Designing bioactive molecules that serve a purpose – such as binding specifically to a protein – is central to medicinal chemistry and a common practice in drug discovery. Although automating design has tremendous promise, general-purpose methods do not yet exist. Here we explore a simple, fast, and robust approach to inverse design which combines learned forward simulators based on graph neural networks with gradient-based design optimization. Our approach solves high-dimensional problems with complex physical dynamics, including designing molecules that induce proximity between proteins forming a ternary complex.

With more Heterobifunctional Small Molecules (hSM) being developed, the importance of linker design and optimization has been broadly recognized. Here we report an approach to virtually screen the linker designs by proactively using the available hSM ternary structures. Combined with expansive chemical building blocks, this approach helps us rapidly triage design ideas and expedite the discovery process.

While Targeted Protein Degradation (TPD) can expand druggable targets, how protein degradation is dysregulated in cancer and how TPD drugs counteract this effect is incompletely understood. First, we developed a deep-learning model (deepDegron) to identify mutations that result in loss of protein degradation signals. Second, we developed a machine learning model (MAPD) to predict which protein targets are likely degradable by TPD compounds from unbiased proteomic experiments of the kinome.

We will discuss an approach that applies AI and ML to design, test, and optimize lead molecules rapidly in silico and to suggest what compounds to synthesize and screen next in an ‘active learning’ process. A comprehensive platform approach offers tight integration between the virtual and real cycles (V+R). 3D modeling and simulation methods enhance the accuracy of predictions for drug potency, efficacy, and selectivity, while also addressing multi-target effects.

The mainstream view in improving model predictions is identifying better modeling algorithms and assembling larger training sets curated to get more reliable data. In addition to these approaches we use other measures that are very effective: (1) eliminating data points that confuse the model algorithm, (2) biasing property distributions in training set with the goal to maximize performance metrics. Examples will be given with demonstrated material improvements.

We built a map of the electrophysiology of human excitatory neurons, with applications to high-throughput screening, drug repurposing, and target deconvolution. First, we measured hundreds of parameters from millions of iPSC-derived neurons perturbed by tool compounds, yielding a database of activity profiles. Then, we used representation learning to distill these data down to concise fingerprints, where similar compounds form neighborhoods and several important tasks (e.g. drug repurposing, target deconvolution) reduce to path-matching problems.  

Accurately predicted binding affinities can be a powerful tool for virtual screening of ligands. Protein flexibility plays an important role in ligand binding. Hence, using parameters that determine protein flexibility will lead to better prediction of binding affinities in ML algorithms. This work will evaluate the use of such independent variables to determine binding affinities.

The gap between drug discovery and information science continues to close, hence there has never been a better time to leverage the power of AI/ML techniques to advance our understanding of the relationships between chemical and biological space. NCATS has identified, through the input of the greater scientific community, focus areas that need to be addressed in order to transform the design-synthesize-test cycle to transition to be more data-driven. The ASPIRE Program was created to support the development of AI/ML tools to process captured data to inform the next iteration of the process.

To effectively differentiate huge quantities of molecules in data-driven drug research, we have developed a low-dimensional representation of molecules and utilize it in deep learning for predicting molecular properties and designing new drugs. The concept centers on transforming a molecule's 3D electronic attributes of local hardness and softness to lower-dimensional manifold embeddings. The representation carries the inherent information of intermolecular interaction strength and specificity. Both predictive and generative deep learning models have been developed out of the concept of data-driven drug research.

CACHE is a benchmarking exercise modeled after CASP where every four months, computational chemistry and AI experts predict up to 100 compounds for a predefined protein target. Hit candidates are then procured, tested experimentally at CACHE, and all data and a generic description of the methods are released publicly. The emerging landscape of the most successful computational hit-finding approaches so far will be outlined.

There is no unified API for molecular property predictors (MPPs), which makes it difficult to share, distribute, version, retrain and manage predictors. We present Oloren ChemEngine (OCE), an open-source Python library with a unified, reproducible, and easy-to-integrate API for MPPs. Using OCE, we create models with the best leaderboard performances on 19 ADME/Tox benchmarks with MPP ensembling strategies. Using such API, we integrate model-agnostic uncertainty quantification and interpretability methods.

BigHat Biosciences is designing safer, more effective antibody therapies for patients using machine learning and synthetic biology. Machine learning guides the search for better molecules by directing and learning from each cycle of our high-speed, automated wet lab that synthesizes and characterizes hundreds of antibodies each week. We’ll highlight key features of our platform and share several case studies of protein engineering using this novel platform.

Successful peptide drug discovery programs today require attainment of multiple performance metrics to progress a compound to clinical stage. To aid decision-making, Koliber has developed an AI peptide platform based on state-of-the-art machine learning methods to analyze peptide properties, profile positions, and predict new variants. The capabilities and wet-lab validation of the AI platform will be demonstrated with examples from immunology and antimicrobial peptide discovery and optimization.