Summary
Cortex Discovery is a startup based in Munich focused on providing extremely accurate machine learning models for hit discovery and lead optimisation phases of drug discovery. Vladimir Chrorošajev, Machine Learning Researcher at Cortex Discovery, posed the question Why bother focusing on this task? Many structure activity related models are typically trained to do a single task, such as predicting activities for one specific assay or one specific modality, which limits accuracy.
Meanwhile, at Cortex Discovery, the aim is to break accuracy barriers by capitalising on deep learning methods to train models on all the publicly available data and adding proprietary data to this. Chorošajev explained that this allows them to break accuracy thresholds on traditional quantitative structure-activity relation models.
The platform is an extensive neural network that ingests millions of separate molecules over thousands of separate assays to derive the most efficient compressed representation, a multi-dimensional vector that would encode all the information related to activity. These models are trained on all available assays, which comprise over 5,000 assays, 3.5 million molecules, and 750 million data points.
Chorošajev explained that these deep learning models generate context-aware, multi-modal representations of molecules that can generalise across assay types, including binary, regression, and dose-response. Although one could argue that this context-aware representation applies to standard models, Chorošajev suggests that it is more applicable in Cortex Discovery’s models because of their multi-modal nature and ability to learn binary and continuous regression data and dose response curves. Furthermore, the models match experimental reproducibility accuracy, achieving AUC scores around 94-95%.
The key value proposition of this service is the ability to extract results from small-scale HTS assays of around 25,000 – 50,000 molecules, which can save time and costs. Additionally, the models provide predictions for efficacy but also for off-target effects, ADME tox, and counter screens.
Cortex Discovery’s in-house pipeline mostly focuses on age-related disorders and longevity, but the company has also had success in the oncology field. Chorošajev and his team have discovered novel oncology compounds and GP41 inhibitors. Alongside this, they have identified NRF2 activators, 4 mTOR inhibitors, and DNA enhancers for aging-related research.
Now, Cortex Discovery is developing methods for binding affinity prediction without prior HTS data. Chorošajev uses reinforcement learning and force-field optimisation to speed up free energy calculations
