Guangdong Innovation Center For Small Molecule Drug Discovery and NewDEL Biotech recently collaborated on publishing a paper titled "3CLpro inhibitors: DEL-based molecular generation" in Frontiers in Pharmacology (Impact Factor: 5.98). This paper utilizes DNA-encoded library (DEL) technology to screen for active compounds targeting 3CLpro, a proteinase essential for viral replication. Subsequently, based on the DEL screening data, the paper establishes pre-trained molecular generation models such as RNN, VAE, AAE, MCMG, etc. These models are fine-tuned using molecules with high affinity from DEL to generate a series of novel compound structures, achieving both drug discovery and compound structure optimization based on DEL. It is noted that this is the first paper internationally to achieve structure optimization through molecular generation using DEL data.

Research Background

One fundamental problem in drug structure optimization is designing new molecules with improved specific properties based on known information. In the early stages of drug development, the structural modification of new molecules often relies on the accumulated experience and knowledge of medicinal chemists, which greatly limits the explored chemical space. The application of deep learning in drug design was recognized as one of the top ten breakthrough technologies by MIT Technology Review in 2020. Utilizing this tool to its fullest potential can explore a broader chemical space and generate new molecules with desired chemical and pharmacological properties.

Molecular generation through machine learning (ML) has proven to be an effective tool for drug structure optimization. For example, Zhavoronkov et al. (2019) discovered a kinase inhibitor for DDR1 within 21 days using the deep learning molecular generation architecture GENTRL. However, this is mainly due to DDR1 being a target with a large amount of reported activity data. For new targets, molecular generation for structure optimization often yields poor results due to the lack of extensive activity data.

Rapidly generating a large amount of affinity data for specific targets, including data for new targets, through DEL technology, and then utilizing this data for molecular generation may help address the problem of rapid optimization of lead compounds for new targets.

Methodology

First, the paper constructs two DEL libraries and screens these libraries against the 3CLpro target, resulting in 1.04 billion affinity data points for 3CLpro and the discovery of some active compounds.

Subsequently, molecular generation models were established based on DEL data to generate potentially active molecules. Due to the vast amount of affinity data generated by DEL, low-count data were first truncated, and the remaining 3.7 million data points were used to construct pre-trained models such as RNN, VAE, AAE, etc. Finally, transfer learning was applied to the 18,129 high-count molecules to generate molecules with high affinity, thereby increasing the probability of obtaining active molecules. The paper compared the molecules generated with and without transfer learning and set up a series of filtering criteria (e.g., drug-likeness, toxicity groups, pharmacophores, docking). Additionally, it combined the experience and knowledge of medicinal chemists to select potentially active molecules. Currently, further research is being conducted on these potential active molecules.

Results:

Performance of various models: The MCMG model performed well in Novelty, while the other three models performed poorly in Novelty but well in Validity and Uniqueness. VAE, AAE, and RNN all utilized relatively simple architectures, while the MCMG model referenced research results from Wang et al. (2021), indicating that using more complex models may be necessary to improve the novelty of generated molecules. However, molecules generated by the MCMG model performed poorly in affinity testing, and the reasons for this still require further consideration.

Affinity distribution analysis: Molecules generated by models using transfer learning achieved higher scores in affinity testing. This indicates that when lacking active data, using molecules with high counts for model fine-tuning during molecular generation with DEL data is more likely to produce active molecules. This paper combined affinity score prediction, pharmacophore filtering, evaluation of the rationality of ligand-target complex formation, and the experience and knowledge of medicinal chemists to select potentially active compounds. Some of these compounds were synthesized and tested for biological activity.

Summary:

According to the authors, this study represents the first application of DEL dataset in molecular generation research, which will advance the development of combining DEL with AI. The study preliminarily found that using DEL data for molecular generation is feasible, especially for new targets. Affinity models and molecular generation models based on DEL data are expected to become a dual-function tool for drug discovery and structure optimization. Due to the scarcity of datasets for new targets, relying on public databases makes it difficult to achieve this goal.

Related Literature:

Xiong F, Xu H G, Yu M G, et al. (2022). "3CLpro inhibitors: DEL-based molecular generation." Frontiers in Pharmacology.

Zhavoronkov A, Ivanenkov Y A, Aliper A, et al. (2019). "Deep learning enables rapid identification of potent DDR1 kinase inhibitors." Nature biotechnology, 37(9), 1038-1040.

Wang Jike, Hsieh C, Wang M Y, et al. (2021). "Multi-constraint molecular generation based on conditional transformer, knowledge distillation and reinforcement learning." Nature Machine Intelligence, 3(10), 914-922.