henhaobio
Sign in

bioRxivBioinformaticsPreprint

EVd3x: A Bioinformatics Platform for Evidence-aware Interpretation of EV Cargo

Karima Ait Ouares, Jonathan S. WeerakkodyiD

PDFDOICodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (9 Sept 2026) · CC BY 4.0

EVd3x integrates over 22 million records from 28 public resources so one fixed molecule set can be traced from reported EV cargo through disease, pathway, cell and interaction evidence.

EV profiling yields long RNA and protein lists, but interpreting them needs evidence scattered across cargo repositories, publications, pathway, disease, cell-atlas and interaction resources that are rarely reviewed together. Unlike independent database searches, EVd3x freezes each query into an exportable evidence packet that carries the same molecules through every analysis layer while holding reported EV detection apart from contextual annotation.

Key results

  1. ShownCombines 28 public resources, over 22 million records and interactions in total, into 17 canonical Parquet tables keyed to Ensembl, UniProt and miRBase identifiers, with upstream source fields retained for audit.

    Release row counts: 234,090 EV-cargo rows tied to 3,350 publication records; 2,639,746 scored miRNA-target pairs; 404,758 pathway memberships; 501,157 gene-disease links; 1,587,444 cell-specificity records; 25,779 ligand-receptor pairs; 13,715,404 STRING interactions. Canonicalization version 2026.02.13. · Abstract; 1. Introduction; 5.2 Canonical table construction and source…

  2. ShownA free-text Alzheimer query was routed to familial Alzheimer disease type 3 and produced a fixed PSEN1-centered packet that every downstream module reused; PSEN1 was reported in EV material across several publications and sample types.

    Query (early-onset Alzheimer's disease with behavioral disturbance) mapped to DisGeNET C1843013; packet held 5 visible starting molecules, 104 inferred molecules, 109 nodes and 197 relationships, with 85 nodes having reported EV support. PSEN1 (UniProt P49768) had thirty-four cargo rows from ExoCarta, Vesiclepedia and SVAtlas, 17 linked to EV-TRACK. The packet was loaded from fixed Table S2. · 2.1 A fixed Alzheimer case study demonstrates…; 2.2 Source-linked records define the reported…; 5.3 Query workflow and reviewable evidence…

  3. SuggestedDisease and pathway layers recovered expected Alzheimer biology: familial Alzheimer disease type 3 ranked first, and enriched terms grouped into presenilin/Notch with APP/BACE1 processing, kinase, and neurotrophin/synaptic themes.

    Selected disease record: 483 direct association rows, 349 publications, five source categories, PSEN1 as supporting molecule. Hypergeometric test with Benjamini-Hochberg correction (p ≤ 0.05, q ≤ 0.10) gave 153 displayed terms across 16 categories from GO, KEGG, Reactome and WikiPathways; kinase theme included AAK1, CDK6, CDK8, CSNK1G1, HIPK2, LRRK2 and MAP3K2. · 2.3 Disease associations recover the expected…; 2.4 Pathway convergence organizes annotation…; 5.5 Disease and pathway annotation

  4. ShownComponents exposed weak support: the highly ranked oligodendrocyte precursor context reflected seed coverage and system-level weighting, not marker evidence, and PSEN1 ligand-receptor records were flagged as uncertain since PSEN1 is an integral membrane γ-secretase component, not a conventional soluble ligand.

    Oligodendrocyte precursor card: 100% seed coverage, 44% seed specificity, 38% expanded-gene support, no marker support, communication readiness 51.3. In an oligodendrocyte-to-astrocyte view, PSEN1-NOTCH2, PSEN1-NOTCH1 and PSEN1-CTNNB1 remained after hiding context-only rows; the ligand-receptor workspace held 323 candidate pairs. · 2.5 Cell-context analysis identifies candidate…; 2.6 Integrated relationship evidence identifies…; 5.6 Cell-context and communication…

  5. SuggestedThe Alzheimer gene packet carried more pathway and disease annotation than size-matched random gene sets, which the authors read as annotation coherence within the database backgrounds rather than EV specificity.

    Packet exceeded distributions from 250 random same-size sets (drawn from genes present in both pathway and disease resources, case-study genes removed) for pathway-membership burden, total disease-association burden and neurodegenerative-disease burden; no empirical P values are given in the main text. · 2.6 Integrated relationship evidence identifies…; 5.9 Sensitivity analysis and source-dependence…

Methods, data and reproducibility

A Python/Flask platform over Parquet tables turns molecule, disease, pathway, cell or free-text queries into a fixed evidence packet analysed by deterministic retrieval, hypergeometric enrichment and fixed-weight heuristics. A QLoRA-tuned Qwen2.5-1.5B-Instruct adapter, called after deterministic intent routing and table retrieval, only summarizes records; it selects no molecules, scores or edges.

  • Reported EV-cargo records from ExoCarta, Vesiclepedia and SVAtlas, with EV-TRACK metadata where available · n = 234,090 records linked to 3,350 publications
  • Scored miRNA-target relationships from miRTarBase, DIANA-TarBase and TargetScan · n = 2,639,746
  • Pathway memberships from Reactome, KEGG, Gene Ontology and WikiPathways · n = 404,758
  • Gene-disease and miRNA-disease associations from DisGeNET · n = 501,157 gene-disease; 145 miRNA-disease
  • Cell-context records from Human Protein Atlas, CellMarker and PanglaoDB · n = 1,587,444
  • Ligand-receptor pairs from CellPhoneDB, OmniPath Intercell, CellTalkDB, Cellinker and Reactome · n = 25,779
  • STRING protein-interaction records · n = 13,715,404
  • Fixed PSEN1-centered Alzheimer case-study evidence packet (Table S2) · human (hsa- miRNA identifiers; human molecular alias catalogue) · n = 109 molecular nodes (5 starting, 104 inferred), 197 relationships · DisGeNET C1843013
Methods & tools
Flask, pandas, NumPy, PyArrow, SciPy, Apache Parquet, Qwen2.5-1.5B-Instruct (QLoRA fine-tuning), ExoCarta, Vesiclepedia, SVAtlas, EV-TRACK, STRING
Software & environment
Python with Flask, pandas, NumPy, PyArrow, SciPy; canonicalization version 2026.02.13; adapter fine-tuned from Qwen2.5-1.5B-Instruct via QLoRA: three epochs, learning rate 2 × 10^-4, max length 2,048, LoRA rank 16, alpha 32, dropout 0.05, batch size 4, gradient accumulation four steps.
Reproducibility gaps
Companion repository omits the full production source code and deployment configuration; raw upstream databases and merged production Parquet tables are not redistributed for licensing reasons; API data endpoints need an API key.
Limitations (6)
  • Evaluation rests on one well-characterized Alzheimer query whose PSEN1 packet was loaded from a fixed table; it does not test unconstrained natural-language disease-to-molecule discovery. · stated by the authors
  • Reported EV-cargo records differ in sample type, preparation, characterization, organism and detection technology, and most contextual records come from outside EV, biofluid or disease settings. · stated by the authors
  • Disease, pathway, cell-context, communication and miRNA-bridge scores are fixed heuristics that order review rather than estimate probability, and row counts can reflect publication density or source redundancy. · stated by the authors
  • The small local language model may produce incomplete summaries when context is sparse or overstate broad associations; the default interface accepts at most 25 molecules per graph request. · stated by the authors
  • The main text reports no routing or summarization accuracy for the 424 held-out examples or the 621-prompt stress suite (deferred to Table S8). · noted in this summary
  • The random-set null compares a deliberately well-studied positive-control packet with random genes, so exceeding it may largely reflect literature density; no empirical P values are reported in the main text. · noted in this summary

Read this if you need to triage an EV RNA or protein cargo list against public disease, pathway, cell-type, miRNA-target and interaction evidence while keeping reported EV detection traceable to source studies.

Citations
1

Abstract

Extracellular vesicle (EV) profiling can identify hundreds to thousands of RNAs and proteins, but converting those lists into a coherent biological interpretation remains slow because the relevant evidence is scattered across cargo repositories, publications, pathway databases, cell atlases, and interaction resources. We introduce EVd3x.com, a source-linked bioinformatics platform that brings these evidence layers into one continuous workflow for EV research. EVd3x harmonizes more than 22 million records and interactions from 28 public resources into 17 analysis tables, including 234,090 reported EV-cargo records linked to 3,350 publications, 2.6 million miRNA-target relationships, 404,758 pathway memberships, more than 501,000 disease associations, 1.6 million cell-context records, 25,779 ligand-receptor pairs, and 13.7 million protein-interaction records. Each query generates a fixed, exportable evidence packet that follows the same molecules from reported EV detection through disease, pathway, cellular, regulatory, and interaction analyses. Starting with the complex disease query [early-onset Alzheimers disease with behavioral disturbance], EVd3x arrived at a focused, PSEN1-centered evidence packet and organized thousands of linked records into a source-traceable interpretation. The workflow recovered expected Alzheimer biology while keeping reported EV detection separate from broader contextual evidence, thereby exposing evidence gaps and context-dependent relationships. EVd3x provides EV researchers with a practical route from a cargo list, or from a disease, pathway, or cell-context question to a prioritized set of molecules, relationships, evidence gaps, and unresolved questions that researchers can use to formulate hypotheses and design subsequent experimental studies.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLoria '23Torrejón '26Poupardin '24Rai '24Kuipers '25Muraoka '19
60% semantic matchCell Biology
F. Loria, S. Picciotto, G. Adamo+7 more·bioRxiv·Nov 16, 2023