Located at the Montreal Institute for Learning Algorithms, InVivo AI is developing novel algorithms for low data drug discovery.
We’re pioneering the use of transfer learning, few-shot learning, and meta learning in the prediction of key molecular properties, providing us the unique ability to work with the small and noisy datasets characteristic of early-stage drug discovery.
In collaboration with industry and academic partners, we combine previously generated experimental data with our pipeline for low data learning, enabling rapid iteration towards novel drug candidates meeting prespecified potency, selectivity, and ADMET criteria.
Pre-existing screening data are used to bootstrap our design process. We specialize in low-throughput, physiologically-relevant disease models and have validated our predictions across a range of complex phenotypic readouts.
Scoring functions are built for each of the desired endpoints. We leverage advances in meta learning and few-shot learning to ensure that accurate predictors can be obtained across all desired endpoints.
Novel compounds are then generated that are predicted to satisfy the design criteria. For many projects, this includes key ADME criteria against which only a handful of compounds have been profiled.
Selected compounds are synthesized for assay and profiling. In cases where data is extremely sparse, we can use active learning to leverage newly generated experimental data in subsequent design cycles, allowing for efficient iteration towards compounds satisfying the design objectives.
We integrate multiple deep learning techniques to overcome many of the challenges currently limiting computational drug design, allowing for rapid iteration toward chemically-diverse, lead-like drug candidates.
We use few-shot learning algorithms to learn accurate scoring functions from small, noisy, and highly imbalanced datasets.
Our approaches allow us to extrapolate beyond the chemical space contained within the training set and can be used in the optimization process without any consideration for the underlying target profile.
We use reinforcement learning algorithms to generate novel molecules optimized for multiple objectives, including key potency, selectivity, and ADME criteria.
We can also provide a synthetic route for every molecule we propose by leveraging known building blocks and reactions in the design process.
We use active learning algorithms to ensure that only the most information-rich compounds are selected for synthesis and assay, allowing us to arrive at compounds meeting the target profile in fewer iterations, and with far less data, than otherwise possible.
We can also incorporate newly generated experimental data into future predictions, ensuring our models get smarter over time.
Our interpretable, graph-based molecular representation allows us to quantify the importance of key fragments and functional groups.
By allowing us to see into the “black box” for the first time, we can provide our partners with insight into the impact of specific structural modifications.
Mila - Quebec AI Institute
6666 St Urbain Street
H2S 3H1, CAN
Cambridge Innovation Center
101 Main Street