The promise of AI-driven discovery lies in creating true “closed-loop” systems where models continuously learn from experimental feedback. This panel explores real-world progress toward integrating design, make, and test cycles through automation, active learning, and data infrastructure. Speakers will discuss how chemists, biologists, and AI systems collaborate in real time to improve decision quality, accelerate iteration, and quantify learning efficiency. We’ll examine what’s working, what’s not, and how to scale self-improving discovery systems across diverse therapeutic areas.

Many small-molecule compounds are mischaracterized, creating major risks for both patient care and scientific reproducibility. To solve this problem, we have built Compound VALET, an LLM platform that integrates genetic, biochemical, and biophysical evidence to generate citation-verified reports on a drug’s selectivity and specificity. Using Compound VALET, we evaluated ~2,000 recent publications in cancer research, which revealed the widespread use of promiscuous or unverified chemical probes.

Terray’s platform, EMMI, integrates proprietary ultra-high-throughput experimental hardware capable of generating billions of protein-ligand binding data points with proprietary AI models that iteratively guide experimentation and design at all phases of discovery. For molecular design, the platform uses models that allow us to Reason, Generate, Predict, and Select in driving DMTA cycles for our pipeline programs. This talk will provide an overview of the EMMI architecture and present a case study demonstrating how the platform can accelerate drug discovery for challenging targets.

Using Iambic’s AI–driven platform, we developed a next-generation HER2 inhibitors with exceptional selectivity, broad mutant coverage, and robust brain penetrance, now under evaluation in a Phase 1/1b clinical trial. This compound representsthe first reported Type II HER2 tyrosine kinase inhibitor, binding HER2 in a DFG-out conformation. This program was powered by AI technologies, including PropANE (a precursor to our Enchant platform) and NeuralPLexer, integrated with high-throughput parallel synthesis and screening to accelerate the design and optimization of potent candidates with broad therapeutic potential in HER2-driven cancers.

Serna Bio is redefining what’s possible in drug discovery by opening up RNA as a tractable target class for small-molecule therapeutics. Using proprietary datasets and multiple machine learning architectures, Serna Bio's GenAI engine has outperformed benchmark models in RNA-relevant chemical space. Combined with multi-parametric optimization functions, our platform can rapidly reduce the time to DC nomination.

Lessons Learned from CACHE Computational Hit Finding Challenges So Far

We assess the performance of co-folding methods on fragment screening tasks in varying degrees of difficulty, including their ability to identify fragment binding sites, separate fragment hits from misses, and predict fragment binding modes. We evaluate the utility of these models during the hit-to-lead stages in terms of their ability to predict binding modes and, critically, induced fit. As a benchmark, we use the most comprehensive in-house fragment-based drug-discovery dataset.

This talk will demonstrate how a structural proteomics and AI integrated framework informs degrader design. It will showcase lead optimization strategies to enhance the efficacy and selectivity of degraders and preclinical results from Protai’s KAT6 degrader program will be presented as a case study.

Oral bioavailability of TPDs, especially larger sized bifunctional molecules (i.e. PROTAC), is one of the most challenging properties to be optimized. In this study, a comprehensive in vivo data set for CRBN-based PROTACs was collected from public domain and 2D/3D descriptors were developed to establish predictive models for oral bioavailability prediction. We address the different behavior of predictive models on different types of animal models, like mouse and rat models.