reading notes of《Artificial Intelligence in Drug Design》
文章目录
1.Introduction
- There are few reports of successful combination of the two aspects of design and synthesis of novel molecules.
2.Computional Molecular Generation
- These publications provide training and testing datasets, and a set of metrics to evaluate the quality and diversity of generated structures. The platforms and source codes are available via guacamol and moses.
3.Retrosynthetic Planning and Synthetic Feasibility Assessment
- We applied some of the methods to assess synthesizability of compounds generated by BREED, one of the earlier de novo molecular design methods that generated drug-like molecules based on known inhibitors that bind to the same target. In this publication, four known HIV-1 protease inhibitors were used and produced 9 novel hybrids。
- Early in 2009, Ertl and Schuffenhauer proposed synthetic accessibility score, SAscore for drug-like molecules based on a combination of fragment contributions and a complexity penalty. SAscore ranges between 1 (easy to make) and 10 (difficult to make), which has been validated by the good agreement between calculated and chemist-estimated synthetic accessibility for a set of 40 molecules.
- Instead of considering molecule complexity, Coley at el developed the synthetic complexity score (SCscore) based on the premise that the products of chemical reactions should be more synthetically complex than their corresponding reactants.
- Thakkar et al. published the retrosynthetic accessibility score (RAscore) for rapid estimation of synthetic feasibility as deter- mined from a retrosynthesis planning tool AiZynthFinder.
4.Combination of Synthetic Feasibility and Deep Generative Algorithms
- The penalized logP (logP-SAscore-the number of long cycles) is used as the target objective to optimize the generated molecules in multiple deep generative models, such as grammar variational autoencoder (GVAE), junction tree variational autoencoder (JT-VAE), molecule deep Q-Networks (MolDQN), and Mol-CycleGAN
- However, these synthetic feasibility heuristics scores cannot always give an accurate estimate of the synthesizability for de novo generated molecules. In other words, the high synthetic feasibility score cannot guarantee the ease of synthesis for certain compounds. Additionally, using synthetic feasibility score to guide the generation detracts the molecule optimization from the pri- mary objectives.
- The second approach, which embeds the synthesizability into molecule generation algorithms, has shown potential to overcome these limitations in more recent studies. Horwood and Noutahi developed a Reaction-driven Objective Reinforcement method (REACTOR) that uses a sequence of chemical reactions as state transitions in Markov decision processes to generate molecules within a reinforcement learning framework.
- Gottipati and Sattarov et al. developed a similar reinforcement learning empowered forward synthesis framework called Policy Gradient for Forward Synthesis (PGFS) for de novo molecule design.
- Bradshaw et al. proposed a novel VAE generative model called MoleculeChef.
- Korovina et al. developed a Bayesian optimization framework, ChemBO, which performs a random walk on a synthesis graph iteratively with selected reactants and conditions using the Rexgen reaction outcome predictor to generate final molecules that fit user-defined objectives.
5.Conclusions
- it is not straightforward to compare results obtained by different methods. There are attempts to create standard datasets of compounds (with known synthesis) and metrices for making comparisons with respect to cost, time and favorable reactions. However, these tools are too complex for incorporation into the iterative design and optimization of novel molecules.