System and method for prediction of protein-ligand bioactivity using point-cloud machine learning
Inventors
Bucher, Alwin • Prat, Alvaro • Bastas, Orestis • Pabrinkis, Aurimas • Kamuntavi{hacek over (c)}ius, Gintautas • Demtchenko, Mikhail • Macer, Sam Christian • YANG, Zeyu • Jamieson, Cooper Stergis • Jo{hacek over (c)}ys, {hacek over (Z)}ygimantas • Tal, Roy • Knuff, Charles Dazler
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11256995-B1
Publication Date
2022-02-22
Expiration Date
Abstract
A system and method that predicts whether a given protein-ligand pair is active or inactive, the ground-truth protein-ligand complex crystalline-structure similarity, and an associated bioactivity value. The system and method further produce 3-D visualizations of previously unknown protein-ligand pairs that show directly the importance assigned to protein-ligand interactions, the positive/negative-ness of the saliencies, and magnitude. Furthermore, the system and method make enhancements in the art by accurately predicting protein-ligand pair bioactivity from decoupled models, removing the need for docking simulations, as well as restricting attention of the machine learning between protein and ligand atoms only.
Core Innovation
The invention provides a system and method for prediction of protein-ligand bioactivity using a point-cloud based bioactivity module that receives a molecular structure file comprising molecular and structural information about at least one protein and one ligand. The system generates a graph-based neural network of the protein and a graph-based neural network of the ligand, with edges determined using the molecular structure file, and concatenates vectors from both neural networks for downstream learning and prediction.
The system performs restricted-cross-attention learning on the concatenated vectors, generates a single feature vector from the restricted-cross-attention learning, and uses the single feature vector in a feed-forward neural network to produce one or more outputs. The outputs are selected from active or inactive classification, crystalline-structure similarity, and regression analysis, and are used to produce one or more bioactivity predictions about one or more protein-ligand pairs.
The provided content emphasizes restricting cross-attention to protein-ligand atom interactions in proximity and employing crystalline-structure similarity to provide an assessment such as legitimacy. Related refinements condition the protein representation and/or operations on whether the protein and ligand are coupled or decoupled, including the selection of atoms within a proximity region and binding-pocket atoms under a decoupled representation.
Claims Coverage
The provided independent claims cover a protein-ligand bioactivity prediction system and a corresponding method. Each independent claim includes a common inventive sequence: graph-based neural networks for protein and ligand built from edges determined using the molecular structure file, concatenation of vectors, restricted-cross-attention learning, a single feature vector feeding a feed-forward neural network, and outputs including active/inactive classification, crystalline-structure similarity, and/or regression analysis to produce bioactivity predictions.
Point-cloud based bioactivity prediction system receiving molecular structure file and using protein and ligand graph neural networks
A point-cloud based bioactivity module receives a molecular structure file comprising molecular and structural information about at least one protein and one ligand; generates a graph-based neural network of the protein with edges determined using the molecular structure file; and generates a graph-based neural network of the ligand with edges determined using the molecular structure file.
Restricted cross-attention on concatenated protein and ligand vectors to form a single feature vector
Concatenates a set of vectors from both the graph-based neural network of the protein and the graph-based neural network of the ligand; performs restricted-cross-attention learning on the concatenated vectors; and generates a single feature vector from the restricted-cross-attention learning.
Feed-forward network outputs for active/inactive classification, crystalline-structure similarity, and regression bioactivity analysis
Uses the single feature vector in a feed-forward neural network to produce one or more outputs selected from active or inactive classification, crystalline-structure similarity, and regression analysis, and uses the one or more outputs to produce one or more bioactivity predictions about one or more protein-ligand pairs.
Graph-based neural networks, restricted cross-attention, and feed-forward network for a protein-ligand bioactivity prediction method
A method receives a molecular structure file comprising molecular and structural information about at least one protein and one ligand; generates a graph-based neural network of the protein with edges determined using the molecular structure file; generates a graph-based neural network of the ligand with edges determined using the molecular structure file; concatenates vectors from both neural networks; performs restricted-cross-attention learning; generates a single feature vector; and uses the single feature vector in a feed-forward neural network to produce one or more outputs selected from active or inactive classification, crystalline-structure similarity, and regression analysis for bioactivity predictions about one or more protein-ligand pairs.
Across both independent claims, the claim coverage centers on building separate graph-based neural networks for protein and ligand from a molecular structure file, concatenating their learned vectors, applying restricted-cross-attention learning to obtain a single feature vector, and using a feed-forward neural network to output active/inactive classification, crystalline-structure similarity, and regression analysis to drive protein-ligand bioactivity predictions.
Stated Advantages
Improves prediction by using restricted-cross-attention learning to focus the model on protein-ligand atom interactions in proximity.
Uses crystalline-structure similarity to assess legitimacy of one or more bioactivity predictions.
Avoids using docking simulations as required by the decoupled protein-ligand modeling approach described in the provided content.
Documented Applications
Protein-ligand bioactivity prediction for one or more protein-ligand pairs, producing active/inactive classification, crystalline-structure similarity, and/or regression analysis outputs.
Use of crystalline-structure similarity as an assessment to determine whether one or more bioactivity predictions are legitimate.
Three-dimensional visualizations of molecular interaction properties/saliencies derived from the model described in the provided content.
Interested in licensing this patent?