Artificial Intelligence applications in the drug design-and-development space continue to expand. The rapid integration of AI into numerous industry sectors has forced the United States to rethink its stance on AI- and software-based patent applications, and to shift policy to be more lenient toward applicants seeking to protect their developments in this space.

TuneLab provides access to Lilly-trained artificial intelligence models to participating biotechs. The platform employs federated learning, a privacy-preserving approach that enables biotechs to tap into and contribute to Lilly's AI models without directly exposing their proprietary data. The talk will provide background on federated learning and share examples of how participating companies are accelerating their programs.

Target discovery faces a critical bottleneck centered on cross-functional data democratization and analysis to identify high-confidence therapeutic targets, aligned with modalities offering the highest probability of success. To address this, we developed the Multi-Attribute Ranking System (MARS), an AI-empowered platform for multimodal, internal and external data synthesis. MARS supports efficient decision-making via: 1) customizable multi-target prioritization tailored to therapeutic area strategies, and 2) comprehensive single-target druggability assessments to guide modality selection.

The central challenge in medicinal chemistry is making optimal decisions on which compounds to design and advance, rather than just predicting molecular properties. We describe an integrated framework for program decision making that combines ADMET machine learning, physics-based methods, PBPK models, AI retrosynthesis, and AI chemistry assistants. This system organizes the Design-Make-Test-Analyze (DMTA) cycle around the clinically meaningful endpoint of predicted human pharmacokinetics and efficacious dose, enabling consistent, high-quality decision making. Case studies illustrate its use in lead generation and lead optimization campaigns.

LLM-based reasoning, generative design, and learned property prediction are changing our relationship with data in drug discovery, but these models demand more than convention can give. Zero-shot development candidates from a prompt remain ahead of us; what is emerging is iterative discovery that pairs pretrained models with adaptive learning. Ginkgo Datapoints has been building at that frontier, running automated labs and multi-party consortia to generate the data these loops consume. I’ll share what that has taught us about which data actually moves a model, and why the answer has often involved collaboration.

The Chatterjee Lab at Penn develops generative AI methods for designing functional biologics. We first built language models that created peptides to bind and control undruggable disease targets, with validation across neurodegeneration, pediatric cancer, and viral infection. We then developed discrete diffusion and flow-matching models to generate peptides, proteins, and mRNAs, that balance binding with drug-like properties such as solubility, stability, and safety. More recently, we introduced Schrödinger Bridge models that learn biological trajectories, enabling context-aware design across protein folding, cell-state transitions, and therapeutic intervention. 

We will present an AI-orchestrated lab-in-the-loop framework for oral peptide optimization that coordinates computational assessment and experimental validation across discovery cycles. By selecting the most informative readouts at each step, it enables simultaneous optimization of affinity, membrane permeability, and stability while reducing cost and shortening DMTA cycles. We will also show how hybrid AI + physics-based membrane permeability modeling is translated to GI permeability through delta learning.

Generative models have made it possible to design novel proteins, but translating this capability into medicines requires more than computational generation alone. I will describe how Generate:Biomedicines integrates generative and predictive models and high-throughput experimental measurements to translate biological hypotheses into therapeutic candidates.

OpenFold3 is an open-source AI model for predicting biomolecular structures across proteins, ligands, and nucleic acids, developed through the OpenFold Consortium. Built through collaboration among academic, biotech, and pharmaceutical partners, it combines shared priorities with coordinated investment in data, compute, and model development. Early benchmarking shows competitive protein-ligand performance, while ongoing fine-tuning is focused on antibody-antigen modeling and other translationally relevant applications for future drug discovery and structural biology use.

Predicting ADMET properties is vital for drug discovery, yet retrospective benchmarks often mask model failures in real-world applications. OpenADMET addresses this through blind challenges using prospective, unpublished data, establishing a transparent baseline for the state of the art. By exposing where current models fail to generalize to novel chemical space, these challenges cut through the noise of "optimistic" internal metrics. This presentation details recent outcomes and identifies the specific data standards and architectural shifts necessary to move beyond the hype, transforming predictive modeling into a robust, reliable tool for prospective drug discovery.