Small data, big challenges: Machine- and deep-learning strategies for data-limited drug discovery

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• The drug discovery and development (DDD) pipeline is significantly hampered by a scarcity of high-quality experimental data, a common issue due to high costs, time, and confidentiality.
• Standard Machine Learning (ML) and Deep Learning (DL) algorithms struggle with small datasets, leading to challenges in feature extraction and model generalization.
• Existing reviews on AI in drug discovery often overlook the specific challenges posed by limited data across the entire DDD pipeline.

Purpose of the Study:

• To address the critical challenge of limited data in the drug discovery and development (DDD) pipeline.
• To survey key drug discovery tasks where data scarcity is prevalent.
• To synthesize both traditional ML and advanced DL strategies specifically adapted for small datasets in DDD.

Main Methods:

• Reviewing existing literature on AI applications in drug discovery, focusing on the small data problem.
• Analyzing the limitations of traditional ML and DL models when applied to small datasets.
• Identifying and categorizing ML and DL methods tailored to overcome data scarcity in various DDD tasks.

Main Results:

• Limited data is a pervasive issue across multiple stages of the drug discovery and development pipeline.
• Traditional ML methods are suitable for small data but lack representational capacity, while DL methods risk overfitting.
• Adapted DL strategies and extended traditional ML approaches show promise for handling small datasets in DDD.

Conclusions:

• Addressing the small data problem is crucial for advancing AI in drug discovery and development.
• Integrating task-specific applications with methodological innovations is key to developing robust and generalizable AI models.
• Future research should focus on creating interpretable and trustworthy AI solutions that effectively leverage limited data in DDD.