[4Q3-IS-2d] Machine learning

[4Q3-IS-2d-05] A machine learning model for predicting quantum chemistry based protein-drug molecule interactions

〇Ryosuke Kita¹, Chiduru Watanabe², Masateru Ohta², Naoki Tanimura³, Koji Okuwaki⁴, Mitsunori Ikeguchi^2,5, Kaori Fukuzawa⁶, Teruki Honma², Tsuyohiko Fujigaya¹, Koichiro Kato¹ (1. Department of Applied Chemistry, Kyushu University, 2. RIKEN Center , 3. Mizuho Research & Technologies, Ltd, 4. JSOL Corporaion, 5. Yokohama City University, 6. Department of Pharmacy, Osaka University)

Keywords:Computational Chemistry, AI Drug Discovery, Fragment Molecular Orbital Method

The evaluation of protein-drug molecule interactions through molecular simulation plays a critical role in identifying drug candidates from a vast pool of chemical compounds in computational drug discovery.
To increase the hit rate, which has been a challenge with traditional methods, accurate quantum chemical calculations of protein-drug molecule interactions are necessary.
However, evaluating protein-drug molecule interactions using conventional quantum chemistry calculation methods is challenging.
The Fragment Molecular Orbital (FMO) method allows for the calculation of protein-drug molecule interactions with quantum chemistry accuracy.
Still, even using the "Fugaku" supercomputer, it takes several hours per structure, indicating a need for further reduction in computational costs.
This study presents a machine learning model that predicts interaction values between proteins and drug molecules using the FMO method.
This model is based on a neural network and utilizes vectors of the surrounding environment of each atom in the drug molecule as explanatory variables.
Using a dataset of approximately 2000 structures, the model was trained and tested for predicting interactions in unknown structures.
The model successfully predicted protein-drug molecule interactions with an R2 value of 0.59.