AI is being actively used for drug discovery.
The problems with research and development of small molecule drugs are that it takes a long time and cost, and the success rate is low. The drug R & D process begins with exploratory research to identify the target molecule responsible for the disease and the compound that is the "seed" of the drug, followed by compound optimization and nonclinical testing in animals. Finally, human clinical trials are conducted. By accurately and quickly determining the target molecule and identifying the compound, the duration of subsequent processes will be shortened, and the success rate will be improved. To determine the best combination of a target molecule and a substance that acts on the target, including modality, it is not efficient to carry out exploratory research step by step. The best solution is to prepare a compound that acts on a candidate target molecule before the target molecule is determined, instead of identifying the compound by high-throughput screening after determining the target molecule as in the past. If there is no compound that acts on the target molecule, we can proceed with research and development with a different modality.
To realize this idea, it is necessary to search for compounds that act on every molecule that exists in humans.
However, it is difficult to do that by relying solely on wet experiments. AI is used for that purpose. Machine learning will be performed using existing data, and compounds that act on target molecules will be predicted from public compounds that exist all over the world. However, the problem here is the quality of the data used for machine learning. Usually, the existing data is measured using various biochemical assays, and even if the target molecule is the same, if the measurement method is different, the data will be divergent. Therefore, I would like to propose an evaluation using the binding activity as an index, which enables unified evaluation for all target molecules. All activities can be shown in the same unit and Kd value without being biased as in biochemical assays.
This is achieved by screening using a technique called affinity selective mass spectrometry (ASMS). With this method, it is only necessary to prepare the target molecule, and it is not necessary to build an assay system for each target molecule, and it is possible to perform screening by mixing different target molecules. It is possible to screen dozens of target molecules at the same time, which is the best way to feed machine learning that requires large amounts of data. Proteins are expressed from approximately 20,000 cDNA library and are sequentially screened using ASMS. The compound library used is hundreds of thousands carefully selected. Machine learning will be performed based on the obtained data to optimize it. After that, screening from more than 10 million public compounds will be carried out, selection of different binding compounds and compound search will be carried out for target molecules without hit compounds by AI. As a result, a database for target molecules existing in humans will be completed, and a significant improvement in the drug discovery process can be expected.
2021 is the starting year of this strategy and we will contribute to AI drug discovery through ASMS screening for approximately 20,000 proteins.