Polly Xtract is an advanced, proprietary AI-driven capability developed by Elucidata. It's designed to intelligently extract and structure complex data from a wide array of unstructured and semi-structured sources, including PDFs, images, free text, diverse tables, and combinations thereof.
While not a standalone product for purchase, Polly Xtract serves as a core technological engine that empowers our expert team to deliver unparalleled data curation services. It significantly enhances our ability to process vast quantities of heterogeneous research and operational data – from publications and EMR tables to increasingly complex domains such as chemical structures and regulatory filings. By automating and accelerating the critical first steps of data preparation, Polly Xtract enables us to undertake larger, more ambitious data projects for our clients with greater speed, accuracy, and efficiency, all while maintaining the highest standards of data quality. It represents Elucidata's commitment to leveraging cutting-edge technology to transform complex data into actionable insights for the biopharma and related industries

Polly Xtract follows a modular, multi-agent framework:

• Document Parsing Agents – Ingest structured/unstructured sources including GEO, publications, and supplementary materials.
• Extraction Agents – Operate on text, tables, and figures to extract field-specific values.
• Ontology Mapping Agents – Normalize outputs using vocabularies like MeSH, Cell Ontology, and Disease Ontology.
• Reasoning/Validation Agents – Handle conflicts, perform plausibility checks, and reconcile outputs.
• QC/Review Loop – Human-in-the-loop review for flagged fields.
• Schema-Enforced Output – Structured data is written to Polly Atlas and exposed to downstream tools (e.g., Polly KG).

While generic LLMs (e.g., GPT-4) can parse documents, Polly Xtract is purpose-built for biomedical metadata curation:

• Task-specific orchestration: Each agent is specialized for a field or function—no reliance on one-shot inference.
• Schema-constrained extraction: Outputs adhere to pre-defined downstream schemas.
• Domain-context encoding: Built on six years of curatorial decision-making and QA.
• Coordinated agent behavior: Structured hand-offs ensure system-level reasoning and validation.

Polly Xtract delivers high-accuracy, schema-aware metadata extraction from unstructured biomedical sources. Across 50+ metadata fields spanning study design, trial arms, and outcomes, it achieves:

• ≥90% accuracy for simple (binary) and complex (textual) fields
• ~87% accuracy for moderate (numeric) fields
• F1 scores above 85% across most field types
• 100% consistency across binary fields over multiple runs
• 90% groundedness, with extracted values traceable to source documents
• 100% field-level coverage, with no missing predictions

In multiple cases, Polly Xtract outperformed manual curation - correctly extracting values absent in the ground truth but verifiable from the source. This contributed to a 4× increase in throughput, matching the monthly output of a 3-person expert team. Xtract also preserves explainability, with structured reasoning logs and field-level evidence.

• Omics metadata harmonization – Extract and standardize sample-level metadata across GEO, ArrayExpress, and internal datasets.
• Clinical trial parsing – Auto-extract schema elements from protocols (e.g., arms, endpoints, eligibility).
• EHR/EMR extraction – Structure unstructured patient data for real-world evidence and cohort analytics.
• Toxicology digitization – Transform legacy reports into structured datasets for analysis or submission.
• Scientific literature curation – Pull out study design, compound info, and results from publications and supplements.
• Assay result digitization – Normalize outputs from vendor PDFs or spreadsheets into LIMS-compatible formats.

Polly Xtract is best suited for organizations that manage high volumes of biomedical documents and require domain-specific accuracy. Target users include:

• Curation teams handling omics or clinical datasets
• Informatics/R&D teams building knowledge graphs, data lakes, or FAIR repositories
• Data scientists preparing training datasets
• Clinical ops or biomarker groups working with protocols and lab data
• Pharma and diagnostics teams receiving unstructured data from partners

If required, our team can add the following customizations for cell-type annotation

• Textual: GEO pages, Full-text publications (PMC, publisher PDFs), Supplementary files (PDF, Excel, Word), Trial protocols and case report forms, EHR/EMR exports, regulatory documents.
• Tabular/Embedded: HTML tables (GEO, PMC), Image-based or LaTeX tables in PDFs, CSV/TSV files in supplementary sections
• Unstructured/Free-text: Clinical narratives, figure captions, journal discussions
• Multimodal: Cross-referencing across GEO, supplements, external links, Context-sensitive extraction from multiple documents per study

Xtract is schema-flexible, supporting both pre-defined and user-defined metadata fieldsets.

Polly Xtract supports extraction across 23+ fields (as per preprint), including:

• Disease
• Cell type
• Sample type
• Tissue
• Perturbation or compound
• Platform/technology
• Organism
• Donor ID
• Study accession

It handles both raw entity extraction and ontology-based normalization (e.g., “AML” → DOID:9119).

The pipeline supports cross-document and multimodal parsing. Agents are designed to:

• Navigate between GEO pages, full texts, and supplementary files
• Reconcile field values across sources
• Use table-aware and long-context models (e.g., SciTSR, RAG pipelines) to locate dispersed information

Polly Xtract is not yet a standalone commercial product. However, it is actively used within Elucidata's curation operations and is available for early-access partnerships and also enterprise deployments. Teams working on high-volume biomedical data extraction are invited to reach out for collaboration discussions.