Data-efficient learning for accurate identification of MAPK1 inhibitors using an active meta-deep learning framework.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Limited experimental data hinders machine learning applications in drug discovery, especially for cancer targets.
• Mitogen-activated protein kinase 1 (MAPK1) inhibitors are promising for cancer therapies but require efficient predictive models.

Purpose of the Study:

• To develop a data-efficient active meta-deep learning framework for predicting MAPK1 inhibitors.
• To address the challenge of limited experimental data in cancer drug discovery.

Main Methods:

• Integrated active learning (AL) with a meta-model combining four deep architectures (CNN, attention, GCN, GNN-attention).
• Trained models on molecular descriptors and graph representations, generating probability-based features for a meta-learner.
• Evaluated AL sampling strategies, with entropy sampling showing competitive performance.

Main Results:

• The framework achieved significant improvements in AUPRC (5.12%) and MCC (5.48%) using only 10% of training data.
• Achieved competitive performance (AUPRC: 0.835, MCC: 0.817) with less data compared to traditional methods.
• Demonstrated generalizability on an external MAPK1 dataset, with AUPRC of 0.818 and MCC of 0.403.

Conclusions:

• The proposed framework effectively accelerates predictive performance in data-scarce drug discovery by combining intelligent data selection with deep learning.
• This approach holds significant potential for advancing cancer-related therapies by enabling efficient identification of novel drug candidates.