We discuss a neural network (NN) approach to predicting the density-functional-theory-calculated properties of bioactive molecules. This method is guided by lower-level quantum chemistry and short-range atomic distances. We suppress the overfitting and unstable behaviors of our NNs via early stopping, max-norm constraints, and a sigmoid activation function. The trained models can predict molecular properties well, such as molecular energies, molecular forces, and atomic charges. The molecular forces (energy gradients) are evaluated based on a transformation technique from the atomic distances to Cartesian coordinates. We also execute molecular dynamics simulations to test the stability of NN models.

中文翻译：

---

基于半经验量子化学的引导神经网络方法的生物活性分子性质的密度泛函理论预测


我们讨论了一种神经网络 （NN） 方法来预测生物活性分子的密度泛函理论计算特性。这种方法以低能级量子化学和短程原子距离为指导。我们通过提前停止、最大范数约束和 sigmoid 激活函数来抑制 NN 的过拟合和不稳定行为。经过训练的模型可以很好地预测分子特性，例如分子能量、分子力和原子电荷。分子力（能量梯度）是根据从原子距离到笛卡尔坐标的变换技术来评估的。我们还执行分子动力学模拟来测试 NN 模型的稳定性。