Implementation of AI-enabled specialized search engines for scientists is imperative to enhancing user experience. The search engines, powered by natural language processing (NLP) algorithms to understand and interpret scientific language, enables researchers to find relevant information quickly and easily. Using ML techniques, the search engines learn and refine results over time, making them more accurate and efficient. It identifies patterns and relationships in scientific data leading to new discoveries and breakthroughs.

Drug discovery involves three legs of the triangle: Disease, Target, and Drug. Significant efforts focus on identifying targets (network and pathway-based analysis) and in identifying drug candidates (optimizing the drug-protein interaction). Drug researchers rarely deal with current diagnostic limitations, complex disease etiologies, or disease heterogeneity. Novel methods for disease stratification, e.g., knowledge graphs and quantum computing, redefine the disease process; examples from MS, TNBC, and rare diseases will be presented.

Given the wide variety of AI tools currently being deployed in radiology and pathology, it is feasible to bring AI to clinical care. However, the calibration of these tools (i.e, the data, the models and inference) for cost-sensitive situations like healthcare is an important point. Using the example of tumor segmentation task, we will review a few algorithms and describe their potential modes of failure in addition to scoring rubrics related to data and model veracity. This talk will use case studies from these domains to discuss these aspects and their use for adoption of AI.

Along with target identification, understanding mechanism of action and identification of pharmacodynamic biomarkers are central to early drug discovery. The right information and its intelligent application allow bridging preclinical and clinical development. Here, I will describe how we apply statistics and decision tree-based machine learning to proteomics, transcriptomics, and metabolomics experiments to accelerate those efforts. Through real-life case studies, I will showcase the transformation in data analysis methods that took place in recent years.

Cancerous polyploid cells become resistant to chemotherapy and escape cell death. Doctors at University of Texas Health - San Antonio are developing drugs to be administered along with chemotherapy that are effective at reducing the cancerous polyploid cells. Deep learning is used to identify polyploid cells pre- and post-administration of drug candidates to evaluate their effectiveness.

We benchmarked the ability of commercial and open-source PROTAC virtual screening tools to 1) predict the ternary complexes observed in crystal structures and 2) dissociate active from inactive PROTACs. These tools are better able to predict near-native ternary complex structures than traditional protein–protein complex prediction software, but PROTAC virtual screening efficiency is unclear and highly variable. More experimental data is necessary for further improvement.

We present a method for generating high accuracy structural models of E3 ligase-PROTAC-target protein ternary complexes and of the full degradation assembly. The method is dependent on two computational innovations: adding a “silent” convolution term to an efficient protein–protein docking program to eliminate protein poses that do not have acceptable linker conformations and clustering models of multiple PROTACs targeting the same target. We validate the approach on known systems, as well as blindly on new PROTACs.

Pregnancy-Associated Plasma Protein A (PAPP-A) is a metalloprotease that regulates Insulin-like Growth Factor (IGF) bioavailability and signaling. Here we present the Cryo-EM structures of holo-PAPP-A and PAPP-A in complex with substrate Insulin-like Growth Factor Binding Protein 5 (IGFBP5), which in combination with biochemical experiments explain the mechanisms of substrate recognition and selectivity. The AI-developed AlphaFold structure predictions for PAPP-A and IGFBP5 were crucial for rapid structure determination in this study. Using the PAPP-A structures as a case study, the benefits and caveats of utilizing AI in novel protein structure solution will be discussed.

iBio has developed a machine learning technology for controlling the epitope binding site in antibody discovery. This is accomplished by computational design of peptides that embody the target epitope sequence and structure. These peptides are then used in immunization or in vitro screening to improve the efficiency of discovering antibodies that bind to the therapeutic site of interest.

Generative AI can enable drug discovery for known and novel targets alike, potentially overcoming otherwise prohibitive costs and timelines of early discovery in low-data regimes. Peptilogics’ Nautilus AI platform for peptide drug design combines proprietary generative and predictive AI algorithms for hit ID, hit-to-lead, and lead optimization. Nautilus has been applied in-house and through partnerships to diverse targets including a previously undrugged GPCR and therapeutic areas.