TwistDAN: Twisted Domain Adversarial Network for Synthetic Accessibility Assessment

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Machine Learning

Background:

• Synthetic accessibility (SA) prediction is crucial for guiding molecule synthesis in drug discovery.
• Existing SA predictors lack generalization across diverse chemical domains, limiting their use in virtual screening.
• Accurate SA prediction is essential for efficient hit identification and lead optimization.

Purpose of the Study:

• To develop a novel SA prediction model, TwistDAN (Twisted Domain Adversarial Network), with improved cross-domain generalization.
• To leverage semisupervised learning by combining supervised and adversarial training on labeled and unlabeled molecular data.
• To enhance the utility of SA prediction for virtual screening of large and diverse molecular libraries.

Main Methods:

• Adapted domain adversarial neural networks (DANN) for SA prediction using semisupervised learning.
• Employed supervised learning on 640,000 labeled molecules (Easy-to-Synthesize and Hard-to-Synthesize) and adversarial learning on 2.1 million unlabeled SELFIES variants.
• Utilized identical 2D molecular graph representations, graph attention networks, and gradient reversal layers for domain-invariant learning.

Main Results:

• TwistDAN demonstrated strong cross-domain generalization with an AUROC of 0.951 under severe domain shift.
• Achieved high performance on challenging discrimination tasks (AUROC = 0.938) with structurally similar molecules.
• Exhibited high precision (0.980), reducing false-positive predictions by 12 percentage points compared to leading methods.

Conclusions:

• TwistDAN significantly enhances the generalization capability of synthetic accessibility prediction models.
• The model's high precision minimizes wasted synthesis efforts, accelerating drug discovery pipelines.
• TwistDAN offers interpretable attention-based visualizations for medicinal chemists, aiding in rational drug design.