UCSF

Antigen presentation by class I MHC is a vital process for T cell surveillance in humans. However, peptides with tumor mutations are rarely presented on the cell surface. In this study, we address this challenge through a chemical approach. We aim to identify small-molecule glues (SMGs) that stabilize the binding of peptides derived from cancer driver proteins to class I MHC alleles. We focus our screening efforts on the HLA-B*57:01 allele, as it is amenable to small-molecule regulation by abacavir. We identify SMGs that facilitate the binding of tumor-specific peptides or peptides that cannot naturally be presented to class I MHC, thereby restoring surface presentation. We provide molecular insights into the SMGs using X-ray crystallography and biochemical assays. Additionally, we develop a computational pipeline called Tryptophan Scan to identify other peptides that can potentially form MHC complexes with SMGs. We investigate the immunological response to the SMG, the SMG-peptide-MHC ternary complex, and the broader impact of SMGs on immune peptidomes and therapeutic strategies. The latest phase of this project focuses on exploring the use of a transgenic mouse model to identify specific T cells that recognize SMG-peptide-MHC ternary complexes.

Spatiotemporal control of adherens junction fluidity and integrity is critical for sprouting angiogenesis, but underlying mechanisms are incompletely understood. To identify unappreciated regulators of VE-cadherin, we performed proximity ligation mass spectrometry, revealing significant interaction with the multifunctional scaffold Scribble (Scrib). Utilizing a 3D angiogenesis-on-chip model, I discover that SCRIBKO microvessels display reduced multicellular sprouting and increased single-cell detachments that are associated with adherens junction instability and decreased actomyosin in the junction-proximal cortex. I find that this disruption is not due to defects in VE-cadherin coupling to catenins or actin. Instead, SCRIBKO causes loss of cortical actomyosin clusters, which organize cortical actomyosin architecture and dynamics to stabilize adherens junctions. Moreover, I discover that unconventional myosin 1c is a critical effector linking Scrib cortical dynamics to VE-cadherin to regulate stabilization of adherens junctions during angiogenic initiation. This work demonstrates a new role for Scrib directly regulating cortical actomyosin cluster organization, critical for precise control of adherens junctions during angiogenesis. The thesis illustrates how engineered models of the vasculature can reveal novel cell biology underlying morphogenetic mechanisms. I review how the implementation of such models has also uncovered new mechanisms of vascular-tumor interactions during metastasis.

Complex neurological diseases arise from a combination of genetic and environmental factors, many of which converge on gene regulation within diverse brain cell types. While genome-wide association studies (GWAS) have identified numerous noncoding variants associated with these diseases, their cell type-specific functions and regulatory mechanisms remain poorly understood. In particular, rare noncoding variants—despite their higher effect sizes and substantial contribution to disease heritability—remain largely uncharacterized. Furthermore, the role of genetic variation in brain vascular, perivascular, and immune cells in disease risk has been underexplored.

This work addresses this gap through two parallel studies. First, we constructed a large-scale multi-omic atlas of the human brain, integrating single-nucleus RNA-seq and ATAC-seq across >3.3 million nuclei from five brain regions in 101 individuals with matched 30x whole-genome sequencing. We employed a comprehensive functional genomics framework to map enhancer-gene interactions, train machine learning models predicting variant effects on transcription factor binding and nominate rare noncoding variants with disease relevance. In parallel, we developed MultiVINE-seq, a vessel isolation and nuclei extraction approach, to generate paired transcriptomic and epigenomic profiles of human brain vascular cells across 30 individuals with varying cognitive statuses. By mapping disease-associated variants onto our multiomic atlases, we identified cell type-specific regulatory elements perturbed in neurological diseases.

Our findings expand the role of rare and common noncoding variants in neurological disease, highlighting the importance of vascular and immune contributions alongside neuronal dysfunction. This work provides a roadmap for future studies linking genetic variation to cell type-specific regulatory mechanisms in complex diseases.

Aberrant DNA methylation is a hallmark of many cancers. As such, there has been substantial interest in the development of anti-cancer strategies which modulate epigenetic programs associated with alterations in DNA methylation. In acute myeloid leukemia (AML), decitabine is a clinically-approved DNA hypomethylating agent used for a subset of high-risk patients with poor prognoses. Despite the clinical use of this drug, and clear evidence of a clinical benefit for this patient cohort, the mechanisms by which decitabine acts as an anti-cancer agent through perturbing DNA methylation remains poorly understood. In this research, we describe our approach using functional genomics and multiomics to examine the mechanisms by which decitabine acts to kill cancer cells in the context of AML. More specifically, our results unexpectedly reveal RNA dynamics as key regulators of DNA hypomethylation induced cell death in AML. Specifically, we show that RNA decapping quality control promotes cellular resistance to DNA hypomethylation, and conversely, we also observe that RNA methylation promotes cellular sensitivity to DNA hypomethylation. Overall, our findings linking RNA dynamics to DNA methylation suggests new levels of cellular integration between RNA and DNA regulatory biology that may aid in the design of future therapeutic strategies.

Autism Spectrum Disorders (ASD) are a set of neurodevelopmental disorders with complex biology. The identification of ASD risk genes from exome-wide association studies and de novo variation analyses has enabled mechanistic investigations into how ASD-risk genes alter development. Most functional genomics studies have focused on the role of these genes in neurons and neural progenitor cells. However, roles for ASD risk genes in other cell types are largely uncharacterized. There is evidence from postmortem tissue that microglia, the resident immune cells of the brain, appear activated in ASD. Here, we used CRISPRi-based functional genomics to systematically assess the impact of ASD risk gene knockdown on microglia activation and phagocytosis. We developed an iPSC-derived microglia-neuron coculture system and high-throughput flow cytometry readout for synaptic pruning to enable parallel CRISPRi-based screening of phagocytosis of beads, synaptosomes, and synaptic pruning. Our screen identified ADNP, a high-confidence ASD risk genes, as a modifier of microglial synaptic pruning. We found that microglia with ADNP loss have altered endocytic trafficking, remodeled proteomes, and increased motility in coculture.

Forelimb movements are abundant in our daily movements, and range from simple (swatting a bug) to sophisticated (catching a ball). Between higher and lower motor control centers sits the brainstem region lateral rostral medulla (latRM), which is required for the execution of sophisticated reaching movements. To date, latRM has been largely considered a relay between motor cortex and spinal cord, but questions remain whether these neurons participate in additional computations. In Chapter 1, I present a review of existing literature related to motor control, with particular focus on what is known about brainstem premotor regions, including latRM. In Chapter 2, I present work investigating additional, non-premotor features of latRM reach-related neurons in the context of a novel innate, touch-evoked reaching behavioral paradigm in mice. Experiments demonstrate that latRM neurons, previously implicated in skilled forelimb reaching behaviors, are also required for innate reaching behaviors and, unexpectedly, respond to orofacial somatosensory stimuli. Through a combination of electrophysiological recordings, anatomical tracing, and inhibition experiments we show that somatosensory-motor latRM neurons receive ascending somatosensory information as well as top-down inputs from both motor cortex and the midbrain structure superior colliculus. Together, these findings suggest additional functions for latRM by acting as a site of convergence for multiple streams of sensory and motor information, thus expanding our understanding of the organization of brainstem motor control circuitry.

Actin networks drive cell motility, which is important for essential processes such as embryonic development, wound healing, tissue remodeling, and the immune response. Actin regulatory proteins guide the self-organization of actin into force-generating structures like lamellipodia, filopodia, and ultimately drive cellular motility. Among these regulatory proteins is Capping Protein. Capping protein is a heterodimer that terminates actin filament elongation; it promotes actin network assembly, it competes with Nucleation Promoting Factors (NPF) to bind barbed ends, and is essential for the growth of polarized, force-generating actin networks. How Capping Protein performs these essential functions in the presence of multiple cellular inactivators, V1 and CARMIL, remains a mystery. V1 inactivates Capping Protein, while membrane-bound-CARMIL inhibits Capping Protein. Both the allosteric regulation of Capping Protein by CARMIL and steric inhibition by V1 have been studied from the atomic to the cellular scales. However, the question remains: if CARMIL and V1 inhibit Capping Protein, how does Capping Protein enter the actin network to give rise to cell motility? To answer this question, we used the bead motility assay, which reconstitutes branched actin networks in-vitro with purified components. The assembly of these branched actin networks requires the following key proteins: NPF, the Arp2/3 complex, Profilin, Actin, and Capping Protein. Here, we reconstituted branched actin networks with the addition of the Capping Protein regulators V1 and CARMIL. We immobilized CARMIL on the surface of microbeads to mimic its physiological role, and we used soluble V1 in molar excess with respect to Capping Protein. Our reconstitution assays show that CARMIL, in contradiction to previous work, can activate Capping Protein in the presence of V1 when attached to a surface. In addition, the CARMIL-mediated local activation of Capping Protein also regulates the growth rate and density of branched actin networks. Taken together, our results suggest that CARMIL acts at the cell membrane to activate and deliver Capping Protein to nearby actin filaments, thereby promoting the assembly of force-generating branched actin networks. We further explored the biochemical and biophysical properties of CARMIL, demonstrating its dimeric nature and high-affinity binding to Capping Protein. Using immunofluorescence and scanning electron microscopy, we examined the localization of CARMIL in mammalian cells and its association with purified endosomes. Additionally, we developed CRISPR knock-in strategies to visualize CARMIL endogenously. To complement our experimental findings, we sought agent-based and stochastic models to simulate the dynamic interactions between CARMIL, Capping Protein, and V1, revealing emergent properties in actin network regulation. These experimental and computational approaches offer insights into the self-organization of CARMIL and actin structures while also providing a framework for future investigations of molecular function to explain the emergence of cell motility.

Computational drug discovery is an active area of biophysics research due to the difficulties in estimating the free energy changes of small molecules binding to proteins. Considering the large search space of drug-like molecules, any estimates of binding must be fast as well as accurate. In this dissertation, I present new tools and models to improve computational drug discovery. The first project (Chapter 2) involves converting an expert-curated hierarchical database into a statistical potential for rapidly evaluating torsion strain in docked ligand poses. The second project concerns better understanding a new mode of ligand binding. New cryogenic electron microscopy (cryo-EM) structures of positron emission tomography (PET) radiotracers bound to protein fibrils show long, symmetric stacks of ligands within the fibrils. We present SymDOCK (Chapter 3) for accurately docking molecules to protein fibrils in symmetric, interacting stacks at a fast enough rate for large-scale docking. To better understand the effects on experimental observables from ligand-ligand interactions and entropy from the number of sites, we derive a new model for symmetric ligand binding to protein fibrils (Chapter 4). We lastly attempt to use our new tools and models to prospectively dock molecules against Alzheimer’s Disease tau fibrils (Chapter 5), where preliminary experimental results show some of our predicted hits do bind to the protein.

The immune system is complex, dynamic, and absolutely critical to maintain human health. The immune system comprises an enormous breadth of cell types, each able to respond to numerous extracellular environments. We have lacked adequate tools to dissect the mechanisms maintaining immune homeostasis through regulation of diverse cellular identities and stimulation responses. Human genetics has revealed key genes required to prevent immune dysfunction and many genomic loci associated with immune disease1,2. However, natural constraint, statistical power limitations, and lack of causal data hinder genetic association studies, preventing the identification of all key regulators and limiting conclusions about the relationships between regulatory genes 3–5. Experimental manipulation of gene expression with CRISPR provides the ability to investigate the structure and function of immune regulatory systems without the inherent limitations of association studies6.

In the T cell compartment, distinct lineages must respond to diverse signals to mount effective immune responses and maintain homeostasis, but the dynamic regulatory circuits that respond to extracellular cues in primary human cells remain poorly defined. To reveal the regulators of a core immune gene, IL2RA, expressed dynamically across the T cell compartment and required to prevent immune disease, we applied pooled CRISPR KO screens across cellular contexts7–10. We defined critical context specific regulators of IL2RA expression as well as regulators that affect the overall rest and activation state of the cell. One regulator in particular, MED12, coordinated gene regulatory networks required to maintain both T cell rest and activation. CRISPR ablation of MED12 blunted the cell state transitions between rest and activation and protected from activation-induced cell death, revealing a previously unappreciated gene regulatory mechanism governing T cell activity.

The effects of genetic variation on complex traits act mainly through changes in gene regulation. Although many genetic variants have been linked to target genes in cis, the trans-regulatory cascade mediating their effects remains largely uncharacterized. In a separate study, we investigated the function of and relationship between transcription factors associated with immune disease which are categorized as inborn errors of immunity (IEI) genes. We formed a large regulatory network consisting of regulators of IL2RA, IEI genes, and transcription factors without known immune disease associations1,11. These connections revealed shared paths and novel regulatory nodes that enable control over specific immune traits.

As a whole, this thesis work uses gene perturbation in distinct human T cell populations under different conditions to discover critical mechanisms regulating cell type and state specific gene expression. Detailed gene regulatory maps resulting from these studies, coupled with functional immune assays and biochemical data, provide new insights into human genetic variants linked to immune dysfunction and have potential to predict new modification strategies that enhance immunotherapies for the betterment of human health.

What do we study in neuroscience? Recent advances in our understanding of both non-neuronal cells within the nervous system and contemporary machine learning models necessitate a renewed focus on answering this foundational question in sufficient generality to be able to capture a wide array of phenomena whose deep resonances may escape more restrictive starting points. Here, I present two investigations that both tug at tensions within traditional conceptions.

First, I explore calcium activity in the astrocyte network, where tantalizing results have suggested the possibility that these neglected cells carry wide-reaching consequences for the functioning of brains and organisms. The results of these investigations reveal not only that astrocytes' bidirectional interactions with neurons carry a remarkable amount of structure in relation to a core physiological state (sleep), but also suggest that astrocytes may perform sophisticated spatial and temporal integration of their proximal neuronal inputs, implementing a form of combinatorial logic.

Second, I explore the structure of how language models respond when placed in relationship with highly-evocative naturalistic text. In the early stages of the explosion of interest around language models, a common argument levied against them was that, because they were trained to solve the task of next-token prediction (that is, of learning syntax), language models lacked actual understanding of the meaning of the tokens they were predicting (that is, of semantics). The results here—generating a rich and highly-structured portrait of human emotional life, through simple prompt-based impersonation and question-answering geared toward the practical task of hypothesis generation and symptom annotation in psychiatry—add to a growing body of work suggesting that, on the contrary, by solving the problem of generating language syntax, these models do (and, as implied by recent theoretical work, perhaps must) learn a latent structural image of semantics.

Finally, I draw the threads of these investigations together toward an answer of that foundational question: in neuroscience, we study the dynamics of interacting world models. I sketch a framework, the postmodern synthesis, of how the mechanics of these changing minds---when viewed as consisting solely of their relational structure---might be rigorously modeled using the mathematical language of double categories. This framework touches formal resonances with both quantum field theory and classical genetics. By framing our science the other way 'round, from the perspective of the mental content, rather than its material implementation, we may be able to ask new questions.