Trends in Deep Learning for Property-driven Drug Design.

Area of Science:

• Computational chemistry and cheminformatics
• Systems biology and pharmacology
• Artificial intelligence in drug discovery

Background:

• Drug discovery faces challenges in time and cost reduction.
• Deep learning (DL) and high-throughput screening advance virtual drug screening.
• Deep generative models show promise for exploring chemical space and expediting drug discovery.

Purpose of the Study:

• To bridge drug discovery with systems biology for designing deep generative models.
• To focus on the interface of predictive and generative modeling in drug discovery.
• To quantify trends and discuss progress in generative models for drug discovery.

Main Methods:

• Systematic publication keyword search on PubMed and preprint servers (arXiv, biorXiv, chemRxiv, medRxiv).
• Analysis of trends in molecular representations and generative model architectures.
• Discussion of deep learning for toxicity, drug-target affinity, and drug sensitivity prediction.

Main Results:

• Molecular graphs and Variational Autoencoders (VAEs) are the most adopted representations and architectures for generative models.
• Progress in deep learning for predicting toxicity, drug-target affinity, and drug sensitivity.
• Focus on conditional molecular generative models incorporating multimodal prediction.

Conclusions:

• Future prospects include integrating DL into experimental workflows and adopting federated learning.
• Key challenges involve interpretability, evaluation metrics, and community-accepted benchmarks.
• Bridging drug discovery with systems biology is essential for next-generation generative models.