Synthetic accessibility (SA) prediction guides which computationally designed molecules warrant experimental synthesis during early-stage hit identification and lead optimization. Current SA predictors achieve high accuracy on training data but fail to generalize across chemical domains, undermining their utility for the virtual screening of diverse molecular libraries. We introduce TwistDAN (Twisted Domain Adversarial Network), adapting domain adversarial neural networks (DANN) for SA prediction through semisupervised learning. We combine supervised learning on 640 k labeled molecules (Easy-to-Synthesize (ES, ≤10 steps) and Hard-to-Synthesize (HS, >10 steps)) with adversarial learning on 2.1 M unlabeled SELFIES-generated variants. Both domains use identical 2D molecular graph representations, forcing the model to learn from structural patterns rather than superficial features. Graph attention networks identify synthesizability-critical substructures, while gradient reversal layers ensure domain-invariant representations. TwistDAN achieves strong cross-domain generalization: AUROC = 0.951 under severe domain shift and 0.938 on challenging discrimination tasks with structurally similar molecules. High precision (0.980) reduces false-positive predictions by 12 percentage points versus leading methods, decreasing unnecessary synthesis attempts. TwistDAN is freely available at https://twistdan.denglab.org with interpretable attention-based visualizations for medicinal chemistry applications.
ALJANABI et al. (Thu,) studied this question.