Purpose: Atherosclerosis is characterized by lipid accumulation and inflammatory cell infiltration in large arteries, promoting monocyte adhesion and migration into the arterial intima. Liver X receptors (LXRs) regulate cholesterol homeostasis and inflammation, and recent findings from our laboratory demonstrate that LXR activation in endothelial progenitor cells (EPCs) promotes the secretion of factors that reduce monocyte–endothelial adhesion, conferring protection against atherosclerosis. Soluble triggering receptor expressed on myeloid cells 2 (sTREM2) was identified as a key mediator of this protective effect. We hypothesized that sTREM2 interacts with specific endothelial cell (EC) surface receptors to modulate the EC phenotype, enhance endothelial barrier integrity, and reduce monocyte adhesion. It has been reported that sTREM2 can interact with transgelin-2 (TAGLN2) in the brain, suggesting it is a potential receptor. The goal of this project was to assess the mechanism by which sTREM2 prevents monocyte interaction with endothelial cells.

Methods: Human umbilical vein endothelial cells (HUVECs) were transfected with siRNA targeting TAGLN2 or a non-targeting control and treated with sTREM2 in the presence or absence of TNFα, followed by assessment of monocyte–endothelial adhesion using fluorescently labelled THP-1 monocytes. In parallel, sTREM2-associated proteins were identified by affinity-based pulldown and shotgun proteomics, and endothelial barrier integrity was assessed by transendothelial electrical resistance (TEER) in human coronary artery endothelial cells (HCAECs).

Results: Under inflammatory conditions (TNFα, 10 ng/mL), there is a 5-fold increase in the binding of monocytes to HUVECs compared to control. With the addition of sTREM2, monocyte adhesion was decreased by ~50% compared to the +TNFα control. This protective effect of sTREM2 was abolished in the siTAGLN2 treated HUVECs. Additionally, sTREM2 enhanced endothelial barrier integrity, as evidenced by increased TEER under TNFα stimulation. Proteomic analysis revealed enrichment of RhoA signalling pathways and cadherin-binding proteins among sTREM2-associated proteins, implicating cytoskeletal regulation and junctional stabilization as key mechanisms. Notably, TAGLN2 was not detected in the sTREM2 pulldown.

Conclusions: Together, these data demonstrate that sTREM2 suppresses endothelial activation and monocyte adhesion while enhancing barrier integrity under inflammatory conditions in a TAGLN2-dependent manner, highlighting sTREM2 as a promising therapeutic target for preserving endothelial function in atherosclerosis.

Acute lung injury, and its most severe manifestation, acute respiratory distress syndrome (ARDS), is a common diagnosis in critical care units and carries a mortality rate approaching 40%. ARDS can arise from diverse etiologies, including pathogen-induced pneumonia, aspiration, and inhalational injury. Pathophysiologically, ARDS is characterized by disruption of the alveolar–capillary barrier, leading to increased vascular permeability, alveolar flooding, and severe hypoxemic respiratory failure. Therapeutic strategies that restore or strengthen the lung endothelial barrier may therefore be critical to limiting disease progression and preventing clinical deterioration.

Lipid nanoparticles (LNPs) have emerged as a safe and effective platform for mRNA therapeutics, enabling transient, endogenous production of targeted proteins. In this study, mRNA constructs encoding reporter proteins or endothelial tight junction proteins were generated via in vitro transcription from linearized T7 promoter plasmids, incorporating co-transcriptional capping, poly(A) tailing, and chemically modified uridine nucleotides. Using microfluidic mixing, a library of LNP formulations was produced and systematically screened for in vitro transfection efficiency, cytotoxicity, and endosomal escape.

In vivo biodistribution was evaluated using firefly luciferase mRNA in both healthy and ARDS mouse models, while localized protein expression was assessed through immunostaining. We identified an optimized LNP formulation that achieved greater than 85% lung specificity with robust protein expression following intravenous administration. This formulation demonstrated a strong safety profile across multiple concentrations and cell lines, with no detectable toxicity observed by histological analysis or oxygen saturation monitoring. Importantly, lung-specific protein expression was preserved across healthy and diseased mice at different stages of ARDS, with no significant differences observed between male and female animals. Immunostaining confirmed preferential expression within lung endothelial cells.

Based on these findings, we are advancing this lung-targeted LNP platform to deliver therapeutic mRNA encoding endothelial tight junction proteins. Ongoing studies are focused on evaluating safety and efficacy in larger animal models to support translational development. This work highlights a lung-selective mRNA delivery strategy with potential application in ARDS, offering a safe and selective approach to restoring lung barrier function. Beyond ARDS, this platform highlights the broader potential of mRNA therapeutics, already validated by FDA-approved vaccines, to address critical unmet needs in pulmonary medicine.

MicroRNAs (miRNAs) are short, non-coding RNAs that regulate gene expression and have emerged as promising biomarkers for several diseases, including sepsis and fitness prediction. Detecting multiple miRNAs simultaneously is essential for accurately identifying biomarker profiles for complex conditions. Furthermore, these applications would benefit from point-of-care miRNA detection to enable timely clinical intervention. CRISPR-Cas12 systems offer fast, programmable, and sensitive detection of nucleic acids that can be integrated into portable formats. However, challenges remain in achieving high specificity for direct RNA detection and simultaneous detection of multiple targets in a single reaction. Our group has developed RAPID; a platform that enables sensitive, PAM-free detection of miRNAs using AsCas12a. RAPID uses a split-activator design and a novel reporter sequence to detect synthetic miRNA biomarkers in under 30 minutes, with a detection limit of 97 pM and single-nucleotide specificity. Here, I aim to expand the RAPID platform towards multiplexed miRNA detection and high-fidelity detection of single nucleotide polymorphisms (SNPs). Towards the aim of supporting multiplexing, I purified and screened nine Cas12a enzymes from different bacterial species to assess their trans-cleavage activity on diverse reporter sequences. Cas12a variants possessed distinct substrate preferences, which we plan to leverage for simultaneous detection of multiple targets by further reporter discovery using next generation sequencing and reaction optimization of orthogonal enzyme-reporter pairs. Towards the aim of SNP detection, I screened guide RNA designed with modified nucleic acids to reduce the off target activation of synthetic miR-21 targets containing single point mutations. While the mutation detection is typically dependent on mutation position, I determined the optimal position and number of DNA modifications required to improve the specificity across all microRNA target positions. These modified guides are promising for highly specific detection of SNP mutations. By combining high specificity with the potential for multiplexed detection, we envision a platform that supports comprehensive disease profiling from a single sample, enhancing both diagnostic precision and the speed of clinical decision-making.

Developing improved predictive models for multi-molecular systems is crucial, as nearly every chemical product used results from a mixture of chemicals. While being a vital part of the industry pipeline, the chemical mixture space remains relatively unexplored by the Machine Learning community. In this paper, we introduce CheMixHub, a holistic benchmark for molecular mixtures, covering a corpus of 11 chemical mixtures property prediction tasks, from drug delivery formulations to battery electrolytes, totalling approximately 500k data points gathered and curated from 7 publicly available datasets. CheMixHub introduces various data splitting techniques to assess context-specific generalization and model robustness, providing a foundation for the development of predictive models for chemical mixture properties. Furthermore, we map out the modelling space of deep learning models for chemical mixtures, establishing initial benchmarks for the community. This dataset has the potential to accelerate chemical mixture development, encompassing reformulation, optimization, and discovery.

Extracellular vesicles (EVs) carry disease-specific molecular signatures, making them promising candidates for non-invasive cancer diagnostics. However, the sensitive and reproducible detection of EV-associated biomolecules presents significant technical challenges. In this study, we investigated the performance of random silver nano-island surface for detecting cholangiocarcinoma-derived EVs using surface-enhanced Raman spectroscopy (SERS), with synthetic liposomes serving as a lipid-only control group.
EVs were isolated from the KKU-23 cholangiocarcinoma cell line, and liposomes with a defined lipid composition were deposited (using 2 µL of the samples) onto random Ag nano-island surface. These samples were then analyzed using excitation wavelengths of 532, 633, and 785 nm. The random Ag nano-island SERS platform provided significantly enhanced Raman signal intensity, improved signal-to-noise ratios, and richer spectral features. Importantly, clear spectral differences were observed between the EVs and liposome controls, even at nanomolar concentrations. Using iterative multi-pass peak deconvolution, we identified a total of 36 Raman peaks associated with the EV samples, of which seven peaks were unique to the EVs and not present in the liposome controls. This indicates that molecular contributions extend beyond lipid membranes, likely arising from proteins and surface-associated biomolecules enriched in the EVs. In contrast, the spectra from liposomes were primarily dominated by lipid-associated vibrational modes, showing fewer unique features.
These findings demonstrate that a random Ag nano-island SERS platform enables sensitive and label-free differentiation between extracellular vesicles and lipid-only controls, highlighting their potential for detecting cancer biomarkers based on EVs and for use in translational diagnostics.

All cells sense and respond to their environment through membrane receptors. To help facilitate this process, ligand-bound receptors often group into nanoscale "microenvironments" that act as hubs for cell signaling and decision-making. Current techniques to study these dynamic assemblies leverage mass spectrometry-based proximity labeling proteomics (PLP), enabling unbiased profiling of receptors within the microenvironment. While these platforms have been foundational in mapping receptor interaction networks across many biological systems, existing PLP can only be applied to single ligand-receptor interactions, preventing researchers from studying biology that emerges under simultaneous ligand-receptor engagements.
To address this, we developed a PLP tool that uses DNA origami, a method of folding DNA into precise shapes, to act as a molecular "pegboard." This enables plug-and-play presentation of signaling molecules with controlled valency and nanometer precision. On this DNA pegboard, we installed light activated Eosin Y as our labeling tool, alongside T cell specific ligands that target T cells as a model system. By using this DNA pegboard to engage T cells, light irradiation allows us to map the immediate protein microenvironment surrounding the pegboard. Our preliminary data demonstrates that we can successfully perform this labeling in live T cells, and that the process is both light- and ligand-specific. Together, this work establishes a useful platform to systematically map and compare receptor microenvironments across different multivalent ligand systems, with the potential to discover proteins with previously uncharacterized biological function, including proteins that may represent druggable targets.

Background: Breast cancer is one of the most common types of cancer, with MYC overexpression occurring in 40-60% of triple-negative breast cancer (TNBC) cases. MYC overexpression drives tumorigenesis by promoting proliferation and inhibiting apoptosis. Consequently, disruption of this dysregulated network is a promising avenue for treatment. We developed MEF and MEF/C93, protein-based MYC inhibitors, based on the binding domain of Max. Our proteins function as homodimers and competitively bind to E-box, suppressing Myc transcription. MEF and MEF/C93 have shown selective and high-affinity E-box binding (Kd ~ 8nM).

Aim: Our objective was to evaluate the efficacy and targeted effects of these proteins in breast cancer cell lines expressing either increased (MDA-MB-231, MDA-MB-468) or basal (MCF-7) levels of Myc. Initial in vivo testing in healthy immunocompetent mice assessed tolerable doses and routes, and initial estimate of half-life.

Methods: IC50s were determined via MTT assay in cell lines treated with 0.01-25 μM of MEF, MEF/C93, or control protein (Sc-5) for 24 or 48 hours. MYC protein expression was measured in cell lines via western blot immunodetection. Target engagement was evaluated through qPCR analysis of MYC-regulated genes with treated cells. CD-1 mice were administered escalating doses of MEF, MEF/C93 or control protein via intravenous or subcutaneous routes. Animals were monitored daily for changes in body weight, behaviour, and physical well-being. After the sacrifice, histopathological analysis and complete blood counts were completed.

Results: Cytotoxicity studies demonstrated IC50 values of 1-2 μM in Myc overexpressing cells, which were over 10-fold lower than those seen in MCF-7 cells. MEF and MEF/C93 significantly decreased, and levels of MYC target genes in MDA-MB-231 cells. Immunofluorescence demonstrated nuclear localization of MEF within 6 hours. In vivo administration of 1 – 10 mg/kg IV dose of MEF or MEF/C93 to CD-1 mice was well tolerated with no significant impact on weight, tissue histology, or complete blood counts. A plasma half-life of 13 hours was seen for MEF in a pilot pharmacokinetic study, indicating relatively extended circulation time.

Conclusion: These findings support MEF and MEF/C93 as promising MYC-targeted protein therapeutics with strong preclinical efficacy and tolerability, offering potential for treating MYC-driven cancers.

Three-dimensional microvascular vessel network (MVN) is a highly physiologically relevant tool for modeling vasculature, with numerous applications in vascularized tissue engineering, vascular pathology modeling, studies of the vascularized tumour microenvironment, and drug delivery and screening. Typical perfusable MVNs are created by endothelial cells self-assembly in 3D hydrogel with supporting cells (i.e. fibroblast), angiogenic factors and mechanical cues (i.e. perfusion flow). To facilitate protocol selection, we interrogated the effect of fibroblast concentration, fibroblast conditioned media, angiogenic factors, and luminal flow on MVN morphology, its functional perfusability and vessel wall integrity. To improve accuracy of the quantitative morphology analysis, we quantified the morphology of the empty spaces between vessel branches using novel void-based metrics. The void-based metrics captured the subtle morphological features of a wider range of MVN protocols with more accuracy compared to the traditional metrics. Overall, we found that high fibroblast concentration increased MVN growth rate and lead to excessive vessel fusion. MVNs cultured without fibroblast using only fibroblast conditioned media or angiogenic factors developed patch-like morphology instead of the physiologically relevant thin-branch morphology. Therefore, fibroblast in the MVN culture system is crucial for MVN development and cannot be substituted by its conditioned media or angiogenic factors. Our novel void-based analysis approach for MVN morphology quantification also offers a more efficient alternative to traditional branch-based methods, providing higher information density per metric and reduced data volume for machine learning and AI-assisted image analysis. The MVN culture in this work has been adapted into a 3D microfluidic tumour model for studying tumour-vasculature interaction.

Adult neurogenesis is restricted to specific regions, leaving the brain with limited capacity to repair itself after stroke-induced neuronal loss. Recently, cellular reprogramming of resident astrocytes into new neurons has emerged as a promising therapeutic strategy to treat stroke by replacing dead neurons. This approach utilizes adeno-associated viruses (AAV) to express proneural factors in astrocytes. Recently, we developed a designer version of the reprogramming factor Ascl1, ASCL1(SA6), which reprograms astrocytes into neurons with a higher potency in the uninjured mouse cortex than the native protein. However, AAV targeting of astrocytes is imperfect, and off-target expression in neurons raises important questions about whether the functional recovery observed with Ascl1 treatment is genuinely driven by neurogenesis or whether ectopic expression of reprogramming factors in pre-existing neurons is the cause. We hypothesised that cellular reprogramming is required to drive functional recovery following a stroke, and that neuronal expression of Ascl1 is insufficient to promote recovery. To address these questions, we employed the inducible astrocyte-specific Aldh1l1-CreERT2 Cre driver mice crossed to Ai14 ROSA-tdTomatoFlx reporter mice to conduct stringent lineage tracing studies of cellular reprogramming. We observed that the designer reprogramming factor ASCl1(SA6) induces astrocytes-to-neuron reprogramming in the uninjured adult mouse cortex (p = 0.0004 compared to EmGFP controls). Next, using the endothelin-1 ischemic stroke mouse model, we observed that reprogramming promotes functional motor recovery one month after stroke (p < 0.05 for the grip strength, foot fault, and cylinder tasks compared with EmGFP controls). To test if off-target neuronal expression of reprogramming factors influences behavioural recovery, we directed Ascl1 expression to neurons in the stroke-injured mouse cortex using the neuron-specific human synapsin (hSyn) promoter. However, we observed no differences in stroke recovery between neuron-directed Ascl1 and EmGFP controls (no significant differences in the grip strength, foot fault, and cylinder tasks). In conclusion, our studies support the hypothesis that cellular reprogramming, rather than ectopic expression of reprogramming factors in neurons, enhances functional recovery following ischemic stroke.

Adult neurogenesis is restricted to specific regions, leaving the brain with limited capacity to repair itself after stroke-induced neuronal loss. Recently, cellular reprogramming of resident astrocytes into new neurons has emerged as a promising therapeutic strategy to treat stroke by replacing dead neurons. This approach utilizes adeno-associated viruses (AAV) to express proneural factors in astrocytes. Recently, we developed a designer version of the reprogramming factor Ascl1, ASCL1(SA6), which reprograms astrocytes into neurons with a higher potency in the uninjured mouse cortex than the native protein. However, AAV targeting of astrocytes is imperfect, and off-target expression in neurons raises important questions about whether the functional recovery observed with Ascl1 treatment is genuinely driven by neurogenesis or whether ectopic expression of reprogramming factors in pre-existing neurons is the cause. We hypothesised that cellular reprogramming is required to drive functional recovery following a stroke, and that neuronal expression of Ascl1 is insufficient to promote recovery. To address these questions, we employed the inducible astrocyte-specific Aldh1l1-CreERT2 Cre driver mice crossed to Ai14 ROSA-tdTomatoFlx reporter mice to conduct stringent lineage tracing studies of cellular reprogramming. We observed that the designer reprogramming factor ASCl1(SA6) induces astrocytes-to-neuron reprogramming in the uninjured adult mouse cortex (p = 0.0004 compared to EmGFP controls). Next, using the endothelin-1 ischemic stroke mouse model, we observed that reprogramming promotes functional motor recovery one month after stroke (p < 0.05 for the grip strength, foot fault, and cylinder tasks compared with EmGFP controls). To test if off-target neuronal expression of reprogramming factors influences behavioural recovery, we directed Ascl1 expression to neurons in the stroke-injured mouse cortex using the neuron-specific human synapsin (hSyn) promoter. However, we observed no differences in stroke recovery between neuron-directed Ascl1 and EmGFP controls (no significant differences in the grip strength, foot fault, and cylinder tasks). In conclusion, our studies support the hypothesis that cellular reprogramming, rather than ectopic expression of reprogramming factors in neurons, enhances functional recovery following ischemic stroke.

Over 60% of primary prostate cancer patients harbor dormant cancer cells (DCCs) that have disseminated to bone marrow (BM). DCCs enter a reversible, non-dividing state which enables metastatic recurrence. Studies have identified how cells within the BM perivascular niche regulate quiescence or reactivation of DCCs. However, the role of the mechanical environment in regulating prostate cancer cell quiescence in BM remains undefined. The aim of this study is to investigate how osteocytes, mechanosensitive cells in bone, regulate prostate cancer cell quiescence under mechanical stimulation, and how changes to the BM extracellular matrix (ECM) impact cancer cell growth.

BM mesenchymal stem cells (MSCs) and human umbilical vein endothelial cells (ECs) were mixed at a 5:1 ratio and seeded into 96-well-plates to induce organotypic vessel formation over 10 days. 400 PC-3 prostate cancer cells were seeded into each well on either Days 3, 7 or 10 during vessel formation. A laminin-rich hydrogel was introduced to co-cultures at either 2 or 5mg/ml to modulate ECM protein concentration. Alginate at either 0.25 or 0.5% was incorporated with the hydrogels to modulate stiffness. Conditioned media collected from either static or oscillatory fluid flow (OFF)-stimulated MLO-Y4 osteocytes (1 Pa, 1 Hz, 2 hours) were added to co-cultures to determine effects of osteocytes on prostate cancer cell quiescence. Prostate cancer cells were co-cultured in BM microenvironments for an additional 10 days.

Organotypic vessels formed 10 days after seeding. OFF-stimulated osteocytes supported the formation of microvasculature. In 2mg/ml hydrogels, seeding PC-3 cells on Day 3 resulted in elevated outgrowth compared to later timepoints. Of interest, PC-3 outgrowth was elevated in 5mg/ml hydrogels across all timepoints, however OFF-stimulated osteocytes exhibited an inhibitory effect when PC-3 cells were seeded on Days 7 or 10. Addition of alginate to co-cultures resulted in similar trends observed in 5mg/ml hydrogels, indicating that PC-3 outgrowth within the BM perivascular niche is dependent on substrate stiffness.
We anticipate this study will provide novel insight into the role of the mechanical environment on cancer cell quiescence in BM, and motivate further studies on how mechanical alterations to the BM matrix may influence DCCs.

Background: Growing evidence has identified vascular dysfunction as an early event in Alzheimer’s Disease (AD), preceding pathological hallmarks such as amyloid deposition. However, the mechanisms underlying vascular pathology in AD have yet to be fully understood, in part, due to the lack of physiologically relevant models of the cerebral vasculature. To address this unmet need, we applied a novel microfluidic blood-brain barrier model to discover mechanisms of vascular dysfunction in AD and demonstrate utility of the platform to uncover novel therapeutic targets.

Methods: Stem cells derived from a patient with familial AD and a control line (fControl) were differentiated into brain endothelial-like cells (BECs) and neurons. BECs were cultured statically or exposed to 12 dynes/cm2 of shear stress for 72 hours prior to functional assessment of BEC marker expression, efflux transporter function, and monocyte adhesion assays. Proteomics and secretome evaluation were conducted in BEC-neuron cocultures to identify paracrine mediators of vascular dysfunction.

Results: BECs from the donor with AD (AD-BECs) demonstrated impaired efflux transport by p-glycoprotein, breast cancer resistant protein, and multidrug resistant protein compared to controls (fControl-BECs, p = 0.0015, p = 0.0004, p = 0.0002, respectively). Under shear stress, no differences were observed suggesting physiological shear maintains efflux transport function. AD-BECs exhibited increased monocyte adhesion (1.9-fold; p<0.01) which was reduced under shear stress in both lines (AD: p<0.001; fControl: p<0.01). Proteomic assessment revealed differential protein expression in BECs and neurons related to mitochondrial dysfunction and oxidative phosphorylation in both static and shear AD models. Secretome analysis identified shear sensitive and potentially protective candidates, such as insulin degrading enzyme (IDE) and ephrin A receptor.

Conclusion: We report the lack of shear stress (reflective of hypoperfusion) reveals BBB dysfunction, specifically identifying altered BEC efflux transport function, immune interactions, proteome, and secretome in a model of the cerebral vasculature in AD. The identification of potentially protective factors upregulated by shear stress suggests novel mechanisms by which hypoperfusion could lead to loss of protection, opening new therapeutic avenues to combat vascular dysfunction in neurodegeneration.

Protein-based therapeutics have led to new advances in disease treatment. Myc is a transcription factor that is overexpressed and "goes rogue" in over 70% of known cancers and tumours. As a heterodimer with Max, Myc binds cooperatively to the Enhancer-Box (E-box) DNA site, resulting in cell proliferation and suppression of differentiation. We designed MEF to competitively inhibit Myc/Max binding to the E-box site, interfering with Myc’s tumour promotion. MEF mimics the basic region/helix-loop-helix (bHLHZ) domain of Myc/Max to bind to the E-box site. We are re-designing MEF variants to be more stable and long lived in the body. Our quantitative PCR studies revealed that MEF and its variant MEF/C93 selectively downregulated Myc target genes in Myc-dependent MDA-MB-231 breast cancer cells, but not in Myc-independent MCF-7 cells. Fluorescence colocalization results have demonstrated transport of our proteins into cell nuclei. MEF and MEF/C93 have displayed IC50 values of 1−2 μM in cell viability assays in MDA-MB-231, compared with∼25 μM in MCF-7. These results demonstrate the specificity of targeting Myc-dependent cancer cells and mark significant progress toward protein-based therapies aimed at Myc-driven cancers.

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disorder characterized by the progressive loss of motor neuron function. ALS is a multifactorial disease which can originate from complex genetic and environmental factors. The identification of risk factors and their molecular contribution to ALS expand our understanding of the disease mechanism.

Here, we describe a family with dominantly inherited degeneration, which carries a mutation in the serine/arginine repetitive matrix 2 gene (SRRM2). SRRM2 is essential for nuclear speckle formation and a constitutive member of the RNA splicing machinery. To investigate how the mutation in SRRM2 contributed to the ALS pathogenesis, we examined its effect on a model cell line, where the point mutation was introduced in the endogenous gene. Surprisingly, we found that the resulting single amino acid exchange led to the loss of one protein-protein interaction, between SRRM2 and the splicing factor ACIN1. Transcriptome studies further revealed wide-spread differential gene expression, which converged on the dysregulation of synapse-associated pathways. Together, our findings identify SRRM2 as a novel ALS risk factor and provide mechanistic insights into how its mutation can be linked to ALS pathology.  

Genome-scale metabolic models (GEMs) are a cornerstone for simulating cellular metabolism, but their predictive power is limited by the absence of enzyme kinetics. Incorporating enzyme constraints to these models (ecGEMs) narrows the solution space and links molecular parameters to organismal phenotypes. However, the lack of experimentally measured kinetic constants (e.g., kcat) severely restricts scalability and transferability.
We introduce an integrated machine learning and modeling framework that balances precision and accuracy in ecGEMs. Our approach combines CPI-Pred, a deep learning model that predicts kinetic parameters from protein language model embeddings and compound representations, with kinGEMs, a pipeline that integrates these predictions into GEMs and refines the constraints through simulated annealing.
We evaluate the framework across three axes:

Cross-organism generalization: The same pipeline builds models that are more precise for a variety of organisms.
Together, these results demonstrate that ML-predicted kinetic parameters tuned in-vivo to match cellular contexts can systematically improve both the internal precision of metabolic models and their external accuracy in predicting phenotypes. Ongoing work scales the framework to the AGORA microbiome resource, enabling large-scale, interpretable simulations of microbial communities and perturbation studies. The integration of modern ML with mechanistic modeling offers a path toward more precise and accurate ecGEMs, broadening their impact in systems biology, metabolic engineering, and synthetic biology.

Precision (flux variability analysis): Incorporating CPI-Pred predictions reduces median flux variability compared to unconstrained GEMs, yielding more defined and interpretable solution spaces.

Accuracy (E. coli genetic and substrate perturbations): Using RB-TnSeq data, kinGEMs improves gene lethality prediction accuracy and ROC-AUC compared to baseline GEMs.

Huntington’s disease (HD) is an inherited neurodegenerative disease with progressive development of motor, cognitive, and psychiatric symptoms. It is caused by a trinucleotide repeat expansion of CAG above a threshold of 35 repeats in the gene encoding the huntingtin (HTT) protein, resulting in an expansion of the HTT polyglutamine tract. Despite over 30 years having passed since the discovery of this mutation, HTT biology and the exact mechanism of HD pathogenesis remains incompletely understood. There remains a need to develop high-quality chemical tools to target, detect, and modulate HTT. Here, we report the identification and characterization of a small molecule ligand for the soluble, full-length HTT-HAP40 protein complex, the most abundant endogenous proteoform of HTT. Through a high-throughput screening method, affinity selection mass spectrometry (AS-MS), we identified a ligand with enantioselective binding which was biophysically validated using surface plasmon resonance (SPR). This ligand has stoichiometric (1:1) binding to HTT-HAP40 in vitro with single-digit micromolar affinity, shows specificity for the full HTT-HAP40 complex when tested with various HTT subdomains and unrelated proteins, and is independent of polyglutamine tract expansion. Hydrogen-deuterium exchange mass spectrometry (HDX-MS) revealed the binding location at an interface between HTT and HAP40, and the ligand-bound complex was visualized by cryo-electron microscopy (cryo-EM) with 2.3 Å resolution. This visualization, along with computational free energy perturbation (FEP) studies, revealed pi-pi interactions to be the largest contributor of binding affinity and explained the mechanism of enantioselectivity for the (R)-enantiomer. Preliminary structure-activity relationship (SAR) was established through a selection of chemical analogs which revealed potential avenues to improve binding affinity and demonstrated the tractability of this ligand for downstream chemical synthesis applications. This novel chemical scaffold serves as a starting point for the development of direct HTT-targeting chemical tools (e.g. tracers) and pharmacological modulators (e.g. small-molecule degraders).

Introduction. Brain arteriovenous malformation (bAVM) is a devastating disease and a leading cause of hemorrhagic stroke in young people due to the formation of abnormally tortuous blood vessels within the brain. To date, there is limited knowledge regarding the molecular mechanisms driving bAVM pathogenesis, and consequently, no approved pharmacological treatments are available. Through whole-exome sequencing, our group identified that >50% of bAVM patients carry somatic activating KRAS gene mutations (p.G12V, p.G12D, p.Q61H) in their endothelial cells. Subsequent studies have detected additional low-frequency variants (e.g. p.G12C) and mutations in other genes within the same signaling cascade (e.g. BRAF p.V600E). These discoveries stimulated an emergence of clinical studies exploring the therapeutic potential of inhibitors targeting mutant KRAS activity and downstream signaling pathways. While preliminary results are encouraging (i.e. some symptomatic improvements and disease stability), patient response to these inhibitors is highly variable, and the most suitable therapeutic avenue for long-term bAVM management is not clear.

Hypothesis. We hypothesize that the clinical heterogeneity in patient response to treatment is influenced by the underlying KRAS mutations and that a precision medicine approach is necessary for bAVM therapeutics.

Methods/Results. In collaboration with Drs. Ann Mansur, Ivan Radovanovic (neurosurgeons/UHN) and Milica Radisic (biomedical engineer/UofT), we are developing a novel three-dimensional vasculature model using patient-derived endothelial and perivascular cells. We confirmed the KRAS mutations in collected clinical samples using ddPCR and characterized the identity/purity of isolated cells through immunofluorescence staining of endothelial, mural and neuronal markers. These vascular networks are further supported by a mesenchymal stem cell population that is capable of neurogenesis and differentiation to more closely recapitulate a neuronal microenvironment. Utilizing this model, we are performing a high content screening study with a custom drug library (consists of inhibitors currently under preclinical testing for bAVMs, novel KRAS inhibitors from Dr. Igor Stagljar, and novel kinase inhibitors identified from our completed screening study using endothelial monolayers). We will assess these inhibitors with vascular networks generated from patient cells carrying different KRAS mutations. Our study will provide novel insights regarding mutation-specific sensitivity towards therapeutic candidates for bAVMs to better inform clinical decision-making and future treatment strategies.

Huntington’s disease (HD) is a fatal, autosomal dominant neurodegenerative disorder caused by expansion of CAG repeats in the HTT gene. Although CAG repeat length is a major determinant of disease onset, it does not fully explain the wide variability in clinical progression observed among patients. This has led to growing interest in genetic modifiers that influence disease severity and timing independently of the primary mutation. Genome‑wide association studies have identified multiple such modifiers, yet most remain largely unexplored at the functional level, limiting their translational impact.

One of the strongest yet understudied modifiers emerging from large patient cohorts is CCDC82, a gene associated with delayed functional decline in HD but notably independent of somatic CAG repeat expansion. This observation suggests that CCDC82 modulates disease progression through a mechanism distinct from canonical DNA repair pathways that dominate current HD research. Despite its genetic significance, little is known about the biological role of CCDC82, particularly in the context of the nervous system.

In this work, we aim to establish a foundational understanding of CCDC82 as a novel factor influencing neuronal health and disease resilience. By integrating structural biology, proteomics, and human cell-based approaches, we are systematically characterizing CCDC82 at the molecular and cellular levels. Our ongoing studies focus on defining its biochemical properties, cellular localization, and interaction networks, as well as its regulation in human cells carrying pathogenic HTT expansions.

By shifting attention from widely studied DNA repair modifiers to a previously uncharacterized protein linked to clinical progression, this work highlights an alternative framework for understanding HD pathogenesis, one centered on neuronal resilience rather than mutation burden alone. Elucidating the function of CCDC82 may reveal new biological pathways underlying selective neurodegeneration and open opportunities for therapeutic strategies that complement current HTT‑lowering approaches. More broadly, this study illustrates how systematic functional dissection of GWAS‑identified modifiers can uncover overlooked mechanisms relevant to neurodegenerative disease.

The emergence of Targeted Protein Degradation (TPD) has revolutionized the approach to "undruggable" proteins. While traditional small-molecule inhibitors are limited by stoichiometric binding, PROteolysis TArgeting Chimeras (PROTACs) act catalytically to hijack the native Ubiquitin-Proteasome System (UPS). However, the clinical translation of PROTACs is often hindered by poor solubility, limited cell permeability, and the high-dose requirements of low-affinity bifunctional molecules.

This work proposes a novel Nano-Linked PROTAC system designed to degrade a target protein of interest (POI). By tethering bifunctional degraders comprising a POI-specific ligand and a known E3 ligase recruiter (e.g., CRBN or VHL) onto a biocompatible nanoparticle scaffold, we aim to achieve a high local density of recruiters. This multivalent architecture is designed to facilitate the rapid formation of the ternary complex (Target-PROTAC-E3 ligase) within the cytoplasm, theoretically lowering the threshold for robust polyubiquitination and subsequent proteasomal degradation. We hypothesize that this nano-formulation will significantly improve the pharmacokinetic profile of the PROTAC molecules, overcoming the physical limitations of standalone small molecules. Furthermore, this platform provides a modular "plug-and-play" architecture in which the nanoparticle surface can be functionalized for specific cell-type targeting. This integrated approach seeks to provide a scalable, precision-medicine strategy to eliminate oncogenic proteins, which are often resistant to conventional stoichiometric inhibition.

Introduction/Significance: Retinal degenerative diseases are characterized by progressive degeneration leading to blindness, brought on by genetic predisposition, environmental factors, and advanced age. Insulin-like growth factor 1 (IGF-1) exerts neuroprotective effects within the central nervous system, including the retina. However, barriers to its clinical translation include poor pharmacokinetics, stability and tissue penetration, and off-target effects. As such, sustained local delivery of IGF-1 using hydrogels is a promising treatment strategy to suppress retinal cell death, preserve retinal function, and reduce the frequency of invasive intravitreal injections. We developed a biocompatible hydrogel with properties comparable to the vitreous humour that is an ideal vehicle for the affinity-based sustained release of IGF-1 to the retina. We determined the efficacy of this controlled release system in preserving photoreceptor survival and retinal neuron function in the rd10 mouse model of retinal degeneration.

Methodology: Ketone-modified hyaluronan (HAK) polymers were immobilized with a weak binding peptide (WBP) that has sub-micromolar affinity (Kd = 2.7×10−5 M) for the SH3 domain of SH3-IGF-1 fusion protein. The polymer was covalently crosslinked using oxime chemistry to form HAK-WBP¬, the hydrogel delivery vehicle for affinity-based release of SH3-IGF-1. We used the rd10 mouse model of retinal degeneration to evaluate the efficacy of the SH3-IGF-1 loaded HAK-WBP (I-SR) system. I-SR was intravitreally injected at post-natal day 18-20 (onset of photoreceptor degeneration). We used histology/immunohistochemistry to assess retinal cell death and thickness of the outer nuclear layer (ONL), and electroretinography to assess functional preservation of retinal neurons at 2- and 3-weeks post-injection respectively.

Results: In vitro sustained release of SH3-IGF-1 from HAK-WBP occurred over 7 days. Mice injected with I-SR showed preserved ONL thickness compared to untreated controls and SH3-IGF-1 bolus at 7 days post-injection. I-SR suppressed cell death in the ONL compared to the untreated and SH3-IGF-1 bolus arms. Functional responses of photoreceptors and bipolar cells were improved in the I-SR arm. Future work will assess the mechanisms of IGF-1 on retinal neurons and advance the delivery strategy.

Molecular fingerprinting methods use hash functions to create fixed-length vector representations of molecules. However, hash collisions cause distinct substructures to be represented with the same feature, leading to overestimates in molecular similarity calculations. We investigate whether using exact fingerprints improves accuracy compared to standard compressed fingerprints in molecular property prediction and Bayesian optimization where the underlying predictive model is a Gaussian process. We find that using exact fingerprints yields a small yet consistent improvement in predictive accuracy on five molecular property prediction benchmarks from the DOCKSTRING dataset. However, these gains did not translate to significant improvements in Bayesian optimization performance.

Colloidal drug aggregates are a promising drug delivery platform due to their ultra-high drug loading. However, stabilizers are usually required in physiological conditions, and only a limited number of colloid aggregators have been successfully formulated. This study uses computationally guided drug modification to stabilize nintedanib colloidal aggregates for sustained delivery. To target fibrotic scarring associated with spinal cord injury repair, we propose an in-situ depot for localized, controlled antifibrotic release while minimizing off-target toxicity.

This study employs a random forest-based colloidal aggregates model to design nintedanib analogs as colloidal aggregates with hydrophobic, aromatic, or electronegative nonpharmacological cores. Analogs were synthesized using a robust three-step synthesis approach. These analogs self-aggregate into amorphous nanoparticles detected via dynamic light scattering. Stability of formed nanoparticles was assessed by screening various stabilizers, including small molecule co-aggregators, proteins, and lipids, for size distribution over 7 days in vitro. Antifibrotic activity was evaluated in vitro using a mouse embryonic fibroblast model, and stable colloidal aggregates were embedded in a hydrogel to assess their sustained release profile over 7 days.

A novel and robust three step synthesis approach generated the full library of nintedanib analogs. Modifying nintedanib with hydrophobic or electronegative nonpharmacological cores significantly lowered the critical aggregation concentration in aCSF from 44.83 µM to 1.29 µM. The stability of colloidal aggregates is intrinsic to their physiochemical properties: π-π stacking, hydrogen bonding or van der Waals interactions; increased intermolecular interactions introduced by chemical modifications resulted in more stable nanoparticles formation. While unmodified nintedanib colloidal aggregates precipitated after 2 hours at 37˚C, analogs remained stable for up to 3 days when co-formulated with small molecule dye. Embedding these colloidal aggregates in a hydrogel enabled sustained release of small molecules over 7 days.

In conclusion, chemical modification of nintedanib enabled sustained release of antifibrotic molecules for 7 days from a colloidal aggregates-embedded hydrogel. These findings highlight the potential of a colloidal aggregates-based platform for drug delivery, paving the way for improved formulations in drug optimization and controlled drug delivery.

Generative Flow Networks (GFlowNets) have emerged as a powerful framework for molecular generation by enabling proportional sampling of diverse high-reward molecules. In practice, however, GFlowNets rely heavily on proxy reward models, such as docking scores or learned property predictors, that are often trained on limited data and subject to substantial label noise. These sources of uncertainty can severely distort reward signals, leading to suboptimal training dynamics and degraded molecule quality. Despite their importance, uncertainty in proxy rewards remains largely unaddressed in existing GFlowNet-based molecular design pipelines.

In this work, we propose an uncertainty-aware GFlowNet framework that explicitly quantifies and incorporates proxy uncertainty using conformal prediction. Conformal prediction provides distribution-free uncertainty quantification, enabling us to construct confidence intervals around proxy predictions without assuming a specific underlying model. We leverage conformal prediction intervals to adjust proxy rewards during GFlowNet training, reducing the influence of unreliable predictions while preserving informative reward structure. Also, by conformally adjusting the reward signal, enabling us to create a more robust and calibrated guidance mechanism that better reflects the actual reliability of the proxy model across different regions of the chemical space.
We identify two major sources of uncertainty commonly encountered in molecular generation tasks: (1) data scarcity during proxy model training, and (2) intrinsic noise in molecular property labels, such as stochastic docking outcomes. To systematically evaluate our method, we construct controlled experimental settings by artificially varying the size of proxy training datasets and injecting different levels of stationary noise into reward signals.

Our experimental validation on the soluble epoxide hydrolase (sEH) target, a clinically important enzyme implicated in inflammation, stroke, Alzheimer's disease, and obesity-related disorders, demonstrates that the proposed uncertainty-aware approach effectively mitigates the impact of both uncertainty sources and consistently outperforms standard GFlowNet baselines across all uncertainty regimes. In particular, our method improves the quality of generated molecules, enhances robustness under severe data scarcity, and maintains stable performance in the presence of noisy rewards. These findings highlight the importance of uncertainty calibration in proxy-driven generative models and suggest conformal prediction as a practical and effective tool for improving the reliability of GFlowNet-based molecular design.

Proper control of protein synthesis is critical to cell survival and function. Metabolic incorporation of amino acid analogues is a powerful method for interrogating proteomic changes directly at the translation level, but existing labeling strategies typically require either canonical amino acid deprivation or genetic engineering of translational machinery, which limit applicability in complex biological systems. We aim to develop an alternative strategy which addresses current challenges.

We have previously established TePhe, a tellurophene-bearing analogue of phenylalanine, as an effective protein synthesis probe for mass cytometry. Using only endogenous translation machinery, TePhe competes effectively against phenylalanine for incorporation into the proteome, enabling tagging of the nascent polypeptides under minimally perturbing conditions. Since then, we have also developed a tellurophene-selective bioconjugation reaction for the covalent modification of TePhe-containing proteins. This allows newly synthesized proteins to be visualized by fluorescence microscopy or subjected to affinity enrichment for further analysis.

This poster presentation will showcase the robust TePhe labeling that has been achieved in murine tissues, with particular focus on labeling within the brain. TePhe is easily administered via intraperitoneal injection; no other special treatment of the animals is required. Clear signal in neuronal cell bodies can be observed in formalin-fixed, paraffin-embedded sections obtained at 2 hr post-TePhe injection. Our chemistry is compatible with antibody staining, enabling simultaneous visualization of protein synthesis and biomarkers of interest. Progress toward affinity tagging and capture of TePhe-containing proteins for mass spectrometry-based proteomics, with would enable more in-depth comparisons of biological states, will also be discussed.

Intrinsically disordered regions (IDRs) constitute over 40% of products of human essential genes, which are indispensable to the viability of cells and frequently intertwined with critical cellular functions and diseases. Meanwhile, In silico structure-based protein design algorithms offer designs for reagents in research and disease treatment strategies, and structured protein designs have gained substantial recognition, featured in the 2024 Nobel Prize in Chemistry. However, IDRs are understudied due to their flexibility and lack of stable tertiary structures; as a result, strategies for computational protein design for IDRs have been ambiguous. Although significant breakthroughs have been made for structured protein designs, a lack of experimental data has been a major limitation for precise disordered protein design algorithms. Therefore, I propose to build a high-throughput imaging platform with a pioneer library containing synthetic IDRs resembling molecular features of disordered regions from human essential and non-essential fitness genes and their pathogenic variants.

Our first screen is designed to evaluate synthetic IDRs for their ability to translocate to stress granules (SGs). Cytoplasmic condensates such as SGs are central to post-transcriptional regulation and cellular stress responses. These structures are enriched in IDR-containing proteins, and the IDRs exhibit characteristic patterns of molecular features that are thought to contribute to their localization and function. They are also readily observable in live-cell imaging and play biologically and pathogenically important roles. Recognizing that in vitro phenotypes do not always translate directly to cellular contexts, we hypothesize that native IDRs sufficient to direct proper cytoplasmic localization do exist. Accordingly, we focused on IDRs from well-characterized SG proteins and assessed their localization behaviour in HeLa cells.

Suppressor tRNAs are sequence-modified human tRNAs that can induce translation through a premature termination codon introduced by a nonsense mutation. Due to the universality of this concept, suppressor tRNAs have the potential to treat ~10% of all genetic diseases, making them a promising potential RNA therapeutic. However, current suppressor tRNAs suffer from low readthrough efficiency due to competition with the endogenous translation termination machinery.

We propose that increasing the local concentration of charged, translation-ready suppressor tRNAs near the site of the disease-causing premature termination codon could enhance translational readthrough and circumvent the potential readthrough of endogenous stop codons. To this end, we engineered recruiter RNAs (rcRNA), which contain motives binding to the desired aminoacyl-tRNA synthetase and regions complementary to the disease-causing mRNA. Successful recruitment of the aminoacyl-tRNA synthetase to the specific mRNA would increase the local concentration of charged suppressor tRNA and thereby enhance readthrough efficiency and safety. We built a computational pipeline to optimize rcRNA fold and complementarity to the target mRNA. Resulting sequences were in vitro transcribed into rcRNAs and transfected into reporter cells alongside suppressor tRNAs. One of our rcRNA designs showed a promising trend towards increasing the readthrough efficiency of sup-tRNAs at premature termination codons. We aim to extend this principle into a platform to develop more potent and stable rcRNAs. Ultimately, rcRNA could be applied to increase the potency and specificity of suppressor tRNAs across a wide range of human genetic diseases and could be used in biotechnological applications involving stop codon readthrough.