Over the last decade artificial intelligence (AI) and machine learning (ML) in general have demonstrated remarkable performance in many tasks, reaching from image processing over natural language processing up to predicting protein structure. Well trained ML models can support the expert scientist in decision making and solve many complex problems. In general ML algorithms become more effective as the size of training datasets grows. At Selvita we take the hybrid approach of pairing our medicinal chemistry experts with an ensemble of AI and ML solutions; these tools help streamline the drug discovery process in many ways.
What can Selvita AI do?
HTS Analysis
- Identify false negatives/positives
We can identify up to 75% of misclassified HTS results through ensemble AI methods and Monte Carlo simulations.
- Hit expansion
With the use of several ML models we can analyze the results of an HTS and estimate the likelihood of new compounds to score as active. Our active learning approach selects iteratively compounds to be tested which leads to a consistent edge over a random sampling method. For instance, we obtained a top 1000-recall of over 55% while prescreening 16000 molecules.
Feature Estimation
- Structure property relation (SPR) and structure activity relation (SAR) models help us to estimate the physico-chemical properties of new compounds as well as their activity in biological assays.
- Based on a structure of a compound we can estimate its features, such as logS or BBB permeation (for which our AI obtained 90% accuracy).
- Our experience in predictive modelling allows us to build customized solutions for various endpoints based on our client’s data. Prescreening can reduce considerably the workload of our experts, focusing their experiments on the most promising compounds with respect to the desired properties
Lead Optimization
- Linker Building or Substituting
Starting from a ligand-linker system that exhibits good interactions with its receptor, but has suboptimal structure (e.g. excessive bond torsion, low novelty), our AI model can provide within an hour dozens of examples of alternative linkers , ready for our team of experts to be analyzed.
- Feature improvement
Our ML models can improve specific suboptimal compound properties while keeping the desired ones close to origin This scenario can also be used to introduce structural diversity, without compromising favorable medchem properties.
De Novo Design
- New Chemical Entities generation
Algorithms for generation of novel, but synthetically accessible structures based on active compounds, using Neural Networks to predict iteratively the optimal reagents in the process to create new compounds with similar properties to the input structures provided, along with the proposed synthetic routes.
- Parallel synthesis planning
The appropriate reagents and conditions are selected and grouped for parallel synthesis according to their reactivity using unsupervised Kohonen networks.
Docking Prediction
- Our AI model is capable of predicting the energy of interactions between any protein’s pocket with known crystallographic coordinates and any small molecule. This prescreening potentially increases throughput of our docking protocols to millions of compounds. Additionally, we can use it to find the most likely binding site of a protein with a known small molecule binder.
High Content Screening
- AI-supported image analysis improves our ability to identify and measure significantly more cellular parameters (min. 200) compared to a built-in analysis module (up to 30 parameters).