Authors
Brian McEllin, Julia Robbins, Evan Hubbard, Daniele Canzani, Gaelle Mercenne, Sebastian J Paez, Lilly TL Tatka, Will Fondrie, Alex Federation, Lindsay K Pino

Institutions
Talus Bioscience

Abstract
Intro: Neuroblastoma is a highly aggressive pediatric cancer that accounts for a significant proportion of childhood cancer-related deaths. Neuroblastoma tumor progression is driven by MYCN gene amplification. In cases where MYCN gene is amplified, rapid tumor progression and more severe clinical outcomes are commonly observed. Although it plays a critical role in oncogenesis, MYCN transcription factor is a very difficult therapeutic target due to its intrinsically disordered structure and lack of binding sites for small molecule inhibitors. A promising therapeutic approach for Neuroblastoma drug discovery stems from identifying small molecules that can disrupt MYCN activity by disrupting associated pathways or even the protein itself.

Methods: We developed TF-Scan, a semi-automated high-throughput screening approach to quantify changes in the chromatin-associated proteome by mass spectrometry after drug treatment. The TF-scan platform involves treatment of live cells, followed by nuclear isolation. Isolated nuclei are subjected to washes of increasing salt concentration in order to enrich for DNA-bound protein complexes found in the chromatin subcellular fraction. Chromatin proteins are then digested via SP3 digestion on a KingFisher. The resulting chromatin peptides are then quantified using DIA mass spectrometry using an EvoSep One-timsTOF Ultra mass spectrometer with diaPASEF. Data is searched using our in-house cloud pipeline, quantMS, in order to quantify perturbations in MYCN-activity.

Preliminary Data: To optimize TF-Scan for neuroblastoma research, we systematically evaluated multiple neuroblastoma cell lines representing diverse genetic backgrounds, including MYCN-amplified and non-amplified models. This allowed us to refine assay conditions—such as cell density, salt extraction gradients, and digestion protocols—to maximize the recovery of chromatin-associated proteins while maintaining reproducibility across experiments. Our proteomic analyses revealed distinct chromatin-bound protein signatures between cell lines, highlighting key oncogenic pathways beyond MYCN amplification. For example, we observed differential enrichment of transcription factors, chromatin remodelers, and signaling molecules implicated in neuroblastoma progression, such as ALK, PHOX2B, and GATA family proteins. These findings suggest that neuroblastoma exhibits heterogeneous chromatin landscapes shaped by underlying genetic alterations. Additionally, our platform enabled the detection of dynamic changes in chromatin-associated proteins under various experimental conditions, providing a high-resolution view of transcriptional regulation in neuroblastoma. This depth of analysis supports not only target identification but also the discovery of context-dependent regulatory networks that could be exploited therapeutically. The scalability and robustness of TF-Scan position it as a versatile tool for both functional genomics studies and drug discovery applications in neuroblastoma and other cancers. Utilizing our high-throughput screening approach, we were able to identify small molecule compounds which killed neuroblastoma cells, through possibly MYCN or MYCN-associated protein interactions.

Novel Aspect: A high-throughput proteomics platform quantifying chromatin-associated MYCN, amongst other transcription factors, to uncover therapeutic vulnerabilities in neuroblastoma.

Authors
Joseph S. Pak(1), Seamus R. Morrone(2), Karla Opperman(1), Neal D. Mathew(1), Mukul K. Midha(2), Charu Kapil(2), Damon Page(1,3,4), Ning Zheng(5,6), Robert L. Moritz(2), Brock Grill(1,3,5)

Institutions
(1)Norcliffe Foundation Center for Integrative Brain Research, Seattle Children’s Hospital, Seattle, WA 98101 (2)Institute for Systems Biology, Seattle, WA 98109 (3)Department of Pediatrics, University of Washington, Seattle, WA 98195 (4)Department of Neurobiology and Biophysics, University of Washington, Seattle, WA 98195 (5)Department of Pharmacology, University of Washington, Seattle, WA 98195 (6)Howard Hughes Medical Institute

Abstract
Johanson-Blizzard Syndrome (JBS) is an autosomal recessive spectrum disorder associated with the UBR-1 ubiquitin ligase that features developmental delay including motor abnormalities. Here, we demonstrate that C. elegans UBR-1 regulates high-intensity locomotor behavior and developmental viability via both ubiquitin ligase and non-ligase mechanisms. Super-resolution imaging with CRISPR-engineered UBR-1 and genetic results demonstrated that UBR-1 is expressed and functions in the nervous system including in pre-motor interneurons. To decipher mechanisms of UBR-1 function, we deployed CRISPR-based proteomics which identified a cadre of glutamate metabolic enzymes physically associated with UBR-1 including GLN-3, GOT-2.2, GFAT-1 and GDH-1. Proteomics, multi-gene interaction studies, and pharmacological findings indicated that UBR-1, GLN-3 and GOT-2.2 form a signaling axis that regulates glutamate homeostasis. Developmentally, UBR-1 is expressed in embryos and functions with GLN-3 to regulate viability. Overall, our results suggest UBR-1 is an enzyme hub in a GOT-2.2/UBR-1/GLN-3 axis that maintains glutamate homeostasis required for efficient locomotion and organismal viability. Given the prominent role of glutamate within and outside the nervous system, the UBR-1 glutamate homeostatic network we have identified could contribute to JBS etiology.

Authors
Belinda B. Garana1; Hsin-Yun Lin2; Marija Veličković1; Samantha A. Obermiller1; Camilo Posso1; Marina A. Gritsenko1; Karin D. Rodland1,2; Elie Traer2; Cristina E. Tognon2; Jeffrey W. Tyner2; Paul D. Piehowski1; Anupriya Agarwal2; Sara J.C. Gosline1

Institutions
1Pacific Northwest National Laboratory, Richland, WA 2Oregon Health & Science University, Portland, OR

Abstract
Acute Myeloid Leukemia (AML) is a hematopoietic cancer caused by a block in cell differentiation resulting in uncontrolled proliferation of immature blasts in patient blood and bone marrow. However, this block is incomplete and, therefore, leads to a heterogeneous mix of cell populations in varying stages of maturity, making it difficult to treat the entire blast population using a single targeted therapy. Specifically, patients with a large population of mature monocyte-like (CD14+) cells are more likely to be resistant to drugs such as venetoclax, a drug that targets the B-Cell lymphoma-2 (BCL2) protein, and azacitidine, an inhibitor of DNA methylation, that have proven to be highly effective in combination for those patients whose AML is dominated by primitive (CD34+) cell populations. Recent advances in single-cell RNA-seq have highlighted the population-level complexity of AML patients corresponding to poor prognosis, but how the behaviors of individual populations associate with protein abundance is poorly understood, particularly in the context of drug resistance. Here, we introduce a proteomics-based approach that sorts patient samples into distinct populations by cell surface markers and examines proteins that give rise to drug resistance within the monocytic and primitive cell populations. We collected bone marrow from 22 AML patients, each with varying responses to venetoclax and azacitidine according to our ex vivo assay. We sorted these samples by expression of two cell surface markers: CD34+ to represent primitive cell populations and CD14+ to represent monocytes. Using a data independent acquisition (DIA) liquid chromatography tandem mass spectrometry (LC-MS/MS) proteomics approach, we identified 6,887 proteins, of which 2,597 were differentially expressed between CD14+ and CD34+ populations. Using gene set enrichment analysis of 50 hallmark gene sets from the Human Molecular Signatures Database, we identified increased TNF-alpha signaling via NF-kappa B and decreased expression of MYC targets in both the monocytic CD14+ samples compared to the more primitive CD34+ samples and samples derived from AML resistant to venetoclax versus sensitive. We also developed a monocyte score which predicts the fraction of monocyte populations and ex vivo sensitivity to venetoclax and azacitidine in bulk measurements of 210 AML patient samples. Recent advances in the sensitivity of LC-MS/MS proteomics have enabled the analysis of small (~15,000 cell) patient samples with increased depth and reproducibility, enabling us to further our mechanistic understanding of drug resistance in AML. This study offers a novel protein-based approach to understand cell maturation heterogeneity in AML and may uncover new potential targets to inhibit in drug-resistant patients.

Authors
Saman Rahmati, Jacob Porter, Elie Traer, Jeffrey Tyner, & Andrew Emili

Institutions
Division of Oncological Sciences, Knight Cancer Institute, Oregon Health & Science University

Abstract
Cell surface receptors play a central role in acute myeloid leukemia (AML) by driving malignant signaling, microenvironmental crosstalk, and therapeutic response. However, defining the protein-protein interaction networks surrounding these receptors in their native cellular context remains technically challenging, particularly for identifying transient, membrane-associated interactions. Traditional methods like immunoprecipitation and affinity purification-mass spectrometry often fail to detect key binding partners. To overcome these limitations, we have deployed complementary proximity labeling (PL) approaches to precisely map the interactomes of FLT3 and PEAR1—two key oncogenic receptors implicated in AML pathogenesis and treatment resistance. Using an antibody-guided proximity labeling approach, we are targeting these receptors in AML cell lines with recombinant Protein A–TurboID to enable selective biotinylation of receptor-proximal neighborhoods without genetic modification. In parallel, using TurboID-receptor fusion-based PL in Ba/F3 cells, we have biotinylated receptor-proximal interaction networks under physiologic conditions, and have performed quantitative mass spectrometry to define local receptor-specific interactomes.

This dual strategy enables high-resolution profiling of dynamic surface-associated multi-protein signaling complexes, revealing distinct receptor-specific functional networks. Our initial data highlight novel potential co-regulators and modulators of leukemogenic signaling, demonstrating the utility of proximity proteomics for dissecting receptor biology in hematologic malignancies. Overall, our preliminary findings support the integration of multi-pronged PL-based interactome mapping approaches into functional AML research and therapeutic target discovery pipelines.

Authors
Julia Robbins, Gabriel E Boyle, Andrea Gutierrez, Daniele Canzani, Jay Sarthy, Alexander J Federation, Lindsay K Pino

Institutions
Talus Bioscience, Seattle Children's Research Institute

Abstract
Introduction Anthracyclines, drugs that target the DNA repair system, are prescribed for many pediatric cancers but their use must be balanced against lifelong side effects and potentially lethal cardiotoxicity. Identifying similarly effective anthracycline derivatives with reduced toxicity could drastically improve outcomes for pediatric cancer survivors who receive anthracycline chemotherapy. Talus’ functional proteomics assay TF-Scan circumnavigates the low-abundance and disordered structure characteristic of transcription factors (TFs), using subcellular fractionation to isolate active chromatin-bound proteins in the native cell environment. The platform’s ability to quantitatively interrogate the relationship between proteins and compounds makes it an optimal tool for investigating new therapeutics that impact genome regulation and repair.

Methods Leukemia (THP-1) and rhabdomyosarcoma (RH4, RH30, and RD) cells were seeded in 96-well plates and treated with an anthracycline, anthracycline derivative, or a vehicle control for 2 hours. The cells were collected simultaneously using a VIAFLO96, incubated with lectin-coated magnetic beads, and their nuclei were isolated and subjected to subcellular fractionation on a KingFisher Apex. The resulting fractions were digested in-plate using AutoSP3, and approximately 100 ng of peptides from each sample were loaded onto EvoTips before diaPASEF analysis on a timsTOF Ultra-EvoSep platform. Data processing was carried out using QuantMS (an in-house pipeline) and Python scripts.

Preliminary Data Our analysis revealed both expected and novel protein dynamics after anthracycline treatment. Doxorubicin is thought to cause chromatin disruption by DNA intercalation and histone ejection, and it induces DNA double-strand breaks by inhibiting topoisomerase II (TOP2A). Analogues like aclarubicin disrupt chromatin but do not directly damage DNA, making them mechanistically distinct and therapeutically relevant.

Doxorubicin, dimethyl-doxorubicin, aclarubicin, and etoposide treatments in THP-1 cells resulted in marked decrease in chromatin-bound TOP2A, substantiating the known mechanism of anthracyclines. Further, doxorubicin, dimethyl-doxorubicin, and aclarubicin all created significant increase in ribosome biogenesis factors like WDR43/45 in their chromatin fractions relative to DMSO, suggesting a compensatory stress or repair response. Our analysis showed an increased abundance of histone variants following dimethyl-doxorubicin treatment, aligning with literature reports that anthrocyclines cause histone exchange for alternative isoforms, suggesting that nascent histone binding is part of the cell response. Additionally, proteins like ribosome biogenesis factor HEAT3 shifted from the cytoplasm to chromatin upon aclarubicin treatment, while DNA-binding proteins like FUBP3 significantly decreased in chromatin after dimethyl-doxorubicin and aclarubicin exposure.

In rhabdomyosarcoma cells (RH4, RH30, RD), dimethyl-doxorubicin treatment caused a substantial reduction levels of the oncogenic TF MYOD1, and a pronounced decrease in RNA polymerase subunits RPB1/2 within chromatin fractions, highlighting disruptions in transcriptional regulation. Rhabdomyosarcoma lines showed less profound TOP2A loss, except with etoposide treatment.

These findings indicate that anthracycline activity is more complex than previously understood; suggesting it extends beyond TOP2A inhibition to reshape transcriptional programs, ribosome biogenesis, and chromatin architecture in a cell-type-specific manner. By mapping protein dynamics with high quantitative accuracy, we uncover previously unrecognized mechanisms of drug response that may underlie both therapeutic effects and toxicities. These insights lay the groundwork for designing next-generation anthracycline derivatives with improved safety profiles, bringing us closer to more effective and less harmful treatments for pediatric cancers.

Authors
Gennifer E. Merrihew-1, Deanna L. Plubell-1, Michael Riffle-1, Julia E. Robbins-1, Bo Wen-1, Nick Shulman-1, Jea Park-1, Christine C. Wu-1, Kathleen L. Poston-2, Thomas Montine-2, Michael J. MacCoss-1

Institutions
1- University of Washington, Seattle, WA; 2-Stanford University, Stanford, CA

Abstract
Neurodegenerative diseases often present with complex comorbidities and mixed pathologies, making diagnosis challenging. To investigate biological signatures associated with these diseases, we employed a systematic approach to gather quantitative proteomics data from lumbar cerebrospinal fluid (CSF) samples of 330 patients. These samples were carefully selected to be age- and sex-matched, including both individuals with and without cognitive and motor impairments. To ensure accurate and consistent peptide quantification across different patient samples, we developed a series of controls to evaluate system suitability, sample preparation quality, and batch-level quantitative accuracy.

CSF samples were diluted in SDS buffer containing an internal control yeast enolase protein to evaluate digestion, followed by reduction and alkylation. Proteins were captured on MagResyn Hydroxyl beads, washed, and digested into peptides using a Thermo KingFisher Flex. Peptides were then separated using reverse-phase chromatography with a Thermo Easy nano-LC system coupled to a Thermo Orbitrap Eclipse Tribrid mass spectrometer, where they were analyzed using a 12 m/z staggered DIA isolation scheme. The resulting data were demultiplexed to 6 m/z using Proteowizard. A fine-tuned library for each batch was created using Carafe maintaining batch specific predictions using gas-phase fractionated libraries. Peptides were identified and peak boundaries were assigned with EncyclopeDIA, and the data were imported into Skyline. An additional analysis is being completed with DIA-NN instead of EncyclopeDIA.

CSF from 330 patients were divided into four major groups: 1) Healthy Control (HC), 2) Alzheimer’s Disease (AD), 3) Mild Cognitive Impairment (AD/MCI), and 4) Lewy Body Diseases (LBD). An external inter-batch control from a pool of 50 patients representing all 4 groups was used to assess the technical precision within and between each batch prior to and following normalization and batch adjustment. We also use internal process controls to assess sample preparation and monitor data collection. The use of these controls was crucial given that the data spanned seven batches, analyzed on separate LC columns and traps over a three-month period.

To quantify the same peptides across these batches, fine-tuned batch specific libraries with batch level predictions were created using Carafe from the union of the peptides found in all batches. These libraries were then used to identify peptides in each sample with EncyclopeDIA resulting in 29296 quantitative peptides mapping to 3718 proteins. After normalization and batch correction, we performed feature selection using Boruta and binary classification using XGBoost. The mean cross validation ROC curve between AD/MCI versus samples without AD was 0.91 +/- 0.08 and Healthy Controls versus AD and Lewy Body Disease was 0.90 +/- 0.05.

Authors
Kristine A. Tsantilas, Jeffrey R. Whiteaker, Lei Zhao, Regine M. Schoenherr, Uliana J. Voytovich, Richard G. Ivey, ChenWei Lin, Zhangfang Guo, Cynthia X. Ma, Amanda G. Paulovich

Institutions
Fred Hutchinson Cancer Center, Washington University in St. Louis School of Medicine

Abstract
Antibody-drug conjugates (ADCs) link cytotoxic treatments to monoclonal antibodies to deliver their “payload” to tumor cells with greater specificity. For HER2 non-amplified breast cancer, patient selection for HER2-targeting ADC therapy, such as trastuzumab deruxtecan, depends on quantifying the target protein antigen, typically using immunohistochemistry (IHC). However, HER2 IHC scoring is not sufficiently quantitative and is often subject to interobserver variability at the lower range. Thus, there is an unmet need for an improved assay for patient selection. To demonstrate the potential for this approach, we developed a multiplexed immuno-MRM assay targeting HER2, as well as the target of the payload (DNA topoisomerase 1, TOP1) and DNA-damage response (DDR) proteins (reflecting the pharmacodynamic response to TOP1 inhibition).

Proof-of-principle studies were conducted using two HER2-low (IHC 1+ or 2+, ISH negative) breast cancer patient-derived xenograft (PDX) models implanted into mice. Tumors were harvested from 3 mice each following short-term treatments with either vehicle (DMSO), HER2 ADC (trastuzumab deruxtecan or DS-8201a), Olaparib, or the combination of DS-8201a+Olaparib. Frozen tumors were processed by extracting whole protein lysate and spiking with cleavable stable isotope-labeled peptide standards. Following trypsin digestion, targeted peptides and their standards were enriched by peptide immunoaffinity enrichment. LC-MRM analysis of 660 transitions across a single 25-minute gradient was performed with an Ekspert nanoLC 425 (Eskigent) coupled to a 5500 QTrap (SCIEX).

The HER2-ADC assay quantifies 91 peptides derived from 52 proteins by leveraging peptide immunoaffinity enrichment with 69 antibodies. The panel quantifies HER2 (tumor antigen), the payload target (TOP1), and concurrently evaluates pharmacodynamic responses to the TOP1 inhibitor (deruxtecan) by measuring components of the DDR network. We applied the HER2-ADC panel to HER2-low breast tumors from murine PDX models (WHIM2 - IHC 2+/ISH negative and WHIM6 - IHC 1+/ISH negative) with and without treatment. In treated samples, the HER2-ADC panel identified activated DDR pathway targets based on phosphorylation, which were expected upon successful delivery of the ADC payload (TOP1 inhibitor). The targets elevated by treatment with the HER2 ADC alone and in combination with Olaparib regiment included pS1524 on BRCA1, pS343 on NBN, pS395 NUMA1, and pS317 CHEK1. Additional targets with increased responses in the PDX models included elevated levels of FANCD2, ATR, PCNA, LIG1, and PCNA.

Ongoing work will focus on completion of fit-for-purpose bioanalytical validation studies in frozen tissues using CLSI and CPTAC guidelines. This includes verification of selectivity, stability, and reproducible detection of endogenous analytes. This proof-of-principle demonstration with HER2 sets the stage for expansion into other ADCs, since we have developed MRM assays to additional proteins targeted by ADCs. These assay panels could contribute to pharmacodynamic and mechanism-of-action studies and facilitate patient selection efforts for these exciting, targeted therapeutics.

Authors
Justin Sanders[1], Melih Yilmaz[1], Jacob H. Russell[1], Wout Bittremieux[2], William E. Fondrie[3], Nicholas M. Riley[1], Sewoong Oh[1], and William Stafford Noble[1]

Institutions
[1] University of Washington [2] University of Antwerp [3] Talus Biosciences

Abstract
Machine learning has shown great promise to improve the analysis of mass spectrometry data. Here, we propose unifying various prediction tasks which take spectra as input under a single foundation model. In many settings, insufficient training data and noisy training labels make it challenging to learn a rich understanding of mass spectra in isolation for a given task. We hypothesize that learned spectrum embeddings, pre-trained on a large dataset of high-confidence spectrum annotations, may prove a valuable starting point for varied downstream prediction tasks.

We pre-train a spectrum encoder using de novo sequencing as a pre-training task. We then evaluate the usefulness of these pre-trained spectrum representations on the four downstream tasks of spectrum quality prediction, chimericity prediction, phosphorylation prediction, and glycosylation status prediction. On each task, we compare the performance of a small model trained on pre-trained spectrum representations to a series of task-specific baselines trained from scratch.

We find that our foundation model for mass spectra consistently outperforms the baselines, yielding state-of-the-art performance on all four downstream tasks. On the task of predicting whether a given spectrum contains signal from a phosphorylated peptide, we achieve an AUROC of 0.988, outperforming previously published methods AHLF (0.920) and phostar (0.916), and do so while relying on a fraction of the training data. Predicting whether a given glycopeptide spectrum contains an N- vs. O-glycosylated peptide, we achieve an area under the precision-recall curve of 0.914, compared to 0.753 for an existing strategy for glycoproteomics dissociation-type selection based on relative abundances of pre-selected oxonium ions.

We demonstrate that the spectrum encoder learned by a model trained on the de novo sequencing task learns generally applicable spectrum representations. In our analysis of downstream tasks, we demonstrate improvements on PTM prediction tasks which can help improve the ability of targeted experiments to identify and localize modifications.

Authors
Gwenneth Straub, Varun Ananth, Will Fondrie, Daniela Klaproth-Andrade, Michael Riffle, Justin Sanders, Bo Wen, Melih Yilmaz, Michael J. MacCoss, Sewoong Oh, Wout Bittremieux, William Stafford Noble

Institutions
Department of Genome Sciences; University of Washington, Department of Computer Science; University of Antwerp, Talus Bioscience, Technical University of Munich, Paul G. Allen School of Computer Science and Engineering; University of Washington

Abstract
Casanovo is a state-of-the-art deep learning model for de novo peptide sequencing from proteomics mass spectrometry data. Here we report on a series of enhancements to our open source Casanovo software, aimed at improving the interpretability of the scores assigned to predicted peptides, general- izing the software for use in database search, speeding up the software, and providing workflows and visualization tools to facilitate adoption of the tool and interpretation of the results. Our goal is to make Casanovo accurate and easy to use for applications such as metaproteomics, antibody sequencing, immunopeptidomics, and discovery of novel peptide sequences in standard proteomics analyses.

Authors
Jiayi Wang 1, Jules Nde 2, Andrei G. Gasic 3, Jacob Haseley 1, Margaret S. Cheung 1,4*

Institutions
1. Department of Physics, University of Washington at Seattle 2. Department of Cancer Biology, University of Kansas Medical Center 3. R&D department, GOWell International LLC 4. Environmental Molecular Sciences Laboratory, Pacific Northwest National Laboratory

Abstract
The intracellular environment is highly dynamic and crowded, where proteins form short-lived complex assemblages to regulate gene expression that vary with chromatin states. Investigating the factors that organize proteins within complexes in distinct chromatin states remains challenging. While macromolecular properties such as valency and intrinsic disorder are known to drive the formation of dynamic assemblage, the principles governing the spatial hierarchy of protein constituents remain unclear, particularly within crowded conditions that vary with cell states. Here, we present a computational framework that integrates experimental insights on proteins’ abundance and their interactions to reveal the emergence of subunit hierarchies under crowding. Using yeast INO80 chromatin remodeler complex as an example, we deployed a computational approach that integrate the mass spectrometry-based data to model protein assembly. We use a grand canonical coarse-grained model to iteratively infer subunits’ interaction energies and chemical potentials. We observed the emergence of macromolecular clusters whose size and composition vary with crowding. We grouped the constituents of INO80 complex into two classes, “converged” and “diverged”, according to their contribution to cluster organization: converged subunits contribute to the growth of an existing cluster, while diverged subunits initiate a new cluster. We speculate that diverged subunits, with have rich interaction and low abundance, are the candidates for regulating the spatial organization of chromatin remodeling complexes that drive chromatin state. We provide an agonistic framework to investigate other transient protein dynamics that vary with cell states.

Authors
Aaron J. Maurais (1), Gennifer E. Merrihew (1), Julia E. Robbins (1), Brian Connolly (1), Simona Colantonio (2), Joshua J. Reading (2), Joseph K. Knotts (2), Rhonda R. Roberts (2), Sandra S. Garcia-Buntley (3), Hongyan Ma (3), William Bocik (3), Jesse Stottlemyer (3), John Hamre (3), Eunkyung An (4), George E. Craft (5), Peter G. Hains (5), Roger R. Reddel (5), Phillip J. Robinson (5), Qing Zhong (5), Vincent Richard (6), Christoph Borchers (6-9), Kiminori Hori (10), Hiroki Shinchi (10), Koji Ueda (10), Hiroshi Nishida (11), Kosuke Ogata (11), Yasushi Ishihama (11), Kyu Jin Song (12-13), Jae-Won Oh (12-13), Kwang Pyo Kim (12-13), Hazara Begum Mohammad (14), Ye-Ji Do (14), Min-Sik Kim (14), Shinyeong Ju (15), Hankyul Lee (15),16, Cheolju Lee (15-16), Jingi Bae (17), Chaewon Kang (17), Sang-Won Lee (17), Ignasi Jarne (18-19), Joan Josep Bech-Serra (18-19), Tatiani Brenelli Lima (18-19), Carolina De La Torre (18-19), Lazaro Hiram Betancourt (20), Nicole Woldmar (20), Roger Appelqvist (21), Gyorgy Marko-Varga (21), Sandra Goetze (22-24), Bernd Wollscheid (22-23), Yi-Ju Chen (25), Hsiang-En Hsu (25), Hao Fang (25), Yu-Ju Chen (25), Yu-Tsun Lin (26), Kun-Yi Chien (26-27), Jau-Song Yu (26-28), Sara Ten Have (27), Angus Lamond (29), Nathan J. Edwards (30), Ana I. Robles (4), Henry Rodriguez (4), and Michael J. MacCoss (1)

Institutions
(1) Department of Genome Sciences, University of Washington, Seattle, Washington, United States. (2) Antibody Characterization Laboratory, Cancer Research Technology Program, Frederick National Laboratory for Cancer Research, Frederick, Maryland United States. (3) Proteomics Characterization Laboratory, Cancer Research Technology Program, Frederick National Laboratory for Cancer Research, Frederick, Maryland United States. (4) Office of Cancer Clinical Proteomics Research, Division of Cancer Treatment and Diagnosis, National Cancer Institute, National Institutes of Health, Rockville, Maryland 20850, United States. (5) ProCan, Children's Medical Research Institute, Faculty of Medicine and Health, The University of Sydney, Westmead, NSW, Australia. (6) Segal Cancer Proteomics Centre, Lady Davis Institute for Medical Research, Jewish General Hospital, Montréal, Quebec, Canada. (7) Division of Experimental Medicine, McGill University, Montréal, Quebec, Canada (8) Gerald Bronfman Department of Oncology, McGill University, Montréal, Quebec, Canada. (9) Department of Pathology, McGill University, Montréal, Quebec, Canada. (10) Cancer Proteomics Group, Cancer Precision Medicine Center, Japanese Foundation for Cancer Research, Tokyo, Japan. (11) Division of Medicinal Frontier Sciences, Graduate School of Pharmaceutical Sciences, Kyoto University, Kyoto, Japan. (12) Department of Applied Chemistry, Institute of Natural Science, Kyung Hee University, Yongin, Republic of Korea. (13) Department of Biomedical Science and Technology, Kyung Hee Medical Science Research Institute, Kyung Hee University, Seoul, Republic of Korea. (14) Department of New Biology, DGIST, Daegu, 42988, Republic of Korea. 15Chemical & Biological Integrative Research Center, Korea Institute of Science and Technology, Seoul, Republic of Korea. (16) Division of Bio-Medical Science & Technology, KIST School, University of Science and Technology, Seoul, Republic of Korea. (17) Department of Chemistry, Center for Proteogenome Research, Korea University, Seoul 136-701, Republic of Korea. (18) Proteomics Unit, Josep Carreras Leukaemia Research Institute (IJC), Badalona, Barcelona, Spain. (19) Josep Carreras Leukaemia Research Institute (IJC), Badalona, Barcelona, Spain (20) Department of Translational Medicine, Lund University, Skåne University Hospital Malmö, Sweden. (21) Department of Biomedical Engineering, Lund University, Lund, Sweden (22) Institute of Translational Medicine at the Department of Health Sciences and Technology, ETH, Zürich, Switzerland. (23) Swiss Institute of Bioinformatics, Lausanne, Switzerland. (24) ETH PHRT Swiss Multi-Omics Center (SMOC), Zürich, Switzerland (25) Institute of Chemistry, Academia Sinica, Taipei, Taiwan. (26) Molecular Medicine Research Center, Chang Gung University, Taoyuan Taiwan (27) Laboratory of Quantitative Proteomics, Division of Molecular, Cell, and Developmental Biology, School of Life Sciences, University of Dundee, Dundee, Scotland, UK (28) Department of Biochemistry and Molecular & Cellular Biology, Georgetown University, Washington, DC, USA.

Abstract
Quantitative proteomics is a powerful tool in both translational and basic biological research, however variability in sample preparation, instrumentation, and data acquisition can pose significant challenges when performing cross-laboratory comparisons. To demonstrate that proteomic data acquired in different laboratories can be leveraged in large-scale analyses, we coordinated a multi-site study where member labs of the International Cancer Proteogenome Consortium (ICPC), were sent aliquots of 9 different cell lysates. Each participating lab prepared and analyzed the samples using their own sample preparation and acquisition methods yielding a total of 22 distinct datasets. Our results demonstrate that median normalization combined with batch correction effectively reduces inter-lab variability and enables clustering by biological phenotype rather than technical origin. This work provides a framework for large-scale, cross-laboratory proteomics studies and highlights best practices for data harmonization in multi-site studies.

Authors
Lincoln Harris, Michael Riffle, William Stafford Noble, Michael J. MacCoss

Institutions
Department of Genome Sciences, University of Washington Paul G. Allen School of Computer Science and Engineering, University of Washington

Abstract
The traditional approaches to handling missing values in DIA proteomics are to either remove high-missingness proteins or impute them with statistical procedures. Both have their disadvantages—removal can limit statistical power, while imputation can introduce spurious correlations or dilute signal. We present an alternative approach based on imputing peptide retention times (RTs) rather than quantitations. For each missing value, we impute the RT boundaries, then obtain a quantitation by integrating the chromatographic signal within the imputed boundaries. Our method yields more accurate quantitations than existing proteomics imputation methods. RT boundary imputation also identifies differentially abundant peptides from key Alzheimer’s disease genes that were not identified with library search alone. RT boundary imputation improves the ability to estimate radiation exposure in biological tissues. RT boundary imputation significantly increases the number of peptides with quantitations, leading to increases in statistical power. Finally, RT boundary imputation better quantifies low abundance peptides than library search alone. Our RT boundary imputation method, called Nettle, is available as a standalone tool.

Authors
Lillian T. Tatka, Sebastian J. Paez, Lindsay K. Pino, Alexander J. Federation, William E. Fondrie

Institutions
Talus Bioscience, Seattle WA

Abstract
Modern mass spectrometry has increased proteomics throughput, yet most experimental designs still focus on pairwise comparisons, which scale poorly in high-throughput screens. Traditional approaches struggle to scale to testing thousands of conditions due to computational bottlenecks or failing to account for the complexity of varied molecular interactions. As a result, unbiased approaches like data-independent acquisition remain underutilized in drug discovery, where millions of drug-protein interactions are measured to call “hits”.

We benchmark three statistical approaches for hit-calling in large-scale proteomic screens: an ordinary least squares (OLS) linear model, a linear mixed-effects model (LMM), and two-sided independent t-testing. The OLS model captures protein abundance changes as functions of treatment compound and well plate, while the LMM introduces additional nuance by incorporating well plate, instrument, and replicate as random effects. Both linear models estimate the significance of compound effects on protein abundance and are designed to scale to large datasets with complex experimental structure.

Using simulated datasets, we assessed model performance across varying noise levels by computing precision and recall. LMM and OLS models performed similarly and both showed improved precision and recall compared to t-testing at all simulated noise levels. To evaluate sensitivity, we simulated datasets with a fixed noise level and a single “hit” of varying effect size. At 80% power, both OLS and LMM detected log₂ fold changes as small as 0.45, whereas the t-test required fold changes of 0.9 to achieve equivalent power.

To validate the OLS model on biological data, we used calibration curve measurements in which all proteins are known to exhibit systematic changes in abundance at each point. At each curve point, OLS detected a greater fraction of expected abundance changes compared to t-testing. Finally, we assessed the ability of the OLS model to identify downstream reductions in androgen receptor (AR) levels in response to the NONO inhibitor (R)-SKBG-1. While both methods estimated similar effect sizes, only the OLS model classified this reduction as statistically significant (p < 0.001).

Overall, our results demonstrate that linear modeling offers improved precision, sensitivity, and confidence compared to traditional t-testing. These models enable more reliable hit calling and better prioritization of promising compounds in large-scale proteomic drug discovery screens.

Authors
Anastasiya Prymolenna, Sebastian Paez, Kyle Siebenthall, Lilly Tatka, Lindsay Pino, Alex Federation, Will Fondrie

Institutions
Talus Bio

Abstract
As AI drives scientific advancements, mass spectrometry proteomics benefits significantly from improved data processing and search methodologies. However, its application remains largely confined to computational tasks rather than biological interpretation. A challenge in leveraging AI for biological insights lies in the limitations of domain-specific AI, particularly the lack of robust metadata. To overcome this challenge, we have streamlined metadata collection, enabling more effective machine learning models for small-molecule screening in drug discovery and development. By placing metadata collection and data governance at the core of our data management, our approach not only enhances accuracy, scalability, and machine learning insights but also streamlines experimental workflows to reduce burdens on our researchers while increasing throughput.

We developed a central metadata database to streamline and standardize data organization across the proteomics workflow, including cell culture, treatment, sub-cellular fractionation, peptide digestion, and mass spectrometry. MySQL, hosted on PlanetScale, serves as our backend database, where tables, indexes, and relationships between data are maintained in a serverless and distributed manner. SQLAlchemy with FastAPI is used in Python to interact with the database by defining the data models, executing queries, and completing transactions. On the user side, an intuitive Streamlit web application registers metadata directly into the database, while Pydantic ensures that data conforms to the defined data model. This centralized system supports in-house workflows for identifying protein-compound interactions and promotes our machine learning algorithms that predict drug-protein interactions.

To support the many facets of data needed to train robust machine learning models, we must accurately capture and structure the entire workflow used to generate mass spectra. We focused on developing our data model to accurately reflect the laboratory workflow. The experimentalists’ ease of digitizing their physical sample manipulations is at the forefront of our approach to collecting metadata. Users simply upload a pre-structured Excel file to the system’s UI and it performs all data validation and sanitization. Compared to commercial options, we sought to prioritize data consistency, centralization of data governance, and operational efficiencies. Further, we adhere to the FAIR principles (Findable, Accessible, Interoperable, and Reusable) in our metadata management practices, ensuring structured, well-documented, and high-quality data for optimized usability within our organization’s multiple disciplines, from medicinal chemists to mass spectrometrists to computational biologists, along with federal compliance requirements. We share vignettes of how our approach reduced time devoted to cleaning, curating, and inputting metadata, eliminated the need to embed metadata into MS file names, and increased time spent generating, rather than documenting, data.

Authors
Bo Wen [1], Shreyash Sonthalia [1], Priank Dasgupta [2], Chris Hsu [1], Michael J. MacCoss [1], and William Stafford Noble [1,2]

Institutions
[1] Department of Genome Sciences, University of Washington; [2] Paul G. Allen School of Computer Science and Engineering, University of Washington

Abstract
One critical step in any targeted mass spectrometry experiment is to select, from each protein of interest, a small number of peptides that will be targeted in the experiment. Ideally, the selected peptide targets exhibit robust quantitative signals that can be reproducibly measured by the mass spectrometer. Some existing methods for selecting which peptides to target rely on existing empirical measurements, but these methods do not work for peptides that have not previously been measured. An alternative approach uses machine learning methods to predict the peptide’s response on the basis of its amino acid sequence. However, existing methods in this category have been typically trained on relatively small datasets without considering peptide charge state. In this study, we leverage publicly available mass spectrometry data to generate a massive set of annotated peptides. We then train a deep learning model with a transformer architecture to rank peptides within a protein. We demonstrate the performance of our method on a variety of datasets.

Authors
Fred Mast, Arti Navare, Paul Olivier, Maxwell Neal, Michael Tyers, Robert Moritz, John Aitchison

Institutions
Center for Global Infectious Disease Research, Seattle Children's Research Institute; Department of Pediatrics, University of Washington; University of Toronto; Institute for Systems Biology

Abstract
Viruses rely on host protein complexes to complete essential steps in their replication cycles, creating points of dependency that can be exploited for therapeutic intervention. To identify these viral vulnerabilities, we developed an automated implementation of Isotopic Differentiation of Interactions as Random or Targeted (I-DIRT) integrated with mass spectrometry. This platform enables robust, quantitative detection of protein–protein interactions in infected versus uninfected cells and distinguishes stable from transient associations. Applying this approach to dengue virus, influenza virus, and other model pathogens, we identify viral hypomorphs — host proteins functionally compromised by infection — and validate them by targeting their synthetic lethal counterparts. Together, these findings establish automated interactome proteomics as a scalable and versatile platform for decoding and exploiting virus-induced vulnerabilities in host systems.

Authors
Chris Hsu, Amber Nolan, Christine C Wu, Dirk Keene, Michael J. MacCoss

Institutions
Department of Genome Sciences, University of Washington, and Department of Laboratory Medicine and Pathology, University of Washington

Abstract
Spatial proteomics for formalin-fixed paraffin embedded tissue (FFPE) at the low input level can provide insights into the complex and interesting morphological changes in tissues that would otherwise be lost in bulk sample experiments. Despite its potential, current spatial proteomics methods suffer from significant challenges in both sample preparation and sensitivity of analysis. These challenges hinder the efforts to effectively characterize complex biological changes from FFPE tissue. Here, we aim to address these challenges by optimizing our sample preparation in low volume (<5µL) in a single-pot to reduce sample loss and to further improve sensitivity of quantitative measurements. We apply our workflow to investigate spatial protein abundances in distinct regions in the cerebellum tissue portion of the brain.

Spatial cutting and gridding was performed on a Leica LMD6 laser capture microscope. Each FFPE “pixel” was lysed with 0.05% DDM, de-crosslinked by heating, reduced, alkylated, and digested. To minimize batch effects, samples were blocked and randomized for spatially gridded sections prior to MS analysis. Controls were used to evaluate sample preparation and mass spectrometer performance. For FFPE tissue size evaluation, the samples were run from smallest to largest volume. Data independent acquisition (DIA) data on the prepped FFPE tissue was acquired on an Orbitrap Astral paired to an EvoSep liquid chromatography system. The data was analyzed by first performing a Carafe search to fine tune the retention time and then searched by DIA-NN using the “Conservative NN” parameter.

We demonstrate the feasibility of our FFPE sample preparation strategy by optimizing for sample recovery at low volume (<5 µL). Key optimizations, including de-crosslinking time, sample clean-up, and digestion conditions, were also evaluated to maximize sample recovery and protein yield. To assess the impact of tissue size on measurement sensitivity, we used the Leica LMD6 laser capture microdissection system to section spatial regions from the white matter of the cerebellum, testing tissue areas ranging from 100x100µm to 1000x1000µm. The smallest section (100x100 µm) yielded an average of ~2,820 detected proteins, and the largest section (1000x1000 µm) resulted in ~5,500 proteins. The results remained consistent across different biological replicates, indicating our method is reproducible across biological replicates.

To demonstrate the feasibility of spatial analysis, we used laser capture microdissection to isolate 16 distinct FFPE samples (225x225 µm tissue area per sample) from the cerebellum cortex and white matter. Enolase was spiked into each sample as an internal quality control metric to ensure consistent sample preparation. We validated the spatial signature of the 16 samples by plotting a heatmap of protein abundance of several cerebellum-specific proteins. These included MAG and MBP in the white matter, CLBN1 and GRID2 in the cerebellum cortex, and CALB1 as a cerebellum protein control. To further assess the quantitative spatial data, we performed unsupervised clustering using principal component analysis (PCA), which successfully revealed signatures of patterns across the 16 samples – distinguishing the white matter and cortex regions by molecular similarity.

To identify protein drivers of the different clustering, we employed Boruta to assess features that are important in this classification. We were able to identify protein factors that best distinguished the cerebellum cortex from the white matter regions.

Authors
Michael R. Hoopmann, Christopher D. McGann, Devin K. Schweppe

Institutions
University of Washington

Abstract
Application programming interfaces, such as Thermo Fisher Scientific’s IAPI has paved avenues to develop software solutions to analyze MS data and alter MS data collection methods in real-time with spectral acquisition. However, real-time MS software development is an iterative process where many code changes occur from conception to release. Testing each of these changes with the use of a mass spectrometer and samples is slow, expensive, and creates a bottleneck in the laboratory.

We developed Corona, a virtual mass spectrometer for real-time MS software development. Corona interacts with the Thermo Fisher Scientific IAPI through the Helios API for rapid, iterative testing of applications without running a mass spectrometer. Instrument operation is replicated using previously acquired data, and spectral data is broadcast to RTMS applications utilizing IAPI. The speed of Corona is adjustable, up to 100x faster in our tests, or set to skip ahead and broadcast only a portion of the data, to facilitate rapid testing and analysis of the external applications with different data rates or specific acquisition events. Using Corona, we are able to operate, test, and develop RTMS applications developed for both the Exploris and Tribrid Orbitrap mass spectrometry platforms.

Authors
Nima Ranjbar Sistani1, Naoki Oshimori2, Jim Korkola3, and Andrew Emili1

Institutions
1 Division of Oncological Sciences, Knight Cancer Institute, Oregon Health Science University (OHSU) 2 Department of Cell, Developmental & Cancer Biology, School of Medicine, Oregon Health & Science University (OHSU) 3 Biomedical Engineering, School of Medicine, Oregon Health & Science University (OHSU)

Abstract
Abstract: Understanding the complex molecular interactions that occur within the tumor microenvironment (TME), particularly crosstalk involving pre-cancerous cell populations, is essential for identifying early therapeutic targets and intercepting malignant progression. Proximity labeling (PL)- based proteomics offers a potentially powerful approach for mapping secreted effector signaling proteins under native TME conditions in vitro and in vivo. Here, we report the development and optimization of a PL workflow for application in both a pre-cancerous mouse epithelial model (COMMA-D) and a stromal cell line (BALB/c-3T3), alongside human HEK293T cells as a reference control. Each cell type was stably transduced with inducible lentiviral constructs encoding either TurboID alone or a TurboID-KDEL (ER-retention fusion), enabling compartment-specific biotin labeling of proteins in the cytosol and endoplasmic reticulum, respectively. Following induction and extensive biotinylation, streptavidin enrichment and on-bead protein digestion were performed, and samples were subjected to LC-MS analysis.

We prioritized label-free quantification (LFQ)-based profiling as a first-pass precursor to TMT labeling, enabling reproducible detection of dynamic proteome patterns without ratio compression. Extensive troubleshooting across labeling conditions, pulldown efficiency, and cell-type–specific expression yielded a robust and adaptable in-house platform for proximity proteomics in early neoplastic mouse models. This optimized workflow provides the foundation for downstream studies aimed at characterizing signaling architectures within senescent and pre-neoplastic microenvironments, both in vitro and in vivo, after ectopic cell transplantation, offering a scalable strategy to functionally map cell-type and context-specific secretion-based cell-cell signaling networks within the TME relevant to early cancer progression.

Authors
Kathryn Kothlow (1), Bo Wen (2), Chris D. McGann (2), Anna G. Duboff (1), Jacob H. Russell (1), Emmajay Sutherland (1), Devin K. Schweppe (2), William S. Noble (2,3), and Nicholas M. Riley (1)

Institutions
(1) Department of Chemistry, University of Washington, Seattle, WA, USA; (2) Department of Genome Sciences, University of Washington, Seattle, WA, USA; (3) Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA.

Abstract
INTRODUCTION Deep-learning enabled predicted peptide spectral libraries (e.g., Prosit, Prism, pDeep) have accelerated numerous applications in proteomics. One such application is real-time library searching (RTLS), where matches to predicted spectra can inform and direct scan acquisition in real time. RTLS offers several potential benefits to glycoproteomics, e.g., prioritizing complementary tandem MS scans on identified precursor ions, but current spectral prediction tools are largely trained on non-modified peptides and not robust for predicting glycopeptide spectra. Here we re-train peptide-centric deep learning models to predict glycopeptide fragmentation, and we use these glycopeptide spectral libraries in RTLS workflows to identify glycopeptides without needing a library entry for each specific glycan structure. We use our RTLS approach to capture glycan structural features for identified glycopeptides.

METHODS All experiments were performed on a Thermo Scientific Orbitrap Ascend Tribrid MS system (Thermo Fisher Scientific, San Jose, CA) equipped with a Vanquish Neo UHPLC System. All analyses used online reversed phase separations. Spectral libraries were created using Prosit (https://www.proteomicsdb.org/prosit/) or a custom-trained version of Carafe (https://github.com/Noble-Lab/Carafe). Glycopeptide samples were derived from commercial glycoprotein standards and from MAX enrichments of human cell lysates (e.g., HEK293, K562). RTLS was performed using custom C# code implemented into Orbiter, which interfaces with the Thermo Instrument API. Library matches were evaluated with cosine similarity scores and used to further inform several scan parameters including dissociation setting and accumulation time.

PRELIMINARY DATA Most glycoproteomics studies are only able to report, for each glycosite, the glycan composition rather than its structure. To elucidate glycan structure, it is often helpful to perform a collision energy ramp to generate fragments that reveal connectivity of monosaccharides. However, these collision energy ramps require multiple tandem MS/MS scans per precursor ion, meaning we would only perform a collision energy ramp on precursor ions ideally where we know that the time cost will translate to useful information. Here, we generate in silico spectral libraries and use them in RTLS methods to guide this process of prioritizing identified glycopeptide precursors for structural characterization. To account for many glycoforms, we require that the library spectra be impartial to the glycan composition, instead making an identification based on the peptide fragments. Previously, we manually adapted Prosit libraries to contain glycopeptide-specific ions. We found that RTLS with these modified libraries is able to identify a majority of database-identified glycopeptides. We improved this method by making a custom RTLS algorithm implemented in Orbiter that automatically filters out glycan-specific oxonium ions using our recently developed software GlyCounter. This decongests the spectra and allows peptide b- and y- ions to be searched more easily. We also examined the b- and y- ions of glycopeptide spectra on a large scale and found that there is a decrease in intensity from their non-glycosylated counterparts. To better predict the intensity of glycopeptide b- and y- ions, we fine-tuned Carafe’s peptide fragment ion intensity predictions using glycopeptide data, and we compared the identifications made to those made in an MSFragger-Glyco database search to evaluate RTLS accuracy. We found that confident RTLS matches can trigger collision energy ramps to generate glycopeptide structural information. Altogether, we demonstrate a method for generating in-silico glycopeptide libraries for use in RTLS methods to elucidate glycoforms.

Authors
Sarah Hadley, Yuqi Shi, Charlie Mundorff, Chistopher Mullen, Rosa Viner, Miklos Guttman

Institutions
Department of Medicinal Chemistry; University of Washington, Seattle, WA 98195 Thermo Fisher Scientific, San Jose, CA 95134

Abstract
Introduction H/D exchange with mass spectrometry (HDX-MS) typically provides peptide-level information on protein structure and dynamics. Achieving higher residue-level information has been limited by the effect of hydrogen scrambling, which causes hydrogen or deuterium to migrate across different amides prior to backbone fragmentation. Classical collision-based fragmentation techniques result in complete scrambling, while electron-based fragmentation methods (ETD) under gentle conditions can achieve fragmentation without scrambling. UVPD has also been shown capable of mitigating scrambling, but with conflicting reports. Here we systematically test different parameters on the levels of scrambling with UVPD in different peptides on a Thermo Orbitrap Tribrid system.

Methods: Peptide P1 (HHHHHHIIKIIK) in either deuterated or Optima water was analyzed by direct infusion on Thermo Orbitrap Ascend mass spectrometers equipped with UVPD at 213 nm. Source and ion transfer conditions were optimized to minimize scrambling of the +3 charge state of P1 with ETD fragmentation. Irradiation times were varied from 2 ms up to hundreds of ms with UVPD performed either in the low- or the high-pressure region of the dual-pressure ion trap. Spectra were analyzed in HX-Express v3 to analyze deuterium content on the relevant fragment ions.

Preliminary Data When performing UVPD in the low-pressure region of the ion trap, excessive scrambling was observed for all b and y fragment ions of the +3 charge state of P1. Using very low irradiation times of 2 to 4 ms resulted in limited scrambling for a and x ions, but with very low fragmentation efficiency. Longer irradiation times, needed for efficient fragmentation quickly resulted in complete scrambling of all peptide ions, largely consistent with prior studies. Comparison of UVPD performed in either the low- or high- pressure regions of the ion trap revealed that a much lower level of scrambling occurred in the high-pressure region. For some ions, even longer UVPD irradiation up to 100 ms could be performed without excessive scrambling, making higher fragmentation efficiency achievable. In addition, we found that even among the undeuterated samples, peak envelopes varied with irradiation time. This was observed for most abundant fragments, regardless of whether UVPD was performed in the low- or high-pressure cell. This is likely due to a change in the frequency of hydride transfers, which are commonly observed with UVPD, to or from the ion in question during fragmentation. As a result, matched deuterated and undeuterated samples are required to accurately assess deuteration. Without this control, scrambling may be over or underestimated, depending on the fragment.

Authors
Kyle T Siebenthall, Daniele Canzani, Andrea Gutierrez, Julia Robbins, Brian McEllin, Evan Hubbard, Anastasiya Prymolenna, Lillian Tatka, J Sebastian Paez, Gaelle Mercenne, William E Fondrie, Lindsay K Pino, Alexander J Federation

Institutions
Talus Bio, Seattle, WA

Abstract
Thousands of nuclear proteins orchestrate the complex process of gene regulation– from transcription factors that bind DNA and recruit co-factors to modify and remodel chromatin to facilitate or hinder the transcription of their target genes, to RNA-binding proteins that regulate the splicing of transcripts into mature mRNAs, to DNA repair factors that correct mutations to ensure faithful outputs of gene expression. Much of this work is carried out by large, multi-protein complexes such as MLL/COMPASS, which is responsible for methylating histone H3 lysine 4 at enhancers and promoters of active genes. Chromosomal translocations involving one of the MLL histone methyltransferases that comprise the enzymatic core of this complex drive multiple leukemias by constitutively activating target genes of the complex.

At Talus Bio, we have developed a proteomics-based drug discovery platform that is able to measure the effects of small-molecule perturbations on the thousands of proteins that regulate gene expression in live, unmodified cells. We report here on the proteomic profiling of the covalent MLL complex inhibitor M-1121, which targets the Menin protein to disrupt its interaction with MLL1 within the MLL complex. We show that M-1121 treatment strongly depletes Menin from chromatin in a dose-dependent fashion and relocalizes it to the nucleoplasm. Interestingly, this inhibition differentially affects the localization of a subset of MLL complex members, which may reflect compensatory stabilization of alternative assemblies or changes in chromatin accessibility that promote retention of the scaffold. We observed strong depletion of transcription factors that interact with the MLL complex (e.g. IKZF1) and that are directly upregulated by MLL-fusion proteins (e.g. HOXA11).

The simultaneous profiling of nearly all components of a chromatin-regulating complex alongside other nuclear proteins that directly interact with and are regulated by such a complex highlights the unique ability of the proteomics-powered Talus Bio platform to aid in the understanding of– and development of drugs against– the proteins that regulate our genome.

Authors
Theresa A. Gozzo, Jianji Chen, Rosalynn Molden

Institutions
Just-Evotec Biologics

Abstract
Introduction: Disulfide bonds are critical in stabilizing the three-dimensional structures of antibody-based therapeutics. Free and improperly paired cysteines can lead to aggregation or immunogenicity.

Other cysteine modifications, like glutathionylation, cysteinylation, or trisulfide and thioether bond formation can be considered product-related impurities, so it is essential to accurately characterize cysteines in biopharmaceuticals. Often, a bottom-up mass spectrometry approach is taken to localize disulfides and cysteine modifications; however, conditions like alkaline pH and high temperatures can lead to method-induced disulfide artifacts that complicate these analyses. Various methods have been proposed to minimize artifacts, but they only focus on the analysis of a subset of cysteine modifications, most often native and scrambled disulfide bonds. We present a method to provide a more comprehensive analysis.

Methods: Peptide mapping was performed on stressed and non-stressed NISTmAb, IgG1 mAb and SigmaMAb K4 samples to localize and quantify free thiols, glutathionylation, cysteinylation, trisulfide and thioether bonds, and scrambled and expected disulfide bonds. Two denaturants were compared: Rapigest™ and guanidine. N-ethylmaleimide was the alkylator. N-methylmaleimide will also be tested to differentiate between native and method-induced free thiols. Alkylating temperature and incubation time were tested to maximize alkylation while minimizing artifacts. Acidic pH was maintained throughout. Trypsin and LysC digestions were performed overnight at 37°C. Samples were split into non-reduced and reduced portions. Reduction was performed with TCEP. pH- and temperature-stressed samples were prepared to evaluate method performance. LC-MS/MS was performed on a Vanquish™ Duo and an Exploris™ 240 mass spectrometer.

Preliminary Data: The S-S workflow was used within Protein Metrics ByosTM Software to identify and quantify cysteine modifications. Choice of denaturant, pH, temperature, and incubation time were most influential for maximizing peptide recovery and minimizing method artifacts. When Rapigest was used, we observed improved digestion and disulfide-bonded peptide recovery for NISTmAb. For example, the signal for the peptide containing the expected C23-C87 disulfide bond improved from 1.7E6 to 1.1E8. Increased signal intensity for the C23-C87 bonded peptide also improved free thiol estimation at C23, bringing the value down from 24% to 3%. According to literature, NISTmAb is not expected to have elevated free sulfhydryl levels greater than 7%. Similar results were observed for SigmaMAb K4 with Rapigest denaturant. For all three non-stressed samples, trisulfide bonds were observed at low levels (at or below 1.1%) at cysteine residues participating in hinge and interchain bonding.

For SigmaMAb K4, trisulfide bonding was also observed at C90-C91 with the non-canonical cysteine residue. When alkylating temperature was increased to 78°C, increased disulfide bond scrambling was observed, even with short incubation times. Additionally, stressed samples showed clear differences from non-stressed samples; increases in scrambled disulfide bonds as well as thioether and trisulfide bonds were observed for samples that underwent incubation at pH 8.4 and 40°C. N-methylmaleimide will be used as a secondary alkylator to further investigate method-induced free thiols.

Authors
Evan Hubbard, Lindsay Pino

Institutions
Talus Bioscience

Abstract
Glycosylation is a modular, post-translational modification which is ubiquitous in the human proteome and necessary for normal biological function. In human cells, the nuclear membrane has a unique glycan profile from other organelles. Talus Bio developed lectin-coated magnetic beads which use these membrane-specific glycans as “handles” for binding and then isolating nuclei. We use this separation method in TF-Scan, an assay which directly observes the chromatin-bound proteome and detects changes to transcription behavior. However, in addition to thousands of chromatin-associated proteins, TF-Scan often identifies hundreds of proteins which are not associated with chromatin or even the nucleus generally. To better understand why these non-nuclear proteins are observed, we explored the possibility that they are lectin-compatible glycoproteins which directly interact with the magnetic beads. Through a series of modified TF-Scan procedures we found which proteins do not originate from the chromatin proteome. We then investigated the potential mechanisms of bead interaction for each of these proteins, from lectin coupling to direct bead binding via hydrophobic interactions. These experiments improve our assay by clarifying which proteins are directly reflective of changes to chromatin and which are indirect indicators of nuclear-localization at best.

Authors
Emmajay Sutherland, Tim S. Veth, and Nicholas M. Riley

Institutions
Department of Chemistry, University of Washington, Seattle, WA, USA

Abstract
Cell surface glycoproteins are a dominant feature of the glycocalyx, yet we know surprisingly little about the glycosylation states of any given protein at the cell surface. This limited progress can be attributed to two key challenges: (1) the inherent heterogeneity of glycosylation, a non-templated process that generates a vast array of proteoforms, and (2) the current limitations of analytical tools, which hinder comprehensive exploration of the cell surface proteome. Broadly speaking, these methods involve membrane purification as well as metabolic, enzymatic, and chemical labeling. One particularly useful technique, often referred to as Cell Surface Capture, involves the covalent modification of cell surface glycans with biotin, facilitating protein capture via streptavidin. A critical step in this workflow requires the release of the glycans from the peptides to prepare the sample for LC-MS analysis. However, this process disrupts the essential link between glycan identity and its corresponding protein glycosite, limiting the biological insights that can be derived from the workflow. This work focuses on leveraging commercially available biotin and streptavidin reagents to investigate the cell surface N-glycoproteome. We selected K562 human lymphoblast cells for their broad applicability, providing a foundation to extend this approach to other cell lines in the future. Biotin reagents consist of three key components: a biotin moiety for streptavidin binding, a spacer to enhance binding efficiency, and a reactive group for protein derivatization. A variety of biotin reagents were evaluated based on their distinct reactive groups, including options utilizing desthiobiotin. Additionally, three commercial options of biotin binding proteins were investigated, each with different affinities for the biotin tag. Overall, this work seeks to highlight the main caveats of traditional cell surface capture and allow us to ultimately develop a method that preserves glycopeptide integrity for analysis of the distinct proteoforms present at the cell surface.

Authors
Alex Zelter1, Michael Riffle1, Batool Mutawe1, Gennifer E. Merrihew1, Aaron Maurais1, Christine C Wu1, Michael J. MacCoss1

Institutions
1: Department of Genome Sciences, UW

Abstract
Current dogma suggests protein quantification is a pre-requisite to LC-MS/MS based proteomics studies. Quantification of protein amount allows the correct ratio of sample to digestion enzyme to be calculated for each sample and additionally enables normalization of the quantity of peptides loaded onto the mass spectrometer for analysis. Most proteomics studies perform such steps prior to sample digestion and analysis. In a world of unlimited resources, these steps would be carried out for every sample. However, it is important to recognize that there are significant costs, in both time, money and experimental complexity, associated with performing such quantification and physical normalization for every sample. This is especially true for large scale studies involving hundreds or even thousands of individual samples. Proteomics data analysis pipelines typically include computational data normalization strategies designed to compensate for systematic biases that are unavoidably introduced during sample processing and data collection. These strategies also have the potential to compensate for avoidable variation introduced during sample processing, such as by omitting sample amount normalization. In the current work we investigate the effects of either normalizing the protein concentration for each individual sample or leaving it unnormalized and accepting this range of concentrations in the hope that it can be sufficiently compensated for using computational normalization strategies applied after the mass spectrometry data is collected.

Our results show the relationship between increased protein concentration variation in sample input and the variance of quantified relative abundances of proteins output after data analysis. The data presented suggest that for many experiments protein quantification and physical normalization steps can be omitted from quantitative proteomic experimental workflows without incurring an unacceptable loss of sensitivity or an unacceptable increase in measurement variability after computational normalization has been applied. This work will enable important time and cost saving optimizations to be made to many proteomics workflows.