OncoCITE | Ali Quidwai

Overview

OncoCITE is a multi-agent AI system for automated genomic evidence extraction from scientific literature, currently submitted to Nature Cancer.

Try OncoCITE Live

Interactive genomic evidence extraction for Multiple Myeloma

Launch Demo →

The Problem

Through systematic exploratory data analysis on the CIViC database (11,312 evidence items, 3,083 publications), I quantified 12 structural bottlenecks:

Curation latency: Median 31 days, P90 >21 months
Pareto inefficiency: Top 100 sources = only 29% coverage
Emerging target gaps: GPRC5D has 0 items despite FDA approval

Technical Architecture

6-Agent Collaborative System

Built using Claude Agent SDK with sophisticated orchestration:

22 custom MCP tools for genomic data extraction
State serialization enabling pause-resume for long-running extractions
Vision-based PDF extraction (300 DPI) replacing traditional OCR
Hierarchical multi-tier architecture with supervisor-worker pattern

Novel Validation Framework

Developed a three-way validation methodology treating source publications as ground truth—identified 24.2% curation errors in expert-curated databases.

Results

Metric	Value
Ground Truth Recovery	84%
Novel Discovery Precision	97.8%
Critical Errors	0% (n=108)
Ontology Resolution	83.12% across 20 Tier-2 fields

Validated on 15-paper corpus and deployed normalizer to enrich all 11,312 CIViC items.

Key Innovations

State Serialization: Enables deterministic replay and debugging of multi-agent workflows
Cross-field Validation: Automated consistency checking across extracted fields
Hallucination Prevention: Input/output guardrails with reasoning chain validation

This project resulted in the following publications:

(Quidwai et al., 2025) - Full paper submitted to Nature Cancer
(Quidwai et al., 2025) - Conference abstract presented at ASH 2025

Technical Stack

Agent Framework:    Claude Agent SDK
Tools:              22 custom MCP tools
PDF Processing:     Vision-based extraction (300 DPI)
Orchestration:      State graphs, pause-resume workflows
Validation:         Three-way ground truth framework
Database:           CIViC (11,312 items)
Ontologies:         DOID, EFO, NCIt, HPO via EMBL-EBI OLS

References

2025

Nature Cancer
OncoCITE: AI-Driven Genomic Evidence Curation for Hematologic Malignancies

Mujahid Ali Quidwai, Santiago Thibaud, Sundar Jagannath, and 2 more authors

Nature Cancer, 2025

Submitted to Nature Cancer

Abs Bib

We present OncoCITE, a multi-agent AI system for automated genomic evidence extraction from scientific literature. Through systematic exploratory data analysis on the CIViC database (11,312 evidence items, 3,083 publications), we quantified 12 structural bottlenecks including curation latency (median 31 days, P90 >21 months) and emerging target gaps. Our 6-agent architecture using Claude Agent SDK with 22 MCP tools achieves 84% ground truth recovery, 97.8% novel discovery precision, and 0% critical errors on a 15-paper validation corpus. A novel three-way validation framework identified 24.2% curation errors in expert-curated databases.
@article{quidwai2025oncocite, title = {OncoCITE: AI-Driven Genomic Evidence Curation for Hematologic Malignancies}, author = {Quidwai, Mujahid Ali and Thibaud, Santiago and Jagannath, Sundar and Parekh, Samir and Lagana, Alessandro}, journal = {Nature Cancer}, year = {2025}, note = {Submitted to Nature Cancer}, }
ASH
Oncodif: An Auditable AI Framework for Automated Genomic Curation and Natural-Language Clinical Querying in Hematologic Malignancies

Mujahid Ali Quidwai, Santiago Thibaud, Sundar Jagannath, and 2 more authors

Blood, 2025

ASH 2025 Conference Abstract

Abs Bib HTML

We present OncoDIF, an auditable AI framework for automated genomic curation achieving 97.8% precision in novel discovery with 0% critical errors. The system enables natural-language clinical querying for hematologic malignancies, providing clinicians with evidence-based genomic insights.
@article{quidwai2025oncodif, title = {Oncodif: An Auditable AI Framework for Automated Genomic Curation and Natural-Language Clinical Querying in Hematologic Malignancies}, author = {Quidwai, Mujahid Ali and Thibaud, Santiago and Jagannath, Sundar and Parekh, Samir and Lagana, Alessandro}, journal = {Blood}, volume = {146}, pages = {2646}, year = {2025}, note = {ASH 2025 Conference Abstract}, }