The Digital Press

Researchers developed a method that improves molecular property predictions by training a Graph Transformer neural network with both real experimental data and synthetic tasks from XGBoost models. This joint training enhances performance without relying on complex pretraining or direct feature injection.

Why it matters: Combining synthetic tasks with real data boosts neural network accuracy on molecular property predictions.

The big picture: Synthetic task augmentation offers a new way to improve multitask learning without extensive pretraining or feature engineering.

Stunning stat: The multitask Graph Transformer outperforms XGBoost on 16 out of 19 molecular property targets.

Quick takeaway: Integrating XGBoost-derived synthetic targets as auxiliary tasks consistently enhances Graph Transformer performance.