Mature CNS neurons fail to regenerate and no approved therapy promotes recovery from spinal cord injury. We identify four evolutionarily-related AGC kinase clusters whose functional overlap creates biological robustness, helping explain why single-target strategies have failed to yield regenerative therapies. Coordinated inhibition of all four clusters maximizes neurite outgrowth in rodent and human neurons. Their shared ancestry underlies structural homology enabling engagement by a single chemical scaffold. Phenotypically guided optimization yielded TMP-316, a clinical candidate potently engaging all four clusters with minimal off-target activity. A single intrathecal-dose produced sustained motor recovery after cervical hemicontusion in rats.

Targeting RNA remains a frontier of medicinal chemistry, and often a detailed structural understanding of a given RNA is lacking. In this talk, I will discuss our group's efforts to bring rational design to RNA-targeting molecules, including leveraging structure, conformational ensembles, machine learning, and new high throughput screening methodologies to identify selective, RNA-targeted probe molecules.

Ultra-large virtual screening (ULVS) is bounded by throughput. Most of chemical space stays out of reach utilizing conventional screening. Throughput itself should be treated as a first-order design variable. Expanding library size improves hit enrichment and quality while lowering per-compound cost by orders of magnitude. Utilizing Google Cloud, Model Medicines used its first-principles AI platform, GALILEO, to complete a record-breaking ULVS of 325 billion compounds in 24 hours. Case Study: the record-breaking ULVS delivered MDL-4102, a BRD4-selective inhibitor with no detectable BRD2/BRD3 activity and scaffold novelty relative to clinical BET inhibitors.

AI is transforming early drug discovery by enabling more efficient and information-rich hit identification and characterization. This talk will describe how modern screening informatics, coupled with machine learning, integrates biochemical, cellular, and phenotypic assays to extract robust signals from noisy, high-throughput data, enabling hit finding and clustering.

Covalent drug discovery is undergoing a renaissance, unlocking targets long considered beyond the reach of conventional approaches. By weaving together AI, chemoproteomics and quantum mechanics into a single integrated workflow, the Frontier platform—powered by Covalent AI—is purpose-built to turn "undruggable" into actionable, opening the door to therapeutic intervention across the vast majority of the human proteome.

Protein kinase inhibitors represent one of the most promising avenues for cancer therapy. However, drug resistance to kinase inhibitors frequently occurs in cancer. Tumor cells can circumvent kinase inhibitors through kinome reprogramming, a process characterized by system-wide changes in protein kinase networks. Here, we combined kinome profiling and phosphoproteomics to define the fraction of the kinome activated by targeted therapies to rationally predict combination therapies in various cancer models.

DegDig is a high-throughput, target-guided proteomics platform for quantitative profiling of degrader activity across the proteome. It measures multiplexed protein abundance changes after degrader treatment, providing direct readouts of degradation efficiency (Dmax/DC50) and off-target selectivity. Using a 384-well workflow, predefined protein panels, nanoscale TMT labeling, and next-generation mass spectrometry, DegDig enables scalable screening, optimization, and mechanism-of-action studies.

Membrane proteins represent the most drugged protein class yet remain underrepresented in quantitative ligand-binding datasets. Using affinity selection mass spectrometry applied directly to native membrane fractions, we generated standardized dissociation constants (Kd) for ~3,400 ligands across ~400 transmembrane proteins. Comparison with AI-predicted affinities reveals variability in prediction–measurement agreement across targets (R = 0.04–0.68), demonstrating that target-specific calibration is required and establishing experimental ground truth for reliable AI-driven drug discovery.

Polo-like kinases (PLKs) are a group of five related serine/threonine kinases involved in cell cycle regulation. While inhibitors of PLKs have entered clinical trials for oncology, serious adverse effects have been noted. Our current understanding of PLK biology is limited by the lack of selective inhibitors. Using machine learning pharmacophore models, we successfully designed potent PLK3 selective probes to further our understanding of the complex PLK biology.

Cryptic pockets are binding sites that are formed or exposed upon a conformational change. They represent an important class of potentially druggable binding sites. Reliably predicting cryptic pockets capable of binding ligands, however, remains a challenge. The use of AlphaFold 3 (AF3) for generating realistic conformational ensembles that include known cryptic pockets will be presented. We find that AF3 is generally able to reproduce the scale of conformational change required for cryptic site formation.

Target safety assessments (TSA) are a critical, resource-intensive gate in early drug discovery. We present a modality-agnostic, agentic AI framework that automates the end-to-end TSA workflow. During early discovery the platform bridges target and compound safety by integrating computational toxicology predictions directly into the assessment, uniquely identifying possible toxicities  across a range of species, and identifying its effect to be compound-specific, scaffold-specific, on-target or off-target activity.

We present a new 3D fibrotic scar microphysiological system (MPS) derived directly from the liver of patients with Metabolic Syndrome Associated Hepatosteatitis (MASH). Unlike animal models or iPSC derived MPS, this patient derived MPS enables human disease-state preservation, spatially resolved cell-type interactions, and functional, cell-selective lipid nanoparticle (LNP) mediated testing of therapeutic RNAs within a human fibrotic liver microenvironment. Therefore this patient derived MPS provides a preclinical tool with direct human relevance and high potential for clinical translation.

We have developed a platform of human 3D neural spheroids and bioprinted tissues assembled in multi-well plate formats using iPSC-derived cells optimized for HTS. These models enable the generation of functional, brain region-specific models containing neurons, astrocytes, and microglia. Functional activity is assessed using calcium imaging, fluorescent neurotransmitter biosensors, fiber-based electrophysiology, and additional HTS-compatible endpoints. This talk will describe their applications in pharmacological testing, neurotoxicity assessment, and neurological drug discovery.

Immunogenicity risk using in silico tools and in vitro assays can be combined to prioritize and mitigate during the biotherapeutics discovery process. An ever-increasing diversity of biotherapeutic modalities requires an understanding of immunogenicity risk factors and modification of in vitro assays and in silico tools used to assess and mitigate these risks. Here we will discuss workflow immunogenicity strategy integration into the biotherapeutic discovery process.

Emerging therapeutic modalities increasingly work by creating, stabilizing, or modulating biomolecular interactions. We present physics-aware AI approaches that integrate deep learning with molecular physics to engineer interactions across scales—from allosteric modulators, molecular glues, and degraders to antibodies, bispecifics, and agonistic biologics. We further extend these approaches to systems-scale interaction networks, connecting molecular design with cellular context to discover new mechanisms and therapeutic opportunities.

Drug discovery produces an enormous range of data, from structure to assays to images to omics. Our goal is to bring these modalities together in a single modeling framework. In this talk, I’ll discuss two steps toward that vision: NeuralPLexer, for understanding protein-ligand binding, and Enchant, for learning broad, multimodal representations of molecules and their biological consequences.

Large-scale perturbational profiling at single-cell resolution is transforming how we identify and validate drug targets. We describe the construction of unified datasets spanning 100 million to one billion cells, integrating thousands of chemical and genetic perturbations across diverse cancer models. We discuss the engineering challenges of building these atlases, computational approaches to extracting biological signal at scale, and how these integrated resources are accelerating target identification and preclinical drug discovery.

We present label-free single-cell buoyant mass profiling as a rapid (<20-minute) phenotypic screening platform for pre-clinical drug discovery. Using a Suspended Microchannel Resonator, we map a biophysical-transcriptomic continuum in resting T cells. Heavier cells exhibit a scalable "translation engine" primed for activation, whereas lighter cells display distinct exhaustion signatures. This stimulation-independent assay provides a novel functional endpoint for evaluating immunomodulators, validating targets, and optimizing next-generation cell therapy manufacturing.

Dysregulated extracellular proteolysis is a hallmark of cancer but remains challenging to systematically profile within intact tissues. We introduce PSurf, a platform that discovers tissue-specific protease sensors by differential selection on tumor versus healthy specimens. Applied to lung metastasis, PSurf yields substrates that distinguish metastatic lesions in tissue slices and, when incorporated into nanobody-targeted urinary biosensors, enable precise, noninvasive, tumor-specific detection in vivo and support development of conditionally activated therapeutics.