Meta Learning with Attention Based FP-GNNs for Few-Shot Molecular Property Prediction

This paper introduces a new method, AttFPGNN-MAML, to predict molecular properties using very little data, which is a common problem in drug discovery. The method uses a combination of different ways to represent molecules and a special type of machine learning to make accurate predictions even with limited information. The model is trained and adapted to new tasks using ProtoMAML, a meta-learning strategy.

• 1
  The AttFPGNN-MAML method outperformed other methods on three out of four tasks in the MoleculeNet dataset, demonstrating superior performance in few-shot learning settings.
• 2
  The method achieved the best performance on the FS-Mol dataset at support set sizes of 16, 32, and 64, further validating its effectiveness.
• 3
  Ablation studies showed that removing the Molecular Fingerprint module or Instance Attention module resulted in performance degradation, confirming their positive impact on the model's efficacy.