Artificial intelligence (AI) and machine learning (ML) experts at Psivant Therapeutics leverage data and cutting-edge algorithms to solve real-world problems in drug discovery. We apply AI/ML broadly to accelerate the progression of projects toward the clinic, minimize development costs and maximize the probability of success. Our models use molecular representations ranging from fast and efficient 1D encodings to 4D dynamic trajectories with spatial-temporal information and property optimization.
We are pioneering a more efficient discovery process through ML-assisted decision making, target/modality selection and focused property predictions. We combine experimental and in silico data to build predictive models that encompass target selection, screening, molecular design and clinical trial decisions. Our integrated platform includes models for binding pocket identification, druggability assessment, binding affinity, selectivity, off-target effects and physicochemical/pharmacokinetic properties. Our design workflows use biological and chemical property models to accelerate the Design-Predict-Make-Test-Analyze (DPMTA) cycle and enrich the discovery process for potent, drug-like molecules. The capabilities in QUAISAR allow us to rapidly explore larger and more diverse swaths of the vast chemical space to select molecules that have promising properties. Molecules with good drug-like properties are further interrogated using our physics-based in silico assays that include quantum physics, statistical thermodynamics and molecular simulation. These calculations run on our extensive high-performance computing (HPC) resources to enable the speed and scale needed to inform decisions in our drug discovery projects.
Exploring Chemical Space
We navigate vast chemical spaces using a combination of reaction-based enumeration and generative chemistry. DPMTA project cycles often involve millions or billions of virtual molecules that we explore each week based on structural hypotheses. We include information about chemical building block availability and synthetic tractability and “live” intellectual property (IP) scoring to minimize the time and cost of making the best novel molecules. Finally, intuitive maps of global and local chemical property spaces allow our drug discovery teams to efficiently navigate the immense property and IP space.
Predicting ADME and Toxicology Endpoints
We develop a combination of global and local models to predict properties that matter in drug discovery. Global models are applied early when project-specific data are not available. As project data accumulate, we augment global models with local, focused models that provide more accuracy for the chemical space of interest. Psivant develops accurate ML models to predict pharmacokinetics, metabolism and toxicity which we continuously deploy and retrain in projects. We also develop permeability and solubility and target binding models for small molecules covering thousands of proteins to address selectivity and secondary pharmacology.
Improving Physics-Based Models
We use AI/ML to improve our physics-based models and are currently generating a data set of high-level quantum mechanical calculations on hundreds of thousands of molecules and millions of their conformations. We have trained ML models using this data set and other molecular databases to predict properties such as molecular energies, 3D geometries and force field parameters at a fraction of the computational cost and with little loss in accuracy when compared to quantum mechanics calculations. AI/ML also helps us elucidate reaction coordinates and collective variables that we use to enhance sampling speed, accuracy and interpretability of our molecular simulations.
Augmenting Human Intelligence
The AI/ML-powered QUAISAR platform is designed to expand human imagination. While scientists continue to drive drug discovery projects and make decisions at Psivant, QUAISAR accelerates this process to an unprecedented scale. Empowering our researchers with the most accurate predictive tools and intuitive graphical representations allows them to make better decisions faster and spend more time on creative, critical thinking. We develop our AI / ML methods with the goal of enabling our world-class research teams to bring medicines to patients as efficiently as possible.