The importance of using AI in healthcare is well emphasized in the clinical community but translating such innovative tools for clinical applications and physician/patient utilization remains a challenge. I will highlight several unique challenges that have impeded the progress of AI in healthcare and will discuss some potential resources that could inform the implementation of AI in clinical practice, allowing clinicians to better diagnose and treat patients.

Currently, real-world data in pharma most frequently supplements clinical trials and complements regulatory submissions, however this may miss the real opportunity to improve target selection and drug development. RWD, focused on clinical data rather than operational data—and applied in early drug discovery—can significantly improve disease (and patient) stratification and reduce failure rates. Disease is a process, not a state, leading to “next-generation phenotyping” that stratifies most current diagnoses and enables differential diagnosis into disease subtypes. Examples will be presented including hypertension, preeclampsia, breast cancer, and menopause—and their impact on women’s health.

The advent of ultra-large screening libraries has created opportunities and challenges for virtual screening. With multi-billion molecule libraries like the Enamine REAL and WuXi GalaXi collections, brute-force evaluation is no longer a viable alternative. To meet this need, computational groups are developing active learning methods that use machine learning models as surrogates for more computationally (and economically) expensive calculations. This presentation will highlight applications of one such method, Thompson Sampling.

We live in a unique and exciting time. This presentation will showcase the latest developments in modeling, generative methods, and drug discovery, using real-life examples. The talk will emphasize that success in this field requires a combination of AI deployment and molecular thinking, as well as adherence to first principles. A culture revolution is currently underway, which enables the acceleration of higher-quality results.

From identifying a dark target implicated in hepatocellular carcinoma to leveraging an AlphaFold2 predicted target for the design of tool molecules, during this talk, I will take on the next chapter in this story, optimization of our CDK20 inhibitors using AI.

The rich structural context of fragment-based drug discovery opens up unique opportunities for artificial intelligence (AI) to assist us with the design of preclinical candidates. Here we discuss ideas around augmented interactive design as a strategy that integrates AI-driven approaches with human expertise—thus adding scale to the tradition of carefully handcrafted design. We will present some of the predictive and generative technology we have developed to incorporate prior knowledge and structural, synthetic, and directional constraints into the design process.

Our team has developed DrugInteLLM, a suite of LLM experts, each focused on different tasks and activities related to early drug discovery. PharosGPT (targets, diseases, ligands from pharos.nih.gov), litGPT (learn from papers), ChEMBLGPT (compounds and bioactivities) and ActivityGPT (predict bioactivity endpoints) provide LLM-based support for our projects. We will describe our platform for target and ligand selection, and how GPTs can support drug discovery. 

There is a defined need to properly manage and disseminate quality assured data. In practice this rarely happens. Beyond data generation, dissemination of high-quality data is often de-prioritized. Data are therefore generated in a format suitable for the experimentalist, but less so for computational methods. This talk will focus on generation and dissemination of such high-quality data, ranging from protein expression to data from patient-derived models, being suitable for data science.

The advent of giga-large make-on-demand combinatorial chemical spaces presents a great opportunity for drug discovery but requires novel computational approaches for fast and accurate screening. We have developed V-SYNTHES, a new iterative synthon-based approach for fast structure-based virtual screening of billions of readily available (REAL) compounds. I will discuss the latest developments of V-SYNTHES technology and its synergistic combination with machine learning tools for efficient exploration of giga-large chemical spaces.

I will provide a critical review of examples in the literature and industry to determine if, how, when and why using AI or machine learning has helped drug lead generation when using DNA-Encoded Libraries.

At Anagenex, we are coupling the high-throughput power of DNA-encoded libraries with Machine Learning in order to drive the hit-to-lead process. We will discuss a case study wherein we were able to drive a target campaign via leveraging focused libraries with Machine Learning to enable rapid chemotype expansion and decision making.

Machine learning models make better predictions of small molecule binders to proteins when they are built on better training sets. Training sets enable better models when they (i) comprise more, and diverse, true positives and negatives, and (ii) when the true positives are more accurately rank-ordered by affinity. We are building DELs and DEL selection methods that produce higher-quality training sets.