Conformation Sampling via Monte Carlo and Neural Network Energy Evaluation

The patent employs a deep neural network model that has been pre-trained on quantum mechanical calculations to rapidly evaluate conformation energies. This allows the system to use approximate but sufficiently accurate energy evaluations many times during the Monte Carlo sampling process, rather than performing expensive quantum mechanical calculations each time. The neural network serves as a computationally inexpensive surrogate that maintains adequate precision for identifying key rotatable bonds.

The patent performs preliminary quantum mechanical calculations to train the deep neural network model before the actual conformational analysis. This pre-training phase captures the complex energy relationships in molecules, allowing subsequent energy evaluations to be performed rapidly using the trained neural network without sacrificing accuracy for the specific chemical space being explored.

The patent implements a feedback mechanism where the Monte Carlo sampling process uses the evaluated conformation energies to guide subsequent sampling decisions. Conformations with lower energies are more likely to be accepted and added to the conformation pool, while higher energy conformations are rejected or accepted with lower probability. This feedback loop allows the sampling to focus on energetically favorable regions of conformational space, achieving reliable results with fewer sampling rounds.

The patent uses a dynamic sampling strategy where the sampling process adapts based on the energy landscape being explored. The Monte Carlo algorithm dynamically adjusts which rotatable bonds are sampled and by how much, focusing computational effort on bonds that significantly affect conformational energy and on conformations that are likely to be energetically favorable, rather than uniformly sampling all possible conformations.