AI pharmaceuticals are maturing, and data sharing needs to be broken

From the target discovery to the nomination of pre-clinical candidate compounds, InSI Intelligence only spent millions of dollars in research and development funds, which took 18 months, which attracted the attention of the industry. In March 2022, the drug was approved to enter phase I clinical trials in China, which is also the first AI drug to enter clinical trials in China. In January 2023, InSI released positive top-line data from its Phase I clinical trial. In February of the same year, InsiliconIntelligence announced that the drug candidate received orphan drug certification from the US Food and Drug Administration (FDA) and was about to start Phase II clinical trials, injecting more imagination into the prospect of AI pharmaceuticals.

A number of Zhangjiang enterprises interviewed by Surging technology said that AI pharmaceutical is not born out of nowhere, but the extension and evolution of traditional computer-aided Drug Design (CADD).

"With the improvement of deep learning technology, the concept of AI pharmaceutical has been put forward more and more specifically, and it turns out that AI is already used in many places." Chen Chunlin, rotating chairman of Zhangjiang AI New Drug Development Alliance and founder and CEO of Medisi, introduced that from the early random screening of candidate drug molecules, to the prediction of physical and chemical properties of small molecule compounds through computer simulation technology, and then to the rise of structural biology, based on the exploration of the interaction mechanism and binding site characteristics between drugs and proteins, The drug discovery and optimization process is beginning to become more predictable and controllable.

"But these are just changes above the point. Human experience can only be limited to one domain, for example, if a person is very good at synthesizing a particular sequence of compounds, another sequence of experience may be inaccurate." Chen Chunlin said that American chemist Christopher Lipinsk had proposed Five basic laws of drug design (" Rule of Five "), "now AI can consider these various factors, which is the biggest breakthrough brought by AI." Because there are more databases, although it's not the most comprehensive."

Fan Mengqi, vice president of the biomedical business group of Deep Trend Technology, said that AI can accelerate molecular virtual screening, help screen a larger chemical space in a shorter period of time, and help the discovery of seedling compounds. At the same time, the free energy perturbation (FEP) algorithm can further accelerate the lead compound optimization stage. "For example, under the traditional paradigm, the optimization stage may have to design and synthesize hundreds of molecules, waiting for several months of synthesis and biological experiment time, and now the designed molecules can first calculate the affinity with FEP, eliminate the molecules with insufficient activity, and finally only need to carry out biological experiments with the first synthetic activity, which can save a lot of time and cost."

Redefining "Structure is King"

In December 2018, AlphaFold, a protein three-dimensional structure prediction platform built by DeepMind (a British artificial intelligence company that was acquired by Google in 2014), showed amazing strength in the CASP13 competition (an international protein structure prediction competition), matching the structural accuracy of laboratory analysis. Less than two years later, AlphaFold2 set another record in the field. In July 2021, DeepMind unveiled a database of millions of predicted protein structures for use by scientists and researchers, marking an important milestone in AI for Science. At the end of July 2022, DeepMind again announced that AlphaFold DB had expanded from 1 million structures to more than 200 million structures, expanding by more than 200 times.

Zhong Wenge, chief scientific officer of Rueger Medicine, pointed out that the prediction accuracy of AI tools has a lot to do with known protein structures. The dynamic properties of protein conformation are particularly important for drug molecular design. "Changes in the binding pocket are key, but the more and more dynamic the protein, the more uncertain the prediction of the AI tool." Small deviations in specific amino acid residues, or even in combination with whether or not there are water molecules in the pocket, may be good enough for AI structural prediction, but for the drug hunter, it is the difference between whether the drug is active or not."

Fan Mengqi also believes that the structural accuracy of the current AI prediction is still a certain distance from the actual requirements of doing medicine, especially near the pocket, and the accuracy of the prediction is higher, so it is also necessary to combine dynamic simulation and other means to further refine it.

In May 2022, Deeppotential Technology released Uni-Mol, a universal molecular representation learning framework based on the three-dimensional structure of molecules, combining 200 million small molecule data and 3 million protein pocket data for common training. Fan Mengqi said that on the basis of this pre-trained model, only a small amount of data needs to be fine-tuned to achieve a prediction accuracy far beyond the previous model, and Uni-Mol has won the first place in 14 of 15 public test sets, and the molecular characterization ability is strong.