基於官能基分類量子化學數據集應用於機器學習預測分子間相互作用能

Abstract

過去幾年，我們實驗室透過量子化學的計算方式，建立了8種常見官能基二聚體的交互作用能數據庫，但隨著分子結構逐漸複雜，使用此方法的計算成本也逐漸上升，目前實驗室數據庫裡的分子二聚體基本停留在十至二十個重原子。如果需要繼續處理更大的分子二聚體，則目前使用CPU-based的量子化學計算方式較難完成計算，為此我們嘗試將現有的二聚體分子數據庫，與機器學習的方法做結合，來得到低計算成本又可靠穩定的替代方式。 我們選擇了一個透過前饋式神經網路建構的機器學習框架，經由具物理意義的描述符來來描述訓練集裡分子二聚體的原子與原子對環境，再透過神經網路將不同特徵進行串聯，訓練出一個可對未見過的分子二聚體資料進行預測其相互作用能的有效模型。 我們認為，實驗室過去以常見的官能基，分子二聚體總共原子數從小到大等規則來建立的二聚體交互作用能數據庫，對機器學習來說能夠更好的訓練出模型，且為了應付更廣泛官能基的分子類型，我們透過其他學者所發表且公開的量子化學數據集中，挑選我們實驗室尚未計算的官能基種類分子進行補充，再將結果拿來測試其他已公布的龐大量子化學數據集，藉此驗證結果的可行度。

In the past few years, our laboratory has established a database of interaction energies for 8 common functional group dimers through quantum chemistry calculations. Molecular dimers in the chamber database basically stop at ten to twenty heavy atoms. If it is necessary to continue to process larger molecular dimers, it is difficult to complete the calculation using the CPU-based quantum chemical calculation method. For this reason, we try to combine the existing dimer molecular database with machine learning methods to Get a reliable and stable alternative with low computational cost. We chose a machine learning framework constructed through a feedforward neural network to describe the atoms and atom-pair environments of molecular dimers in the training set through physically meaningful descriptors, and then concatenated different features through a neural network.To train an efficient model for predicting interaction energies from unseen molecular dimer data. We believe that the dimer interaction energy database established by the laboratory in the past based on common functional groups, the total number of atoms of molecular dimers from small to large, etc., can better train models for machine learning, and in order to cope with a wider range of molecular types with functional groups, we select molecules of functional groups that have not been calculated in our laboratory through the published and public quantum chemical data sets of other scholars to supplement, and then use the results to test other published large quantum chemistry. dataset to verify the feasibility of the results.