literature data, making dataset diversity a central constraint for yield prediction. Harder tests suggest similar risks for the other four tasks, but the evidence remains incomplete. Condition prediction is also among the least publicly benchmarked tasks. We close with priorities for broader datasets that include failed reactions, adaptation to unfamiliar chemistry, and reliable confidence estimates across linked predictions.
