Over the past decade artificial intelligence (AI) and machine learning (ML) have demonstrated remarkable performance in a wide variety of tasks, from image processing to natural language processing to protein structure prediction. Well-trained ML models can provide valuable support to expert scientists in decision-making and can solve many complex problems. In general, ML algorithms become more effective as the size of training datasets increases.
At Selvita we take the hybrid approach of pairing our medicinal chemistry experts with an ensemble of AI and ML solutions; these tools help to streamline the process of lead discovery and lead optimization.
AI augmented HTS
Iterative Chem Space Exploration
An active learning process supports the High Throughput Screening (HTS) by iteratively selecting compounds for activity landscape exploration and allows also the detection of false negatives.
We offer a unique approach to building custom screening libraries by using AI-enhanced chemical space exploration. This approach allows us to focus our search on those regions of chemical space that are most relevant for a given biological target, which increases our chances of finding promising compounds.
Selvita provides focused compound libraries that are built with an AI based virtual screening against specified targets or target families.
Target Aware Drug Affinity Model
The Target Aware Drug Affinity Model (TADAM) is our in-house machine learning model for predicting the activity of a compound towards a given target. The proprietary technology is based on a deep learning approach that was trained on 900k activity data points from the public domain. Ligands and their corresponding protein binding sites are represented as graphs, which allows the model to learn the interactions between them. The TADAM engine is the core application for several use cases.
Large Scale Virtual Screening
- Compound prioritization and selection based on the predicted activity towards a given target
- Help to understand which parts of the molecule contribute to activity
- Provide guidance in the lead optimization process
Protein Pocket Similarity Search
- Analyzing and visualizing a protein database
- Search for targets with similar pockets for reference
- Identification of possible binding pockets
- Identification of residues involved in ligand binding
QSAR and QSPR modelling
Quantitative structure-activity relationship (QSAR) and quantitative structure-property relationship (QSPR) models help us estimate the physicochemical properties of new compounds and their activity in biological assays. Our experience in predictive modelling allows us to build customized solutions for various endpoints based on our clients’ data. This can significantly reduce the workload of our experts, as they can focus their experiments on the most promising compounds with respect to the desired properties.
In addition to classical modeling approaches, we are applying the concept of transfer learning which uses a pre-trained graph-based molecular representation. These models can be trained on large public datasets and then refined with confidential project data.