In recent years, artificial intelligence (AI) has re-emerged due to the continuous advancement of computing abilities, big data and algorithms. With the development of chemoinformatics, AI technology has shown great potential in the field of chemistry. AI brought new opportunities for the development of the prediction of synthetic products and reaction path prediction and has been widely applied in drug design. Alfa Chemistry has applied advanced AI technique to provide accurate prediction of synthetic products and study the path formation.
Figure 1. Synthetic Products Prediction and Path Formation. (Wu, G.; et al. 2016)
Machine Learning
Machine learning builds a hypothesis space in which the program learns common patterns in the data, and builds models and tries to predict the true value of the benchmark. The closer the predicted value is to the baseline true value, the more accurate the model. Through such a process of continuous learning and optimization, the program tries to obtain the learning and problem-solving capabilities of the human brain.
De novo Design
We use algorithms in de novo molecular design methods to virtually design and evaluate a series of molecules that meet specific properties for the prediction of synthesized product molecules.
Our Work Flow
At Alfa Chemistry, we have applied advanced synthetic programs combined with complex algorithm and descriptors to provide reliable results of synthetic products prediction and path formation. Our process is as follows:
- High-level logic-based programs design
The program we designed will output synthesis suggestions for a specific reaction.
- Detailed synthetic planners design
Our synthetic planners will give a complete synthetic route (even reaction conditions). This process requires a module to resolve the target compound, and these modules usually apply rule-based or irregular methods to propose possible transformations.
- Data set preparation
The data set can be obtained from SciFinder, Reaxys, patents, published chemical literature or proprietary databases. We construct a data set to build the model that include data points representing the range of possible parameter values to avoid overfitting or introducing the training set deviation.
- Descriptor selection
Descriptors can be divided into physical-based descriptors or information-based descriptors. When building a reaction prediction model, one must consider how many dimensions or reaction variables need to be modeled and how to most effectively characterize them. Variables commonly used in reaction prediction include substrates, solvents, temperature, additives, bases, and ligands.
- Algorithm selection
The machine learning algorithms used in synthetic chemistry can be divided into linear and nonlinear categories. At Alfa Chemistry, random forest, k nearest neighbors, support vector machine and neural network methods are available.
- Results output
The synthetic products prediction and path formation are obtained.
Products Prediction
- We have established a global prediction model based on machine learning methods using deep neural networks (DNN) to predict the properties and activities of synthetic compounds.
- Our experts screen common machine learning methods by using cross-validation methods and find that the global model trained with neural network or XGBoost algorithm can best predict the synthesized product.
- A SPOC descriptor that combines molecular fingerprints and physical organic parameters has been developed to realize the rapid and accurate prediction of the products of the multi-solvent system.
- Alfa Chemistry also supports the application of molecular diagrams to represent reactant molecules. The nodes and edges of the molecular graph are used to describe the atoms and chemical bonds, respectively. The graph convolutional neural network can be used to calculate the possibility of chemical bond changes between each atom pair. The most likely candidate products are listed to predict the probability distribution of the main products.
Chemical Reaction Path Formation
- At Alfa Chemistry, we apply custom molecular orbital concepts and physical and chemical descriptors as input, and our scientists can predict the reaction with high precision through training on the reaction data.
- Our well-designed strategy takes into account the specific reaction conditions, so it can get more realistic and credible results.
- In addition, our methods explain the elementary process of electron transfer in chemical reactions to a certain extent from the mechanism level, and can identify and predict multi-step reaction processes.
- To construct an effective reaction condition optimization system through machine learning methods, we use a layered neural network model to predict the chemical environment (catalysts, solvents, reagents) and reaction temperature, and the most likely reaction path formation can be obtained.
Features of Our Synthetic Products Prediction and Path Formation
- Open large database
- Access to high-quality and standardized data
- More effective expression of molecules and reactions
- Machine learning algorithms suitable for the study of chemical reactions
- Effective and universal algorithm evaluation benchmark
Our synthetic products prediction and path formation services remarkably reduce the cost, promote further experiments, and accelerate the process of drug design for customers worldwide. Our personalized and all-around services will satisfy your innovative study demands. If you are interested in our services, please don't hesitate to contact us. We are glad to cooperate with you and witness your success!
Reference
- Wu, G.; et al. Metabolic Burden: Cornerstones in Synthetic Biology and Metabolic Engineering Applications. Trends in Biotechnology. 2016: 652-664.