Collaborations

Vanderbilt has a strong reputation for industry-quality small molecule drug development driven largely by the Warren Center for Neuroscience Drug Discovery helmed by Craig Lindsley. The AI revolution has had a substantially smaller impact on CADD and cheminformatics than on protein structure prediction. Indeed, quantitative structure-activity relationship (QSAR) modeling, which forms the basis of modern ligand-based high-throughput virtual screening (vHTS), has not notably improved from approaches developed in the 2000’s – 2010’s. This discrepancy is the result of differences in data availability and the task-specificity of QSAR modeling. The laboratory of Jens Meiler has played an instrumental role in developing and applying these modern cheminformatics tools at Vanderbilt, and in cooperation with Craig Lindsley has a strong history of integrating computational and experimental approaches to discover new small molecule modulators of G-protein-coupled receptors (GPCRs) and other proteins.

​

The time is ripe to further extend Vanderbilt’s unique strengths in small molecule drug discovery by extending CADD services broadly to the campus community in an “in silico first” initiative. There are notable areas in which AI is leading to a new state-of-the-art in CADD. For example, protein-ligand docking is being revolutionized by AI trained with geometric deep learning procedures (e.g., EquBind, DiffDock), enabling the identification of likely binding modes of small molecules 1000x faster and with equivalent accuracy to conventional force field-based docking approaches. These new AI tools in combination with industry-standard conventional methods have the potential to accelerate research in multiple domains at Vanderbilt. Central to the success of the CSB and the PSB Program are outreach efforts to gain the insight and expertise of outside specialists to identify high impact areas of application. Supporting sustainable deployment of CADD efforts across campus will enhance the real-world impact of academic innovation.

​

There remain unsolved grand challenges in CADD that are ripe for reappraisal in the new AI era. Protein dynamics remain a major challenge in CADD. There are no CADD methods suitable for vHTS that can also account for largescale induced-fit changes. Essentially all modern methods, AI included, are limited to rigid backbone predictions with sidechain refinement. Induced-fit and conformational selection still require more expensive simulations that are fundamentally incompatible with screening more than a few hundred molecules in the most generous instances. AI systems incorporating reduced-MSA, RoseTTAFold, and additional algorithmic modifications could, for example, be developed to provide groundbreaking largescale induced-fit capabilities.

​

Another unsolved challenge in CADD is rapid binding affinity prediction. In the mid-to-late 2010’s there was much enthusiasm for training graph-based deep neural networks to predict binding affinities from protein-ligand complexes; however, such methods have since shown to be riddled with training set bias and ill-suited for prospective screening. Thus, the gold-standard binding affinity approaches are still physics-based estimates, such as free energy perturbation or potential of mean force methods, which are computationally prohibitive for vHTS and relegated to small scale lead optimization assistance. Despite these challenges, increasing the generalizability and quality of rapid AI-assisted binding affinity estimates is a tractable problem that could be addressed through experimental and computer-aided data augmentation coupled with debiasing the feature space of protein-ligand interactions, for example through feature engineering or message-passing neural networks.