Biology subjects

Stolovitzky, G.

Publications and source records attributed to Stolovitzky, G..

14 recordsLinked to original sources

bioRxiv2026

The SEA-AD DREAM Challenge: Community benchmarking human and AI agent solutions for Alzheimer's disease neuropathology prediction from single-nucleus transcriptomics

Single-nucleus transcriptomic atlases offer an unprecedented opportunity to connect cellular molecular states with Alzheimer's disease (AD) neuropathology, but whether these profiles encode reproducible, predictive information about pathological burden remains unclear. We present the SEA-AD DREAM Challenge, an open, international, model-to-data competition built on the Seattle Alzheimer's Disease Brain Cell Atlas to predict Alzheimer's disease neuropathological severity from single-nucleus RNA-sequencing data. Participants developed containerized models to predict categorical neuropathological staging, including overall Alzheimer's disease neuropathologic change, Braak stage, Thal phase, and CERAD score, as well as quantitative amyloid-{beta} and phospho-tau burden measured by 6E10 and AT8 immunohistochemistry. Across 17 eligible teams from 15 countries, the crowdsourcing framework enabled systematic comparison of diverse computational approaches and surfaced a broad landscape of modeling strategies and candidate predictive features. Top-performing methods achieved near-perfect prediction of categorical staging, with the best submission reaching a quadratic weighted kappa of 1.0 for the Overall AD Neuropathological Change score (ADNC), and competitive prediction of quantitative pathological burden in held-out data, with a best concordance correlation coefficient of 0.48. Post hoc perturbation analyses revealed that top categorical-stage predictions relied heavily on donor-level metadata-driven signals rather than transcriptomic features, whereas quantitative pathology prediction was more robust and supported by transcriptomic and cell-type-associated features with potential biological relevance to AD progression. The challenge also introduced the first AI Agent Track in a DREAM Challenge, providing an early benchmark for autonomous and human-guided agentic model development in single-cell neuroscience. This work demonstrates that single-nucleus transcriptomes encode substantial information about Alzheimer's disease pathology, establishes a reproducible benchmark for molecular neuropathology prediction, and highlights critical principles for designing privacy-preserving, leakage-aware community challenges using deeply phenotyped human brain data.

neuroscience↗

bioRxiv2025

Benchmarking Large Language Models for Predictive Modeling in Biomedical Research With a Focus on Reproductive Health

Generative AI, particularly large language models (LLMs), is increasingly being used in computational biology to support code generation for data analysis. In this study, we evaluated the ability of LLMs to generate functional R and Python code for predictive modeling tasks, leveraging standardized molecular datasets from several recent DREAM (Dialogue for Reverse Engineering Assessments and Methods) Challenges focused on reproductive health. We assessed LLM performance across four predictive tasks derived from three DREAM challenges: gestational age regression from gene expression, gestational age regression from DNA methylation profiles, and classification of preterm birth and early preterm birth from microbiome data. LLMs were prompted with task descriptions, data locations, and target outcomes. LLM-generated code was then run to fit and apply prediction models and generate graphics, and they were ranked based on their success in completing the tasks and achieving strong test set performance. Among the eight LLMs tested, o3-mini-high, 4o, DeepseekR1 and Gemini 2.0 completed at least one task without error. Overall, R code generation was more successful (14/16 tasks) than Python (7/16), attributed to the utility of Bioconductor packages for querying Gene Expression Omnibus data. OpenAIs o3-mini-high outperformed others, completing 7/8 tasks. Test set performance of the top LLM matched or exceeded top-performing teams from the original DREAM challenges. These findings underscore the potential of LLMs to enhance exploratory analysis and democratize access to predictive modeling in omics by automating key components of analysis pipelines, and highlight the potential to increase research output when conducting analyses of standardized datasets from public repositories.

bioinformatics↗

bioRxiv2025

Multiple Types of Context-Specific Brain Causal Regulatory Networks and their Applications to Autism Spectrum Disorder

One of the primary goals of systems biology is to understand how different cell types are regulated at various developmental stages and how dysregulation in these cells contributes to disease. With advances in high-throughput omics profiling technologies, vast amounts of both bulk tissue and single-cell data have been generated to study cellular regulation. To investigate the types of gene-gene regulatory interactions that can be uncovered from bulk versus single-cell profiling data, we constructed multiple context-specific brain gene regulatory networks using brain bulk and single-cell RNA-seq datasets and applied two distinct computational methods: an integrative Bayesian Network (BN) and a deep learning model, Geneformer. A total of 39 networks were generated, comprising 13 brain region-specific Bayesian networks from bulk RNA-seq data (bulk_BNs), 13 cell-type and developmental stage-specific Bayesian networks, and 13 corresponding Geneformer-based networks from scRNA-seq data (sc_BNs and sc_GFs, respectively). We compared these networks in terms of their topological features, biological pathway regulation, and known regulatory relationships among disease-related genes. Our analysis revealed distinct network properties driven by data source type (bulk vs. single-cell RNA-seq) and network inference method (integrative Bayesian Network vs. deep learning Geneformer). Neurodegenerative disease-related pathways were co-regulated in bulk_BNs, while sc_BNs and sc_GFs captured pathways associated with neuroactive ligand-receptor interactions. To demonstrate the utility of these networks, we leveraged them to investigate potential disease mechanisms of Autism Spectrum Disorder (ASD). We found that ASD risk genes were enriched among hub genes common across networks but were more frequently hubs in context-specific networks, underscoring the significance of inhibitory interneurons during early developmental stages in ASD. Additionally, we showed that these networks could effectively recapitulate context-specific ASD gene signatures. In summary, this study highlights the complementary value of integrating multiple data types through different computational approaches to advance our understanding of disease-related molecular mechanisms, such as those involved in ASD. Our findings provide a framework for using context-specific gene regulatory networks to identify disease-causal genes and key regulators of signaling pathways, potentially guiding future therapeutic strategies.

bioinformatics↗

bioRxiv2021

Extracellular Vesicles Carry Distinct Proteo-Transcriptomic Signatures That are Different from Their Cancer Cell of Origin

Circulating extracellular vesicles (EVs) contain molecular footprints from their cell of origin and may provide potential non-invasive access for detection, characterization, and monitoring of numerous diseases. Despite their growing promise, the integrated proteo-transcriptomic landscape of EVs and their donor cells remain poorly understood. To assess their cargo, we conducted small RNA sequencing and mass spectrometry (LC-MS/MS) of EVs isolated from in vitro cancer cell culture and prostate cancer patients serum. Here, we report that EVs enrich for distinct molecular cargo, and their proteo-transcriptome is predominantly different from their cancer cell of origin, implicating a coordinated disposal and delivery mechanism. We have discovered that EVs package their cargo in a non-random fusion, as their most enriched RNAs and proteins are not the most abundant cargo from their donor cells. We show that EVs enrich for 4 times more cytoskeletal and 2 times extracellular proteins than their donor cells. While the donor cells carry 10 times more mitochondrial and 3 times nuclear proteins than their EVs. EVs predominantly (40-60%) enrich for small RNA (~15-200 nucleotides) molecules that implicate cell differentiation, development, and signaling signatures. Finally, our integrated proteo-transcriptomic analyses reveal that EVs are enriched of RNAs (RNY3, vtRNA, and MIRLET-7) and their complementary proteins (YBX1, IGF2BP2, SRSF1/2), implicating an interrelated mechanism that may protect and regulate transcripts until a biological function is achieved. Based on these results, we envision that the next-generation clinical assays will take an integrative multi-omic (proteomic and transcriptomic) approach for liquid biopsy in numerous diseases.

bioinformatics↗

bioRxiv2019

The transcriptomic response of cells to a drug combination is more than the sum of the responses to the monotherapies

Our ability to predict the effects of drug combinations is limited, in part by limited understanding of how the transcriptional response of two monotherapies results in that of their combination. We performed the first analysis of matched time course RNAseq profiling of cells treated with both single drugs and their combinations. The transcriptional signature of the synergistic combination we studied had unique gene expression not seen in either constituent monotherapy. This can be explained mechanistically by the sequential activation of transcription factors in time in the gene regulatory network. The nature of this transcriptional cascade suggests that drug synergy may ensue when the transcriptional responses elicited by two unrelated individual drugs are correlated. We used these results as the basis of a simple prediction algorithm attaining an AUROC of 0.84 in the prediction of synergistic drug combinations in an independent dataset.

cell biology↗

bioRxiv2019

Predicting cellular position in the Drosophila embryo from Single-Cell Transcriptomics data

Single-cell RNA-seq technologies are rapidly evolving but while very informative, in standard scRNAseq experiments the spatial organization of the cells in the tissue of origin is lost. Conversely, spatial RNA-seq technologies designed to keep the localization of the cells have limited throughput and gene coverage. Mapping scRNAseq to genes with spatial information increases coverage while providing spatial location. However, methods to perform such mapping have not yet been benchmarked. To bridge the gap, we organized the DREAM Single-Cell Transcriptomics challenge focused on the spatial reconstruction of cells from the Drosophila embryo from scRNAseq data, leveraging as gold standard genes with in situ hybridization data from the Berkeley Drosophila Transcription Network Project reference atlas. The 34 participating teams used diverse algorithms for gene selection and location prediction, while being able to correctly localize rare subpopulations of cells. Selection of predictor genes was essential for this task and such genes showed a relatively high expression entropy, high spatial clustering and the presence of prominent developmental genes such as gap and pair-ruled genes and tissue defining markers.

systems biology↗

bioRxiv2018

Stratification of amyotrophic lateral sclerosis patients: a crowdsourcing approach

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease with substantial heterogeneity in clinical presentation with an urgent need for better stratification tools for clinical development and care. In this study we used a crowdsourcing approach to address the problem of ALS patient stratification. The DREAM Prize4Life ALS Stratification Challenge was a crowdsourcing initiative using data from >10,000 patients from completed ALS clinical trials and 1479 patients from community-based patient registers. Challenge participants used machine learning and clustering techniques to predict ALS progression and survival. By developing new approaches, the best performing teams were able to predict disease outcomes better than currently available methods. At the same time, the integration of clustering components across methods led to the emergence of distinct consensus clusters, separating patients into four consistent groups, each with its unique predictors for classification. This analysis reveals for the first time the potential of a crowdsourcing approach to uncover covert patient sub-populations, and to accelerate disease understanding and therapeutic development.

bioinformatics↗

bioRxiv2018

Open Community Challenge Reveals Molecular Network Modules with Key Roles in Diseases

Identification of modules in molecular networks is at the core of many current analysis methods in biomedical research. However, how well different approaches identify disease-relevant modules in different types of gene and protein networks remains poorly understood. We launched the "Disease Module Identification DREAM Challenge", an open competition to comprehensively assess module identification methods across diverse protein-protein interaction, signaling, gene co-expression, homology, and cancer-gene networks. Predicted network modules were tested for association with complex traits and diseases using a unique collection of 180 genome-wide association studies (GWAS). Our critical assessment of 75 contributed module identification methods reveals novel top-performing algorithms, which recover complementary trait-associated modules. We find that most of these modules correspond to core disease-relevant pathways, which often comprise therapeutic targets and correctly prioritize candidate disease genes. This community challenge establishes benchmarks, tools and guidelines for molecular network analysis to study human disease biology (https://synapse.org/modulechallenge).

bioinformatics↗

bioRxiv2017

Combining accurate tumour genome simulation with crowd sourcing to benchmark somatic structural variant detection

BackgroundThe phenotypes of cancer cells are driven in part by somatic structural variants. Structural variants can initiate tumors, enhance their aggressiveness and provide unique therapeutic opportunities. Whole-genome sequencing of tumors can allow exhaustive identification of the specific structural variants present in an individual cancer, facilitating both clinical diagnostics and the discovery of novel mutagenic mechanisms. A plethora of somatic structural variant detection algorithms have been created to enable these discoveries, however there are no systematic benchmarks of them. Rigorous performance evaluation of somatic structural variant detection methods has been challenged by the lack of gold-standards, extensive resource requirements and difficulties arising from the need to share personal genomic information.\n\nResultsTo facilitate structural variant detection algorithm evaluations, we create a robust simulation framework for somatic structural variants by extending the BAMSurgeon algorithm. We then organize and enable a crowd-sourced benchmarking within the ICGC-TCGA DREAM Somatic Mutation Calling Challenge (SMC-DNA). We report here the results of structural variant benchmarking on three different tumors, comprising 204 submissions from 15 teams. In addition to ranking methods, we identify characteristic error-profiles of individual algorithms and general trends across them. Surprisingly, we find that ensembles of analysis pipelines do not always outperform the best individual method, indicating a need for new ways to aggregate somatic structural variant detection approaches.\n\nConclusionsThe synthetic tumors and somatic structural variant detection leaderboards remain available as a community benchmarking resource, and BAMSurgeon is available at https://github.com/adamewing/bamsurgeon.

bioinformatics↗

bioRxiv2017

Origins of fractional control in regulated cell death

Individual cells in clonal populations often respond differently to environmental changes; for binary phenotypes, such as cell death, this can be measured as a fractional response. These types of responses have been attributed to cell-intrinsic stochastic processes and variable abundances of biochemical constituents, such as proteins, but the influence of organelles has yet to be determined. We use the response to TNF-related apoptosis inducing ligand (TRAIL) and a new statistical framework for determining parameter influence on cell-to-cell variability through the inference of variance explained, DEPICTIVE, to demonstrate that variable mitochondria abundance correlates with cell survival and determines the fractional cell death response. By quantitative data analysis and modeling we attribute this effect to variable effective concentrations at the mitochondria surface of the pro-apoptotic protein Bax. Further, we demonstrate that inhibitors of antiapoptotic Bcl-2 family proteins, used in cancer treatment, may increase the diversity of cellular responses, enhancing resistance to treatment.

systems biology↗

bioRxiv2017

Community assessment of cancer drug combination screens identifies strategies for synergy prediction

The effectiveness of most cancer targeted therapies is short lived since tumors evolve and develop resistance. Combinations of drugs offer the potential to overcome resistance, however the number of possible combinations is vast necessitating data-driven approaches to find optimal treatments tailored to a patients tumor. AstraZeneca carried out 11,576 experiments on 910 drug combinations across 85 cancer cell lines, recapitulating in vivo response profiles. These data, the largest openly available screen, were hosted by DREAM alongside deep molecular characterization from the Sanger Institute for a Challenge to computationally predict synergistic drug pairs and associated biomarkers. 160 teams participated to provide the most comprehensive methodological development and subsequent benchmarking to date. Winning methods incorporated prior knowledge of putative drug target interactions. For >60% of drug combinations synergy was reproducibly predicted with an accuracy matching biological replicate experiments, however 20% of drug combinations were poorly predicted by all methods. Genomic rationale for synergy predictions were identified, including antagonism unique to combined PIK3CB/D inhibition with the ADAM17 inhibitor where synergy is seen with other PI3K pathway inhibitors. All data, methods and code are freely available as a resource to the community.

bioinformatics↗

bioRxiv2017

An extensible ontology for inference of emergent whole cell function from relationships between subcellular processes

Whole cell responses arise from coordinated interactions between diverse human gene products functioning within various pathways underlying sub-cellular processes (SCP). Lower level SCPs interact to form higher level SCPs, often in a context specific manner to give rise to whole cell function. We sought to determine if capturing such relationships enables us to describe the emergence of whole cell functions from interacting SCPs. We developed the \"Molecular Biology of the Cell\" ontology based on standard cell biology and biochemistry textbooks and review articles. Currently, our ontology contains 5,392 genes, 753 SCPs and 19,182 expertly curated gene-SCP associations. Our algorithm to populate the SCPs with genes enables extension of the ontology on demand and the adaption of the ontology to the continuously growing cell biological knowledge. Since whole cell responses most often arise from the coordinated activity of multiple SCPs, we developed a dynamic enrichment algorithm that flexibly predicts SCP-SCP relationships beyond the current taxonomy. This algorithm enables us to identify interactions between SCPs as a basis for higher order function in a context dependent manner, allowing us to provide a detailed description of how SCPs together can give rise to whole cell functions. We conclude that this ontology can, from omics data sets, enable the development of detailed multidimensional SCP networks for predictive modeling of emergent whole cell functions.

systems biology↗

bioRxiv2016

A community-based collaboration to build prediction models for short-term discontinuation of docetaxel in metastatic castration-resistant prostate cancer patients

BackgroundDocetaxel has a demonstrated survival benefit for metastatic castration-resistant prostate cancer (mCRPC). However, 10-20% of patients discontinue docetaxel prematurely because of toxicity-induced adverse events, and managing risk factors for toxicity remains an ongoing challenge for health care providers and patients. Prospective identification of high-risk patients for early discontinuation has the potential to assist clinical decision-making and can improve the design of more efficient clinical trials. In partnership with Project Data Sphere (PDS), a non-profit initiative facilitating clinical trial data-sharing, we designed an open-data, crowdsourced DREAM (Dialogue for Reverse Engineering Assessments and Methods) Challenge for developing models to predict early discontinuation of docetaxel\n\nMethodsData from the comparator arms of four phase III clinical trials in first-line mCRPC were obtained from PDS, including 476 patients treated with docetaxel and prednisone from the ASCENT2 trial, 598 patients treated with docetaxel, prednisone/prednisolone, and placebo in the VENICE trial, 526 patients treated with docetaxel, prednisone, and placebo in the MAINSAIL trial, and 528 patients treated with docetaxel and placebo in the ENTHUSE 33 trial. Early discontinuation was defined as treatment stoppage within three months due to adverse treatment effects. Over 150 clinical features including laboratory values, medical history, lesion measures, prior treatment, and demographic variables were curated and made freely available for model building for all four trials. The ASCENT2, VENICE, and MAINSAIL trial data sets formed the training set that also included patient discontinuation status. The ENTHUSE 33 trial, with patient discontinuation status hidden, was used as an independent validation set to evaluate model performance. Prediction performance was assessed using area under the precision-recall curve (AUPRC) and the Bayes factor was used to compare the performance between prediction models.\n\nResultsThe frequency of early discontinuation was similar between training (ASCENT2, VENICE, and MAINSAIL) and validation (ENTHUSE 33) sets, 12.3% versus 10.4% of docetaxel-treated patients, respectively. In total, 34 independent teams submitted predictions from 61 different models. AUPRC ranged from 0.088 to 0.178 across submissions with a random model performance of 0.104. Seven models with comparable AUPRC scores (Bayes factor [≤]; 3) were observed to outperform all other models. A post-challenge analysis of risk predictions generated by these seven models revealed three distinct patient subgroups: patients consistently predicted to be at high-risk or low-risk for early discontinuation and those with discordant risk predictions. Early discontinuation events were two-times higher in the high-versus low-risk subgroup and baseline clinical features such as presence/absence of metastatic liver lesions, and prior treatment with analgesics and ACE inhibitors exhibited statistically significant differences between the high- and low-risk subgroups (adjusted P < 0.05). An ensemble-based model constructed from a post-Challenge community collaboration resulted in the best overall prediction performance (AUPRC = 0.230) and represented a marked improvement over any individual Challenge submission. A\n\nFindingsOur results demonstrate that routinely collected clinical features can be used to prospectively inform clinicians of mCRPC patients risk to discontinue docetaxel treatment early due to adverse events and to the best of our knowledge is the first to establish performance benchmarks in this area. This work also underscores the \"wisdom of crowds\" approach by demonstrating that improved prediction of patient outcomes is obtainable by combining methods across an extended community. These findings were made possible because data from separate trials were made publicly available and centrally compiled through PDS.

bioinformatics↗

bioRxiv2016

Reverse-engineering human olfactory perception from chemical features of odor molecules

Despite 25 years of progress in understanding the molecular mechanisms of olfaction, it is still not possible to predict whether a given molecule will have a perceived odor, or what olfactory percept it will produce. To address this stimulus-percept problem for olfaction, we organized the crowd-sourced DREAM Olfaction Prediction Challenge. Working from a large olfactory psychophysical dataset, teams developed machine learning algorithms to predict sensory attributes of molecules based on their chemoinformatic features. The resulting models predicted odor intensity and pleasantness with high accuracy, and also successfully predicted eight semantic descriptors (\"garlic\", \"fish\", \"sweet\", \"fruit\", \"burnt\", \"spices\", \"flower\", \"sour\"). Regularized linear models performed nearly as well as random-forest-based approaches, with a predictive accuracy that closely approaches a key theoretical limit. The models presented here make it possible to predict the perceptual qualities of virtually any molecule with an impressive degree of accuracy to reverse-engineer the smell of a molecule.\n\nOne Sentence SummaryResults of a crowdsourcing competition show that it is possible to accurately predict and reverse-engineer the smell of a molecule.

neuroscience↗