Knowledge in Data Science Symposium (KIDS23)

Evaluation of a Pooling Chemoproteomics Strategy with an FDA-approved Drug Library

Author: Huan Sun, PhD

Abstract: Chemoproteomics is a key platform for characterizing the mode of action (MoA) for compounds, especially for targeted protein degraders such as proteolysis targeting chimerics (PROTACs) and molecular glues. With deep proteome coverage, multiplexed tandem mass tag-mass spectrometry (TMT-MS) can tackle up to 18 samples in a single experiment. Here, we present a pooling strategy to further enhance the throughput, and apply the strategy to an FDA-approved drug library (95 best-in-class compounds). The TMT-MS-based pooling strategy was evaluated in the following steps. First, we demonstrated the capability of TMT-MS by analyzing over 15,000 unique proteins (>12,000 gene products) in HEK293 cells treated with five PROTACs (two BRD/BET degraders and three degraders for FAK, ALK, and BTK kinases). We then introduced a rationalized pooling strategy to separate structurally similar compounds in different pools, and identified the proteomic response to 14 pools from the drug library. Finally, we validated the proteomic response from one pool by re-profiling the cells under individual drug treatment with sufficient replicates. Interestingly, numerous proteins were found to change upon drug treatment, including AMD1, ODC1, PRKX, PRKY, EXO1, AEN and LRRC58 by 7-Hydroxystaurosporine; C6orf64, HMGCR and RRM2 by Sorafenib; SYS1 and ALAS1 by Venetoclax; and ATF3, CLK1 and CLK4 by Palbocilib. Thus, the pooling chemoproteomics screening provides an efficient method for dissecting the molecular targets of compound libraries.

Computational Reconstruction and Segmentation of Whole Mouse Heart Vasculature

Author: Mia Panlilio, PhD

Co-Authors: Shahinur Alam; Joelle Magne; Nicolas Denans; Aaron Pitre; Aaron Taylor; Douglas Green, PhD; Khaled Khairy, PhD

Abstract: Light sheet fluorescence microscopy (LSFM) enables whole organ and whole organism imaging at cellular and even subcellular resolutions, and represents a major breakthrough in overcoming the traditional trade-off between image resolution and image volume size. It is particularly useful for studies of vascularization, where the structures of interest must be highly resolved, while spanning tissue at the tens of mm scale. However, the technique routinely produces extremely large datasets, terabytes to tens of terabytes in size, which complicates downstream processing. For example, LSFM is able to cover very large volumes by acquiring images of adjacent sub-volumes that need to be assembled seamlessly in a post-acquisition step. Stitching of terabyte large volumes is computationally demanding, and requires specialized hardware and often manual intervention. The latter is especially tedious when the data does not fit into the memory of a typical workstation. Even after stitching, the dataset's size still demands reformatting for efficient visualization and special data traversal strategies for image analysis. 

The above challenges motivated (a) a computational engineering effort to develop and streamline a highly parallelized image post-acquitision processing pipeline that can be executed on high performance compute (HPC)-level hardware with minimal human intervention, and (b) an image analysis effort for fully automated vasculature quantification. For (a), tiled images are stitched using St. Jude's HPC cluster with a modified version of an established volume registration package (BigStitcher). Its output is converted into a hierarchical, chunked file format for efficient multi-threaded read access. Step (b) occurs block-wise along the stitched volume, with filters for background removal, vessel-enhancement, and level-sets segmentation, resulting in a full digital separation of the vasculature from the remainder of the tissue. This is followed by a downstream quantitative vasculature analyses. The pipeline is highly parallel and user-friendly, with even intermediate results readily visualized using common scientific software. For a 0.5 TB volume, it reduces calendar processing time from multiple days to a few hours.

Image Data Management System: Accelerating Scientific Discovery & Collaboration

Author:  Michael Root, MSBME

Co-Authors: Khaled Khairy, PhD; 

Abstract: Research, innovation, and discovery-Science at St. Jude relies substantially on cutting edge imaging technology, that spans data-intense modalities from light and electron microscopy, to MRI, CT, and others. This wide range of modalities is accompanied by a multitude of image data file formats, on-disc data organization conventions, and image data engines (backend tools) optimized for image data access typical of the corresponding field. Such optimal engines and formats are essential for seamless downstream processing and visualization but generate data silos that pose a significant challenge when a more general, and unified, data access path is needed; for example to facilitate collaboration, efficient visualization, computer code consolidation, data discovery, reproducibility, and cross-modality data integration.

In this work, we develop the Image Data Management System (IDMS) that seeks to leverage existing image data engines and data file formats, largely developed and maintained by the biomedical image analysis community, to unify and simplify access and visibility of biomedical image datasets at St. Jude. This effort consists of development of  backend tools to provide a consolidated REST API (Application Programming Interface) for access to image data and metadata. REST-APIs are simple interfaces that make it convenient for researchers to consume image data in a variety of computational environments, computational tools, and programming languages.

Therefore, the IDMS will provide a generic and consistent interface where possible across existing backend tools for ease of image data access.  For users, who wish to use the full power of any backend tool, a Passthru API is provided.  

In either case, the IDMS System, specifically the IDMS Gateway, provides an interface that hides from the user where these backend tools are located and how they are accessed.  In combination with an appropriate user interface (in development), the user conveniently browses their data. Image collections are shown by a unique combination of name, project name, username, and other related metadata tracked by the IDMS database.  It is important to note that the IDMS is designed to provide a thin layer around access to the various image collections and not duplicate image data or the functionality provided by existing tools – simplify, not duplicate.

One other aspect is to furnish an opportunity for adding “plugins” that are managed and callable from the IDMS API.  These plugins will provide programmatic functionality developed by image data scientists and power-users.

Validation of FFPE-derived DNA in target-enriched Enzymatic Methyl-seq profiling for brain tumor classification

Author: Nhu Quynh Tran, PhD

Co-Authors: Sujuan Jia, PhD; Md Zhangir Alom, PhD; Ruth Tatevossian, PhD; Brent Orr, PhD

Abstract: Purpose: DNA-methylation profiling using Infinium Methylation BeadChips has led to significant advances in tumor diagnosis and molecular classification of clinical samples. These microarrays show good performance on both fresh-frozen (FF) and formalin-fixed, paraffin-embedded (FFPE) tissues, but FFPE dominates clinical profiling because it is uniformly produced as part of standard clinical workflows. A significant barrier to adopt DNA methylation arrays in the clinical environment is the requirement of specialized scanners, resulting in a high additional cost and increased laboratory footprint. DNA sequencing-based alternatives are attractive because most clinical molecular pathology laboratories have used sequencers for other molecular assays. To address these clinical needs, we explored the newly developed enzymatic methyl-seq (EM-seq), using the Twist Human Methylome panel. We developed a bioinformatic pipeline to analyze the data and further validated these methods for classification and copy number profiling of brain tumors from FFPE material.

Methods: In this pilot study, we used DNA from FF and FFPE tissues collected from 4 known diagnostic brain cancer patients and 1 control with technical replicates (n=14). DNA libraries were constructed using EM-seq. Methylation detection was targeted to 3.98M CpG sites through 123 Mb of genomic content using the Twist Human Methylome Panel. The EM-seq analysis pipeline was developed based on several open-source tools and packages. To validate the clinical utility of these methylation profiles, the data were used to predict the tumor molecular class using a central nervous system brain-tumor neural net classifier. Copy number data were also evaluated and compared to array-based copy number abnormalities.

Results: The coverage depth ranged from 66 to 346x. Mean per-base coverage for each tumor sample was at least 45x. The CpG calls between FFPE replicates are highly correlated (pairwise correlation coefficients > 0.98). Methylation profiles from 2 different batches were clustered together, indicating consistency and reproducibility. Tumor and control samples were classified into their expected molecular class with high classification scores (> 0.82). Replicate samples with tumor that was not represented in the classifier’s training set yielded subthreshold classification score, supporting high classification specificity. Copy number abnormalities were consistent between sequencing and array-based pipelines.

Conclusions: We successfully developed and validated a protocol to generate and analyze EM-seq data on FFPE samples using the Twist Human Methylome Panel. EM-seq bioinformatic pipeline showed good concordance with existing methods for tumor classification and copy number profiling. This approach has potential to complement existing methods, reducing the barrier to implementation of DNA methylation profiling to more clinical laboratories.

Multi-modality learning with copy number and methylation data from DNA methylation arrays improves classification and risk stratification

Author: Md Zahangir Alom, PhD; 

Co-Authors: Quynh Tran, PhD; Brent Orr, PhD

Abstract: Background: DNA methylation arrays are an important tool for clinically classifying brain tumors.  Methylation-based classification models are trained on differentially methylated probes from a reference series of tumors, and then applied to diagnostic test samples. While copy number profiles are routinely captured from DNA methylation array data, copy number data is reported independent of classification models. 

Certain molecular tumor classes exhibit variability in clinical risk, but current models do not output a within-class risk assessment, analagous to histopathologic grading. Multimodality models have not been established despite the known relationship of specific copy number changes to both tumor class and outcome.

Methods: We established a workflow to extract copy number features from the output of the array-based ‘conumee’ copy number algorithm in order to evaluate the utility of combined multimodality methylation and copy number data for computational modeling. Using a cohort of diffuse gliomas from The Cancer Genome Atlas, we trained deep neural net models on two tasks: tumor classification and survival prediction. We used either methylation-data alone, copy number data-alone, or combined copy number and methylation-data for the models. The accuracy, precision, recall, and F1 score of classification models were compared in cross validation and holdout test sets. The c-index (CI) and prognostic index (PI) were used to compare survival models within and across the tumor classes. 

Results: Our findings indicate that training on copy number data alone is less effective than using methylation or multimodality data for tumor subclass classification (84.7% top accuracy, compared to 95.1 % and 95.83%, respectively). Additionally, our multimodality modeling approach improved survival prediction compared to using DNA methylation or copy number-alone (CI= 86.2, compared to 83.5 and 82.2, respectively). Furthermore, our risk stratification methods accurately demonstrated associative risks within and across the tumor classes.

Conclusions: Our study has shown a systematic approach to integrate copy number data from methylation arrays into clinically relevant computational models. Multimodality models improve model performances, and our survival modeling represents a potential method for molecular grading within tumor classes. 

Integrative analysis of differential expression and differential network for ‘omics data analysis

Author: Yonghui Ni, PhD

Co-Authors: Prabhakar Chalise, PhD; Jianghua He, PhD; 

Abstract: Differential expression (DE) analysis has been commonly used to identify molecular features that are statistically significantly different between distinct biological groups. Differential network (DN) analysis explores molecular network structure changes from one biological condition to the other. Both approaches are very important and complementary to each other. However, they have often been considered separately in ‘omics data analysis.  We developed an integrative method, DNrank, to perform joint DE and DN analysis by incorporating modified Google's PageRank algorithm framework. The purpose of DNrank is to identify the disease-associated features from feature ranking technics. The resampling based cross-validation scheme within DNrank algorithm optimizes the feature ranking by leveraging DE level and DN structure of the features. We illustrate DNrank using both simulated data and three real life data. DNrank shows a better performance in identifying important molecular features with respect to predictive discrimination. Also, as compared to existing feature selection methods, the top-ranked features from DNrank outcome have a higher stability in selection. DNrank allows the researchers to identify disease associated key features accounting for both DE and DN.

Ontology-guided navigation of somatic variants, mutational signatures, gene expression and histology images for pediatric cancer

Author: Alex Gout, PhD

Co-Authors: Stephanie Sandor, MS; Delaram Rahbarinia, MS; Jobin Sunny, MS; Lucian Vacaroiu; James Madson; Kevin Benton, MS; Michael Macias, MS; Samuel Brady; Wentao Yang, PhD; Jian Wang, PhD

Abstract: PeCan knowledge base on St. Jude Cloud (pecan.stjude.cloud), initially developed as a resource of curated genomic variants of pediatric cancer (PeCan), is now significantly expanded to comprise a hub of interconnected data facets for over 9,000 hematologic (heme), CNS, and non-CNS solid (solid) tumor patient samples from around the world. The new data facets, which include gene expression, mutational signatures, and histology, can be explored alongside our existing variants data facet inspiring new hypothesis generation. Variants shows a genomic landscape view (i.e. oncoprint view) and gene -or genome-level view (i.e. ProteinPaint view). Expression data, generated from normalized RNA-seq of ~3,000 samples, can be explored via interactive 2-dimensional maps, revealing distinct subtypes relevant for patient stratification for precision therapy. Mutational signatures, identified from ~2,000 WGS/WES samples, are presented as a heatmap across subtypes, in addition to a summary view for a user-defined cohort or individual sample for which we also display a mutation profile frequency plot together and identified signatures. Histology enables review of histological slide images and associated clinical notes for ~3,000 solid tumors via a searchable interface. As all samples on PeCan have been mapped to a custom pediatric cancer classification based ontology, the user can customize the view of each data facet presented in PeCan by selecting a specific cancer subtype. Integrative analysis between the data facets has enabled new insights into pediatric cancer biology as demonstrated in the following two examples. First is the discovery of two potential subtypes of adamantinomatous craniopharyngioma. These were initially identified via expression analysis which revealed two distinct groups that were confirmed by examination of associated histology slide data which revealed delineation by brain invasion. Second is an analysis that computationally identifies homologous recombination-deficient (HRD) pediatric tumors using data from the mutational signatures data facet. This provides new insights on the applicability of therapies targeting HRD in the pediatric  cancer population. These examples demonstrate the potential value of PeCan in advancing clinical diagnosis classifications of pediatric cancer and exploration of new therapeutic opportunities. PeCan is an evolving knowledge base, as we are continuously expanding the platform and adding data over time to foster scientific discovery for the global research community, with the goal of improving treatments for pediatric cancer. 

DeepMRIRec: Deep-learning-based four-fold acceleration of MRI data acquired with RT coil configurations

Author: Shahinur Alam

Co-Authors:  Jinsoo Uh; Chia-Ho Hua, PhD;  Khaled Khairy, PhD; 

Abstract: Magnetic Resonance Imaging (MRI) is a powerful technique to discover and monitor neurological, musculoskeletal, and oncological diseases. However, MRI is slow and stressful for patients, who must remain motionless in a prolonged MRI imaging procedure that prioritizes reduction of imaging artifacts and an increase in data quality. Prolonged imaging time is particularly challenging for pediatric patients and those that are already vulnerable due to their health conditions. It is therefore important to seek methods that reduce imaging time. One approach is to record fewer measurements and then digitally recover the full information from this limited set of measurements in a post-acquisition reconstruction step. Towards that end, we develop a deep learning-based method called DeepMRIRec for MRI reconstruction from highly under-sampled raw data acquired with RT-specific receiver coils. We demonstrate our method on fully sampled k-space data of T1-weighted MRI acquired from 73 pediatric brains with tumors/surgical beds using loop and posterior coils (maximum 14 channels). We thoroughly evaluated DeepMRIRec and state-of-the-art deep learning-based reconstruction methods using our dataset with/without applying transfer learning from publicly available models. DeepMRIRec reduces image acquisition time by a factor of four, surpassing previous methods (structural similarity (SSIM) score of 0.95, PSNR of 33). Although our model outcome is satisfactory and outperforms state-of-the-art approaches, it still produces images that suffer from significant smoothing and will need further improvement. Importantly, our approach enables significant reduction in acquisition time for children undergoing therapy, and especially those that undergo routine MRI. As our method improves, we plan to perform a thorough evaluation of its effectiveness in the field.

Predicting Cellular Condensation by Fusion Oncoproteins

Author: Swarnendu Tripathi

Co-Authors: Hazheen Shirnekhi, PhD; Bappaditya Chandra, PhD; David Baggett, PhD; Cheon-Gil Park; Benjamin Lang; Hadi Hosseini; Brittany Pioso; Snigdha Maiti; Aaron Phillips; Jina Wang

Abstract: Fusion oncoproteins (FOs) that arise from in-frame chromosomal translocations are drivers of many aggressive pediatric cancers. Several FOs are known to promote oncogenesis by undergoing phase separation (PS) to form aberrant biomolecular condensates linked with oncogenesis. To test the generality of PS, 166 FOs derived from diverse cancers were expressed in HeLa cells, with 58% identified to form condensates. The condensate forming FOs displayed physicochemical features different from those that displayed diffuse localization. Therefore, we hypothesized that distinct patterns of physicochemical features encoded in the amino acid sequences of the FOs could be used to predict their cellular condensation behavior. To test this hypothesis, we applied supervised automatic Machine Learning (ML) and a Gradient Boosting Machines model with 65 trees performed best amongst the 220 models tested. The model accuracy on the cross validated training set of 149 FOs (96 condensate-positive and 53 condensate-negative) was 0.79. We experimentally tested 12 additional FOs for condensate formation in cells and obtained 0.92 predictive accuracy from the ML model. We determined the contributions of the physicochemical features for prediction of FO condensation behavior using a Shapley Additive exPlanations (SHAP) algorithm based on the game theoretically optimal Shapley values. SHAP analysis was used to uncover how physicochemical features contributed to ML predictions, which guided mutagenesis experiments to rationally reverse the behavior of 9 condensate forming FOs. Our ongoing studies seek to understand how the LLPS-prone intrinsically disordered regions within FOs contribute to their cellular condensation behavior and promote oncogenesis.

Repulsive Surfaces for Modeling Complex Biomembrane Morphologies

Author: Meghdad Razizadeh, PhD

Co-Authors: Khaled Khairy, PhD

Abstract: Lipid bilayers are essential components of biological membranes, playing a crucial role in maintaining the structural integrity of cells and organelles. The study of lipid bilayer vesicles, as models for biological membranes, has been an active area of research for the past four decades. The computational modeling of vesicles strives to relate morphology, membrane structure, and membrane energetics. The Canham-Helfrich energy functional and its variants represent a well-established approach to computationally investigate the mechanics and dynamics of lipid bilayers in terms of local principal curvatures of the surface. However, even with the help of such powerful formulations, simulation of lipid vesicles, complex biomembranes, and realistic membrane-bound organelles such as Golgi, ER, or mitochondria, still pose significant computational challenges. For example, the evaluation of bending forces requires approximations of derivatives of surface coordinates up to the fourth order, which can lead to numerical challenges on discretized shapes. Importantly, published works on computational membrane modeling lack reliable mechanisms to prevent self-intersections, making it challenging to extend computational predictions of biomembrane shapes to more intricate highly convoluted forms. This issue is especially prevalent at high surface area to volume ratio shapes, observed for many cell organelles, where minimized shapes can become unphysical due to self-intersections. In this study, we adopt a recently developed method for self-intersection prevention (repulsive surfaces), originally developed for computer graphics applications. We extend its application to model realistic biomembranes by adding a tangent point energy term to the Helfrich functional. Our method yields shape predictions, even at high surface-to-volume ratios, a regime that cannot be reliably explored without self-intersection treatments. Finally, our study demonstrates the ability of this model to predict complex biomembrane morphologies, where lipid bilayers are forced to form tubular or planar shapes; features commonly observed in cell organelles, for example, the ER and inner mitochondrial membrane.

Defining the Condensate Landscape of Fusion Oncoproteins

Author: David Baggett, PhD

Co-Authors: Swarnendu Tripathi, PhD; Hazheen Shirnekhi, PhD; Brittany Pioso; Bappaditya Chandra, PhD; Richard Kriwacki, PhD

Abstract: Fusion oncoproteins (FOs) occur in ~17% of cancers and are often drivers of oncogenesis through one of two molecular mechanisms.  Many FOs promote oncogenesis by changing gene expression, while others alter cellular signaling. Notably, formation of aberrant biomolecular condensates is implicated in the oncogenic mechanism of  several FOs .

To better understand the relationship between FOs and condensate formation, we compiled a database of >3,000 FO sequences (FOdb). We then transfected HeLa cells with mEGFP-tagged forms of 166 FOs selected from the FOdb and found that 58% formed condensates.  These studies show that FOs that formed condensates have physicochemical feature patterns different from those that don’t.  We find that these patterns also correlate with sub-cellular localization.  Further, using conserved domain annotation, we found that FOs localized to the nucleus contained functional terms predominantly associated with regulation of gene expression, while those localized to the cytoplasm  associated with cell signaling terms. 

We then trained a machine learning model that analyzes a FO sequence to accurately predict if it will form biomolecular condensates in cells. This model predicts that 59% of the untested FOs in the FOdb form cellular condensates, many of which have physicochemical features indicative of gene regulation and nuclear localization.  Conversely, a substantial portion of FOs predicted to not form condensates have sequence features that suggest cytoplasmic localization and cell signaling functions.  We conclude that most FOs that drive aberrant gene transcription function in part through condensate formation, and those that alter cell signaling do so without forming condensates. 

Epigenetic regulators TET2 and polycomb repressive complex 2 (PRC2) coordinate gene expression in Myelodysplastic Syndrome (MDS)

Author: Jacquelyn Myers, MS

Co-Authors: Elodie Henriet, PhD; Ana Leal-Cervantes, PhD; LaShanale Wallace, PhD; Amina Metidji, PhD; Jordan Skorupa, BS; Ilaria Iacobucci, PhD; Shondra Pruett-Miller, PhD; Yiping Fan, PhD; Esther Obeng, MD, PhD

Abstract: Myelodysplastic syndrome (MDS) is a pre-leukemic disorder arising from somatic driver mutations that cause the expansion of mutant hematopoietic stem cells (HSCs) and ultimately transformation into Acute Myeloid Leukemia (AML). Epigenetic regulators including TET2 are frequently mutated in MDS. TET2 loss of function mutations cause HSC self-renewal resulting in myeloproliferation and extramedullary hematopoiesis. We know TET2 plays a role in active DNA de-methylation, but how disruption of this process mechanistically results in AML transformation remains unknown. We hypothesize that TET2 loss causes gains in DNA methylation triggering an aberrant epigenetic cascade that dysregulates gene expression in HSCs.  

We collaborated with the St. Jude CAGE Core to develop TET2 knock out K562 cells to evaluate the effects of TET2 loss on DNA methylation and transcription. We identified >5,000 regions that gain DNA methylation upon TET2 loss that reside within intronic regions that are repressed. These gains in DNA methylation occur within genes that are up regulated upon TET2 loss. Taking these data together, we hypothesize that polycomb repressive complex 2 (PRC2) recruitment is blocked upon TET2 loss resulting in gene activation. To test this, we conducted H3K27me3 CUT&RUN and identified a significant loss of H3K27me3 at up-regulated genes. Additionally, we evaluated gene expression changes in TET2-mutant AML samples and identified a conserved set of gene candidates that are putatively regulated by DNA methylation and PRC2. These findings begin to unravel an epigenetic regulatory mechanism responsible for maintaining normal gene expression patterns in HSCs.  

Computational design of multi-specific chimeric antigen receptors

Author: Kalyan Immadisetty, PhD

Co-Authors: Jaquelyn Zoine; Vikas Trivedi; Stephen Gottschalk, PhD;  Giedre Krenciute, PhD; Paulina Velasquez, PhD; M. Madan Babu, PhD; Michaela Meehl

Abstract: Chimeric antigen receptor (CAR) T-cell immunotherapy has the potential to revolutionize the treatment of pediatric brain tumors and leukemias. However, insufficient T cell activation and/or immune escape remain significant obstacles. To address these challenges, bispecific CARs were designed that recognize two antigens IL13Ra2 and/or B7H3, both expressed in pediatric brain tumors. Initial attempts to generate tandem CARs utilizing two single chain variable fragments (scFv) as a bispecific antigen domain resulted in a lack of cell surface expression of the CAR. To test if computational evaluation of structural dynamics of scFv position can predict and/or improve cell surface expression, we developed a computational approach comprising of the state-of-the-art structure prediction tool AlphaFold2. By optimizing the linker length, type, and flexibility, we rationally designed several novel tandem and loop (diabodies) bispecific IL13Ra2-B7H3 CARs. Preliminary testing of the three tandem and one loop CARs in HEK cells showed modest cell surface expression (~20%) for the tandem CARs, while the loop CAR showed nearly full cell surface expression (~90%). We then applied this strategy to optimize several peptide-scFv bispecific including GRP78-B7H3 CARs and to design a novel trispecific peptide-scFv-scFv construct (GRP78-B7H3-CD123 CARs). Validation of their cell surface expression and antitumor activity against acute myeloid leukemia is currently underway. In summary, we have designed an innovative computational approach for multi-specific CAR generation and this strategy can be applied to a broad range of pediatric cancers requiring novel immunotherapeutic options.

Cancer dependency prediction using deep learning

Author: Hadi Hosseini, PhD

Co-Authors: Duccio Malinverni, PhD; Balaji Santhanam, PhD; Bálint Mészáros, PhD; M. Madan Babu, PhD

Abstract: Recent advances in genomic technologies have allowed for determining the genetic dependencies (i.e., gene essentiality) in various cancer cell lines. Cancer dependency is determined by comparing the growth behaviors of cells with and without the knock-out of the target gene of interest. Large-scale efforts have allowed us to determine the dependency of 18,000 genes in ~900 cell lines, resulting in the creation of the DepMap portal. This resource has created unprecedented large-scale opportunities to study cancer vulnerabilities at the genomic level. Despite the immense potential of the DepMap dataset to understand cancer dependencies and vulnerabilities at the cell line level, bringing the power of cancer dependency to personalized medicine requires overcoming significant practical barriers. Determining the dependency profile of a specific patient would require performing ~18K knock-out experiments on all human genes and corresponding cell-growth assays, which is currently out of reach of routine clinical settings. To overcome these limitations, we propose to build an AI-based approach to predict cancer dependency at the individual patient level. Our approach leverages the availability of the DepMap dataset and combines it with the Cancer Cell Line Encyclopedia (CCLE) and protein-protein interaction networks, which are publicly available in the STRING, IntAct, and BioGRID portals.

We are developing a deep learning model for OMICs data to identify novel cancer-dependent genes and make predictions on their role in disease. The question is how to predict cancer dependency using bulk RNA-seq expression and protein-protein interaction data. Predicting the cancer dependencies of the entire human genome will be useful for further research, such as target discovery and designing specific drugs that modulate cancer-dependent genes. We are designing a graph-based neural network (GNN). This model can consider the functionality of several genes together, the assumption being that if the genes interact, they function together. The advantage of GNN models is that they retain the relationship between nodes in each layer. As genes work together in a gene network, the relationship between genes is important for prediction quality. The GNN will predict cancer dependency based on particular gene-expression profiles together with accurate uncertainty estimates.

We aim to apply this model to all the St. Jude pediatric transcriptomics datasets to investigate the resulting genes and determine target dependency. We will specifically focus on AML, T-ALL, and B-ALL using the St. Jude pediatric transcriptomics dataset.

Computational approach for discovery of molecular glues: Expanding the Repertoire of Neo-Substrates for Targeted Protein Degradation 

Author: Vyoma Sheth, BDS, MS

Co-Authors: Besian I. Sejdiu, PhD; Gisele Nishiguchi, PhD; Balaji Santhanam, PhD; Benjamin Leslie, PhD; Zoran Rankovic, PhD; M. Madan Babu, PhD

Abstract: Drug development is not only difficult, time-consuming, and expensive but an inherently unpredictable process. The discovery and validation of an appropriate drug target is crucial to the success of drug discovery efforts. Still, this task is challenging given the fact that approximately 85% of proteins are considered "undruggable." Our aim is to overcome this limitation by adopting the strategy of “Targeted Protein Degradation (TPD)” to target the undruggable proteins. In TPD, a small molecule (e.g., molecular glue) brings a target of interest in proximity to an E3 ubiquitin ligase, thereby ubiquitinating the target (neo-substrate) and degrading it. A critical aim of our project is to develop an integrative data-driven approach to identify lead compounds that can be developed into molecular glues to target therapeutically relevant neo-substrates. This approach provides a scope to expand the repertoire of “druggable” proteins. 

To identify potential molecular glues for pediatric leukemia and medulloblastoma*, we initiated our research by focusing on the C2H2 zinc finger family of transcriptional regulators, as these diseases are characterized by the deregulation of several transcription factors containing C2H2 zinc finger motifs. We considered the structure of the DDB1-CRBN- C2H2 ZNF domain of IKZF1 bound to pomalidomide as a reference starting point (PDB ID: 6H0F). Our structural bioinformatics analyses enabled us to define the signatures of small molecules from the perspective of CRBN and the neo-substrate to predict C2H2 proteins that can be targeted via molecular glue strategy. We developed a virtual screening pipeline that consists of screening the Chemical Biology and Therapeutics (CBT) library (~ 600k compounds) and the enamine library of Molecular Glues (~ 5k compounds). Our pipeline was benchmarked using the above-mentioned reference structure and discriminatory analyses to separate potential binders from non-binders. Further post-docking analysis was conducted on the screening hits to validate novel binders.

We found small molecule hits with less chemical similarity to thalidomide and its analogs while maintaining the same binding pose and higher binding energy. We also rigorously benchmarked our approach as this enhances confidence in compound discovery, which could be prioritized for further experimental validation and investigations. 

The next step is to create a web portal that would disseminate our findings with stratified confidence-score levels for various potential molecular glues against any desired C2H2 zinc finger target. We hope this accelerates future experimental efforts to target therapeutically relevant novel C2H2 zinc fingers. 

*This project is part of the Crazy8 initiative titled “Small Molecule Degraders for Targeting Transcription Factor Drivers of Childhood Cancers” that involves the groups of Charles Mullighan, Zoran Rankovic, M. Madan Babu, Marcus Fischer, Jeffery Klco, Paul Northcott,  and Martine Roussel

A transmission model of avian influenza in central Chile

Author: Lauren Lazure

Co-Authors: Pedro Jimenez-Bluhm, PhD; Christopher Hamilton-West, PhD; Stacey Schultz-Cherry, PhD

Abstract: Select avian influenza virus (AIV) strains continue to cause significant morbidity and mortality worldwide in mammals and birds. A high diversity of AIV subtypes has been identified over central Chile over the past decade, including highly pathogenic avian influenza viruses. Migratory waterfowl are likely the source of novel AI subtypes in the region as central Chile is a location along several migratory bird flyways that link the Americas. This region also contains a high number of backyard poultry and swine farms that are that are susceptible to spillover infections and pose a risk for transmission to other mammals including humans. The purpose of this study is to determine factors that contribute to the AIV prevalence and determine transmission dynamics between migratory birds, non-migratory wild birds, and domestic poultry through agent-based modelling. The development of this model and outcomes of these studies are critical for pandemic preparedness. 

St. Jude Survivorship Portal: A data portal for storing, analyzing, and sharing large and complex cancer survivorship datasets

Author: Gavriel Matt

Co-Authors: Edgar Sioson; Jian Wang, PhD; Congyu Lu; Airen Zaldivar; Karishma Gangwani; Alex Acic; Jaimin Patel; Robin Paul; Colleen Reilly; Kyla Shelton; 

Abstract: Background/Purpose: Cancer survivorship research relies on large-scale cohort studies that inform a wide range of outcomes after the completion of cancer therapy along with demographic, genetic, and clinical data. To maximize the utility of these comprehensive datasets, we must be able to store and share this data in a web-based environment that can be accessed by the broader survivorship research community. Furthermore, this environment should be integrated with analytical tools for performing statistical analyses on the stored data without needing to download the data and import it into third-party analytical software. To address this need, we have created the St. Jude Survivorship Portal (https://survivorship.stjude.cloud), a web-based data portal for exploring, sharing, and analyzing data from survivors of pediatric cancer.

Methods: The portal hosts data collected from two large cohorts of pediatric cancer survivors, the St. Jude Lifetime Cohort (SJLIFE) and the Childhood Cancer Survivor Study (CCSS). All SJLIFE survivors (5,053) were included in the portal. For the CCSS cohort, only those survivors with whole genome sequencing data (2,688) were included, with a future plan for expansion to include all CCSS participants.

Results: The portal hosts both the phenotypic and genetic data of survivors. Phenotypic data include demographic data (e.g., sex, age, race/ethnicity, education, employment), clinical data (e.g., cancer diagnosis, cancer treatment exposures/doses, clinically assessed and survey-based chronic health conditions), and patient-reported data (e.g., healthcare utilization, insurance, quality-of-life assessments). Genetic data include whole-genome-sequencing-derived genotypes for over 160 million variants and polygenic risk scores computed for thousands of traits. This data can be easily explored by the user through a hierarchically organized data dictionary for phenotypic data and a genome browser for genetic data. Charts and plots of variables can be quickly created, customized, and downloaded, all within the portal environment. Statistical analyses, including cumulative incidence analysis and regression analysis, may also be performed within the portal. In cumulative incidence analysis, users can analyze the incidences of various CTCAE-graded adverse events (e.g., cardiovascular dysfunction, neurological disorders, subsequent neoplasms) in survivors. In regression analysis, users can perform either linear, logistic, or Cox regression analysis and can assign any of the variables on the portal as either outcome or explanatory variables. Lastly, we permit users who obtain necessary permission to download the data from the portal for use in local in-depth analyses. The St. Jude Survivorship Portal provides a comprehensive, powerful, and easy-to-use data portal for sharing and analyzing childhood cancer survivorship data that will serve as a valuable research tool for the broader survivorship research community.

Tug-of-war between sticker and zipper interactions in network fluid condensates determines the kinetics of amyloid fibril formation

Author: Tapojyoti Das, MBBS, MMST, PhD

Co-Authors: Fatima Zaidi, PhD; James Messing, PhD; J. Paul Taylor, MD, PhD; Tanja Mittag, PhD

Abstract: Recurrent mutations in the C-terminal low-complexity domain of the stress granule protein hnRNPA1 (A1-LCD) lead to familial amyotrophic lateral sclerosis and a spectrum of related neurodegenerative diseases. These mutations create strong steric zipper motifs that drive fibrillization in vitro and lead to poorly dynamic stress granules in cells. A1-LCD also phase separates in vitro, driven by multivalent sticker-sticker interactions between aromatic residues. Phase separated condensates of hnRNPA1 mutants and some other RNA-binding proteins give rise to fibrils over time. Together with genetic and cell biological data, this suggests that stress granules are crucibles for neurodegenerative disease. We find that disease mutants of A1-LCD have a lower driving force for phase separation than the wild-type and higher fibril stability, reflecting a higher thermodynamic driving force for conversion of condensates to fibrils. Our data further revealed that nucleation is promoted at the condensate interface, but subsequent fibril growth is kinetically inhibited in wild-type condensates. Increasing the aromatic sticker strength in a disease mutant background increased phase separation propensity and markedly slowed down fibril formation in vitro, reflecting a molecular tug-of-war between the two processes. Moreover, using a custom pipeline of automated stress granule detection and analysis of live-cell imaging data, we found that strong stickers in a disease mutant background rescued the disease mutant phenotype of delayed stress granule dissolution in cultured human cells. A1-LCD constructs with strong stickers can prevent fibrillization also in trans, indicating therapeutic potential. Our work suggests that stress granules may suppress fibril formation of constituent proteins in the short term but can become a liability in the presence of fibrillization-prone disease mutants.

JUMP Software Suite (JUMPsuite): An Extensive and Modularized Toolbox for Large-scale Mass-spectrometry-based Proteomics and Metabolomics Data Analysis

Author: Xusheng  Wang, PhD

Co-Authors: Zuo-Fei Yuan, Yingue Fu, Suresh Poudel, Abjihit Dasgupta, Yuxin Li, Ji-Hoon Cho, Handy High, Vishwajeeth Pagala, Haiyan Tan, Ashutosh MIshra, Kiran Kodali, Suiping Zhou, Kaiwen Yu, Huan Sun, Junmin Peng

Abstract: Mass spectrometry-based proteomics and metabolomics are increasingly powerful techniques for identifying and quantifying proteins and metabolites in a wide range of biological samples. However, the analysis of large proteomics and metabolomics datasets requires the use of advanced computational tools and strategies. To address this need, we present the JUMP software suite (JUMPsuite), an extensive and modularized platform specifically designed for processing, analyzing, and interpreting large-scale mass spectrometry-based proteomics and metabolomics data. JUMP not only enables the accurate identification of peptides but also the quantification of differentially expressed proteins across multiple experimental conditions, providing efficient and comprehensive data interpretation capabilities. Additionally, the software suite offers stringent false discovery rate control mechanisms, ensuring the utmost accuracy and reliability in protein identification and quantification processes. The sophisticated clustering and network analysis features within JUMP aid in uncovering functional modules and pathways present in the biological samples under investigation. Moreover, JUMP can effectively prioritize candidate biomarkers and potential therapeutic targets based on their functional relevance to specific diseases. In parallel, we have also developed the JUMPsuite to process metabolomics data. Importantly, JUMPsuite is subject to consistent, ongoing enhancements, including the integration of novel analytical methods to expand its existing functionalities and its deployment on the St. Jude Cloud platform. Collectively, JUMPsuite serves as a valuable resource for the proteomics and metabolomics research communities, delivering a comprehensive, all-in-one solution for processing and analyzing large-scale mass spectrometry datasets.

Exploring homologous recombination deficiency in pediatric high-grade gliomas

Author: Corey Xu, PhD

Co-Authors: Evan Savage; Jason Myers; Evadnie Rampersaud; Ti-Cheng Chang; Gang Wu; Anang Shelat; Christopher Tinkle, MD, PhD

Abstract: Homologous recombination deficiency (HRD) has been extensively studied in breast cancer and ovarian cancer given its correlation with sensitivity to Poly (ADP-ribose) polymerases (PARP) inhibitors. In these cancers, HRD is most frequently caused by biallelic inactivation of BRCA1/2 and tumors often exhibit ‘genomic scars’ such as elevated HRD.sum score and specific DNA mutational signatures, including specific single base substitution (SBS), small insertion/deletion (ID), and structural rearrangement (RS) signatures. Despite its clinical relevance, HRD has not been systematically studied in pediatric cancers and its therapeutic predictive value is unknown. Here, we evaluated HRD and genomic scars in 277 paired tumor and normal samples of pediatric high-grade glioma (pHGG) profiled with whole-genome sequencing (WGS) and 23 tumor-normal paired breast cancer (BC) WGS samples as controls. We profiled variants from over 500 DNA damage response genes and found that the most frequently mutated genes with biallelic loss were TP53, followed by ATM, an essential gene for genome stability. We found that 11% of the samples have HRD.sum score greater than 40, an indicator of HRD in several adult cancers. However, while mutational signature analyses identified HRD signature SBS3 in 25% of the pHGG samples, other HRD-related signatures, including ID6, RS3, or RS5, were not identified. In comparison, we identified SBS3 and ID6 in all BC samples with BRCA1/2 loss. Furthermore, we computed two recently reported composite HRD scores, i.e., HRDetect and CHORD, on the pHGG samples and BC samples. While all BC samples with BRCA1/2 loss were predicted to be HRD by both scores, none of the pHGG samples were predicted to be HRD by either score. Finally, analysis of RNA-seq data from a subset of pHGG samples (n=90) revealed that samples with elevated HRD.sum scores showed upregulation of pathways related to homology-directed DNA repair. Overall, our study suggests that HRD exists in a fraction of pHGG samples, but it is likely driven by variants in genes other than BRCA1/2, resulting in a different set of genomic scars compared to those in breast cancers. Functional analysis of sensitivity to PARP inhibitor and related DNA damage inhibitors is ongoing to evaluate potential therapeutic relevance of these genomic scars.

Radiomic features from apparent diffusion coefficient maps provide additional prognostic value in rhabdomyosarcoma.

Author:  Silu Zhang, PhD

Co-Authors: Chris Goode; Diana Storment; Zachary Abramson, MD; Matthew Krasin, MD; Alberto Pappo, MD; Mary McCarville, MD; Matthew Scoggins, PhD; 

Abstract:  Rhabdomyosarcoma (RMS) is a rare cancer that arises from skeletal muscle tissue and mainly affects children and young adults. The 5-year overall survival (OS) rate for RMS is about 70%. Currently, prognosis depends on clinical factors including age, disease stage, histologic subtype, and tumor location. However, about 30% of patients with intermediate-risk RMS patients experience recurrence with local or metastatic disease and have a poor prognosis. Thus, identifying imaging features that can accurately predict clinical outcome could help improve risk stratification. Radiomics is a technique that extracts quantitative features from diagnostic medical images that are undetectable to the human eye. In this study, we aimed to identify radiomic features on routine MRI apparent diffusion coefficient (ADC) maps that might provide prognostic value in RMS. 

Materials and Methods: We analyzed baseline MR imaging of 67 subjects enrolled in the RMS13 protocol (NCT01871766). A tumor volume of interest (VOI) was manually segmented and approved by a radiologist. Additionally, we generated a tumor-tissue interface VOI by dilating and then subtracting the tumor VOI. Radiomic features were extracted from both the tumor mask and the tumor-tissue interface mask, resulting in a total of 2446 features. We included 6 clinical variables in the statistical analysis: age at diagnosis, race, sex, primary site category (13 sites), clinical risk group (low, intermediate, high-oberlin-low, and high-oberlin-high), and FOXO3 gene translocation status. 

Results: Age at diagnosis, risk group, and translocation status were associated with 3-year OS. When controlling for these factors, 46 radiomic features were associated with 3-year OS. Specifically, larger tumor size, larger ADC range within a single tumor, and other features related heterogeneity were associated with worse outcome. Interestingly, 27 out of the 46 features associated with outcome were identified from the interface mask than from the tumor mask, suggesting potential useful information from the interaction between the tumor and host tissue. 

Conclusions: Although further studies are needed to validate these features, radiomics features extracted from routine clinical MRI may provide important prognostic information to guide treatment and improve outcomes.

Identifying disease-relevant signaling sites by studying their structural evolution using AlphaFold-based predictions across 200 species

Author: Benjamin Lang

Co-Authors: Bálint Mészáros, PhD; Besian I. Sejdiu, PhD; Duccio Malinverni, PhD; M. Madan Babu, PhD

Abstract: Post-translational modifications (PTMs) can dynamically diversify the proteome in response to intracellular and extracellular signals. They are a fundamental means of biological information processing with important functions in development, homeostasis and disease. Since hundreds of thousands of modified residues as well as entirely new modification types are being discovered in proteins, elucidating their biological functions and how they evolve is a fundamental problem. I will present a study of the evolution of modified amino acids in human proteins in their structural context. Using sequence, polymorphism, mutation, and structural data at the species, population and individual levels, we identified significant evolutionary constraints. We also observed an overrepresentation of amino acids which mimic modified residues at orthologous positions, representing clues to their evolution and structural roles.

AlphaSync (http://mbsql.stjude.org/alphasync/) is a still-unpublished internal St. Jude data processing pipeline and web server that enriches DeepMind’s AlphaFold structural database with residue-level accessible surface area and residue–residue contact data. It is also intended to synchronize the infrequently updated public AlphaFold database with the latest UniProt release, which it does by re-predicting the structures of proteins with updated sequences using the St. Jude high-performance GPU computing cluster. Importantly, this pipeline also includes prediction of intrinsically disordered residues using their relative solvent-accessible surface area, which is a recently pioneered approach, as well as building on a novel contact detection Python package from the Babu lab. A MySQL database server serves as the data repository for our refined version of AlphaFold DB. We envision that AlphaSync will be helpful to computational biologists at St. Jude and beyond.

By tracing the structural evolution of human post-translationally modified residues across 200 closely and distantly related species from the Ensembl Compara comparative genomics pipeline using AlphaSync, we aim to identify disease-relevant signaling sites that exert a high structural impact when modified or mutated.

Global Depths for Irregularly Observed Multivariate Functional Data

Author: Zhuo Qu, PhD

Co-Authors: 

Abstract: Two frameworks for multivariate functional depth based on multivariate depths are introduced in this paper. The first framework is multivariate functional integrated depth, and the second framework involves multivariate functional extremal depth, which is an extension of the extremal depth for univariate functional data. In each framework, global and local multivariate functional depths are proposed. The properties of population multivariate functional depths and consistency of finite sample depths to their population versions are established. In addition, finite sample depths under irregularly observed time grids are estimated. As a by-product, the simplified sparse functional boxplot and simplified intensity sparse functional boxplot are proposed for visualization without data reconstruction. A simulation study demonstrates the advantages of global multivariate functional depths over local multivariate functional depths in outlier detection and running time for big functional data. An application of our frameworks to cyclone tracks data demonstrates the excellent performance of our global multivariate functional depths.

St. Jude Neuro-Oncology Data Portal delivers genomic and clinical information from a large pediatric medulloblastoma cohort

Author: Congyu Lu, PhD

Co-Authors: Sandeep Dhanda; Edgar Sioson; Airen Zaldivar; Karishma Gangwani; Gavriel Matt, PhD; Colleen Reilly; Alex Acic; Giles Robinson, MD; Xin Zhou, PhD; 

Abstract: Background/purpose: Medulloblastoma is the most common malignant brain tumor of childhood and is now known to be a highly heterogeneous disease. It is comprised of diverse molecular subgroups and subtypes, each of which is associated with different clinical characteristics, genetic alterations, risk factors, and survival outcomes, necessitating a comprehensive and multidimensional approach to their investigation. In addition, advances in high-throughput genomic analyses have yielded an unprecedented amount of genetic alterations in medulloblastoma tumors, making analysis across multiple variables challenging. Integrating various types of genetic alterations with phenotypic data and presenting them in easily accessible and intuitive visualizations are essential to extrapolate novel insights from this vast dataset. To this end, we developed Neuro-Oncology portal, a web application for multidimensionally visualizing and analyzing data from medulloblastoma patients.

Methods: The portal hosts data collected from three studies undertaken in 898 patients with medulloblastoma (SJMB03, ACNS0331, and ACNS0332). DNA methylome, whole-exome and transcriptome sequencing are available for most of the tumors.

Results: The portal presents demographic, clinical, and genomic data from 898 medulloblastoma patients as hierarchically organized variables inside an interactive data dictionary.  Variables are searchable and can be visualized and analyzed by multiple plots. The summary plots (bar chart and violin plot) visualize the distribution of a given variable. The survival plot presents survival outcomes of the represented patients. The matrix plot provides a compact graphical representation of the overall annotations and genomic alterations from the patients and their samples. The tSNE plot visualizes the tumor clustering patterns based on methylomic state, where user can lasso a cluster of interest and save it as a new sample group to be explored by other plots. Each of the plots is embedded with a multitude of point-and-click functions to facilitate the visualization and analysis. By using the variables jointly in the plots, users can stratify the samples based on simultaneous consideration of these variables and perform multidimensional exploration and integrated analyses. To facilitate the filtering of the samples, we implemented a filtering system, within which any of the variables can be utilized as a filter, and multiple filters can be combined with AND/OR operators as one single filter. The portal also enables users to create novel sample groups using the filtering system and compare the sample groups they have created.

Conclusion: The St. Jude Neuro-Oncology Data Portal provides the research community a powerful tool to multidimensionally visualize and analyze data from medulloblastoma patients. The portal will be made freely accessible upon manuscript publication.