Ligand Docking Pose Ranking Using Neural Network Energy Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current docking methods for predicting small molecule conformations in structure-based drug design are inaccurate due to reliance on physics-based force fields and empirical terms, fail to consider multiple binding sites, and are computationally expensive, missing critical ligand portions and pharmacophores.
Innovation Solution
A deep neural network architecture, OrbitalDock, is used to rank ligand docking poses by learning from large-scale co-crystal structures without prior chemistry or physics knowledge, identifying pharmacophores and predicting ligand conformations efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based force field theories and empirical terms are used for scoring functions, then the docking predictions can be made, but the accuracy is insufficient and computational cost is high
Solution Approach 1:
The patent replaces physics-based force field theories and empirical scoring functions with a machine learning-based scoring function. The ML model is trained on known protein-ligand complex structures to learn accurate binding pose predictions without relying on computationally expensive physics calculations, thereby substituting mechanical/physical computation with data-driven prediction
Solution Approach 2:
The patent performs preliminary action by training the machine learning model on a large dataset of known protein-ligand complexes before actual docking predictions. This pre-training phase allows the model to learn accurate scoring patterns in advance, enabling fast and accurate predictions during actual docking without requiring expensive runtime computations
2Loss of information
If traditional docking methods are used, then docking predictions can be made, but they fail to identify critical ligand portions and pharmacophores
Solution Approach 1:
The patent applies segmentation by analyzing and identifying specific portions of the ligand that are critical for binding, rather than treating the ligand as a whole. The machine learning model learns to recognize and score individual ligand portions and their interactions with the protein, enabling identification of pharmacophores and critical binding regions
Solution Approach 2:
The patent introduces an intermediary mechanism by using the machine learning model as a mediator between the docking prediction and the identification of critical ligand portions. The ML model not only predicts binding poses but also provides information about which ligand portions are critical for binding, serving as an intermediary that bridges structure prediction and pharmacophore identification
3Measurement precision
If standard docking methods are used, then predictions can be made, but they do not achieve high accuracy compared to OrbitalDock
Solution Approach 1:
The patent applies parameter changes by transitioning from traditional physics-based parameters (force fields, empirical terms) to data-driven parameters learned from training data. The machine learning model uses learned parameters from training on known complexes, fundamentally changing the scoring approach from physics-based to statistics-based, achieving superior accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for predicting a conformation of a ligand docked into a protein is disclosed. According to some embodiments, the method may include determining one or more poses of the ligand in the protein, the poses being representative conformations of the ligand. The method may also include determining, using a neural network, energy scores of the poses. The method may further include determining a proper conformation for the docked ligand based on the energy scores.