Skip to content

README.md

Latest commit

 

History

History
407 lines (293 loc) · 15.4 KB

README.md

File metadata and controls

407 lines (293 loc) · 15.4 KB
title OncoMind
emoji 🚀
colorFrom blue
colorTo red
sdk docker
app_file app.py
app_port 8501
pinned false
tags
oncology
cancer
genomics
precision-medicine
bioinformatics
llm
biomedical
genetics
variant-analysis
variant-interpretation
structured-llm
evidence-synthesis
research
streamlit
demo
thumbnail https://huggingface.co/spaces/damigupta/onco_mind/resolve/main/images/starter.png
short_description Cancer Research Copilot - Gap Analysis

OncoMind

Cancer Research Copilot for Gap Analysis

IN PROGRESS

Research intelligence for cancer variants. Find the gaps, not just the facts.

For BRAF V600E, databases already agree. For the next 10,000 variants, the key question is "what don't we know yet?"

OncoMind is a research intelligence platform that identifies evidence gaps in cancer variant knowledge—surfacing where research is thin, conflicting, or missing entirely. It's built for translational teams and small biotechs deciding which variants are worth a project, not for treating individual patients.

How the LLM fits in

The LLM is not the annotation engine—it's the synthesis layer on top of structured data:

  1. Deterministic backbone first: Evidence is aggregated from 8+ databases (CIViC, ClinVar, CGI, VICC, FDA labels, COSMIC, cBioPortal, DepMap) with match specificity tracking (variant vs codon vs gene-level)

  2. Gap detection is rule-based: Missing evidence, source conflicts, and tumor-type extrapolation concerns are computed deterministically before the LLM sees anything

  3. LLM synthesizes, doesn't invent: The LLM receives pre-structured evidence blocks with explicit provenance. It generates:

    • Functional/biological/therapeutic summaries grounded in the evidence provided
    • Research hypotheses tied to specific identified gaps
    • Cross-source drug analysis highlighting corroboration and conflicts

  4. Calibrated to evidence quality: When evidence is sparse, the LLM is instructed to stay generic and highlight unknowns—not fill gaps with training data

Status: Proof-of-Concept / Architectural Demo

This demonstrates an approach to systematic evidence gap detection. It is research use only — not for diagnosis, treatment selection, or any clinical decision-making.

⚠️ SNPs and small indels only. Fusions, amplifications, and copy-number variants are not yet supported.

What Makes It Different

Feature Typical tools OncoMind
Primary question "What is this variant?" "What don't we know yet?"
Knowledge gaps Rarely explicit First-class outputs with severity scoring
Source conflicts Buried in details Detected, surfaced, and explained
Match specificity Not tracked Variant vs codon vs gene-level evidence labeled
Source attribution Often missing Every claim linked to PMID, FDA label, or DB entry
Cancer hotspots Binary yes/no + Adjacent hotspot detection (±5 codons)
Research hypotheses None Generated with evidence basis tags
LLM synthesis Generic summaries Grounded in structured evidence backbone
LLM Cross-source drug analysis Manual comparison Corroboration, conflicts, and emerging targets surfaced
Output Static clinical-style notes LLM-ready context blocks with receipts

Screenshots

Title

Gap Analysis

Gap Analysis

📸 Click to see more screenshots

LLM Research Synthesis

LLM Synthesis

What This Demonstrates

Architecture & Integration

  • Multi-source evidence aggregation (CIViC, ClinVar, CGI, VICC, DepMap, cBioPortal) with conflict detection
  • Evidence hierarchy (variant > codon > gene level) with match specificity tracking
  • Resistance mechanism annotation using cross-database validation
  • Gap detection with severity scoring (CRITICAL → INFORMATIONAL)

Technical Decisions

  • Structured data extraction before LLM synthesis (reduces hallucination risk)
  • Evidence provenance tracking across 6+ databases
  • Context-aware research hypothesis generation
  • Deterministic annotation backbone + optional LLM research layer

Known Limitations (By Design)

This is a proof-of-concept, not production-ready software. Known issues include:

  • Validation: Needs systematic validation, at the technical level, as well as will as by a SME expert.
  • SNPs and small indels only: Fusions, amplifications, and copy-number variants are not yet supported.
  • Negation detection: FDA label parsing may miss negative indicators ("not demonstrated", "not approved")
  • Edge cases: Rare variant-disease pairings may have inconsistent evidence grading
  • Display formatting: Some compound identifiers (CAS numbers) may appear in clinical evidence sections
  • LLM variability: Research hypothesis quality varies; some may be speculative

Production deployment would require:

  • Systematic validation pipeline with domain expert review
  • Robust regulatory text parsing (negation detection, contraindications)
  • Automated testing across representative variant sets
  • Human-in-the-loop review for high-stakes clinical use cases

Use Cases

Good for:

  • Exploring architectural approaches to clinical evidence synthesis
  • Understanding systematic challenges in multi-database genomics integration
  • Generating research directions for understudied variants
  • Portfolio demonstration of domain expertise + technical execution
  • Prioritizing targets, models, or combination strategies for small biotechs
  • Planning functional studies or resistance screens for academic labs
  • Triaging large variant lists from NGS or CRISPR screens

Not suitable for:

  • Clinical decision-making (use validated tools like OncoKB, CIViC)
  • Regulatory submissions
  • Production therapeutic recommendations without expert review

Why This Approach?

Traditional variant knowledgebases focus on summarizing what's known. OncoMind inverts this: it systematically identifies what's unknown to guide research prioritization. The gap detection architecture could inform:

  • Research funding decisions
  • Clinical trial design
  • Functional validation studies

The platform works in two layers:

  1. Deterministic annotation backbone – structured evidence from knowledge bases, trials, cBioPortal, DepMap, Hotspots, literature
  2. Optional LLM research layer – highlights gaps and drafts hypotheses, constrained by that backbone

Quick Start

Install

git clone https://github.com/dami-gupta-git/onco_mind_v0.git
cd onco_mind_v0

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

pip install -e ".[dev]"

If you hit a ModuleNotFoundError after pulling updates, reinstall with:

pip install -e . --force-reinstall

Streamlit UI (Recommended)

The interactive web interface is the easiest way to explore variants:

cd streamlit
streamlit run app.py

Features:

  • Enter any gene + variant + tumor type
  • Browse evidence by source (CIViC, VICC, CGI, FDA, DepMap, etc.)
  • View gap analysis with severity scoring
  • See match specificity (variant vs codon vs gene level)
  • Cross-source drug analysis identifying corroboration and conflicts
  • Export results to JSON

CLI

# Annotation backbone (~7s)
mind insight BRAF V600E --tumor Melanoma

# Literature search only (~15s)
mind insight EGFR L858R -t NSCLC --lit

# Research mode: backbone + LLM (~20s)
mind insight MAP2K1 P124L -t Melanoma --llm

# Full: backbone + literature + LLM (~25s)
mind insight KRAS G12D -t CRC --full

# Save to JSON
mind insight EGFR L858R -t NSCLC --llm --output result.json

# Debug logging
mind insight BRAF V600E --log-level DEBUG

Modes

Mode Flag Output
Annotation (none) Structured evidence from all data sources
Literature --lit + PubMed / Semantic Scholar hits
LLM --llm + Research narrative and gap analysis
Full --full Annotation + literature + LLM layer

What You Get

Evidence Backbone

OncoMind constructs a structured, variant-centric evidence model:

  • Clinical / KB: CIViC, VICC MetaKB, ClinVar, COSMIC, CGI, FDA labels
  • Functional: AlphaMissense, CADD, PolyPhen2, gnomAD
  • Biological: cBioPortal prevalence and co-mutation structure
  • Preclinical: DepMap CRISPR essentiality and PRISM drug response
  • Trials: ClinicalTrials.gov
  • Literature: PubMed / Semantic Scholar summarized in a LiteratureEvidence model

Match specificity is tracked so you can separate variant-level from gene-level signals:

Match level Meaning Example
variant Exact amino-acid change BRAF V600E specific data
codon Same residue, different change BRAF V600K in "V600 variants"
gene Gene-level-only evidence "BRAF mutation" basket trials

Research Insight (LLM Layer)

When enabled, OncoMind adds a research card on top of the evidence backbone:

  • llm_summary – concise synthesis of function, biology, and therapeutic landscape
  • evidence_quality – comprehensive / moderate / limited / minimal
  • knowledge_gaps / well_characterized – structured view of what's missing vs solid
  • research_implications – short, testable hypotheses
  • key_references – PMIDs, trials, and KB IDs supporting the card

Cross-Source Drug Analysis (LLM Layer)

A separate LLM analysis that synthesizes therapeutic evidence across CGI, CIViC, VICC, and Literature sources:

  • Strongest Evidence – Drugs with corroboration across multiple independent sources, with biological rationale
  • Conflicting Signals – Drugs where sources disagree, with likely explanations (tumor type differences, sequential therapy, acquired mutations)
  • Emerging Targets – Single-source preclinical or early-phase evidence worth investigating
  • Key Gaps – Expected drugs not found, tumor type extrapolation concerns

This runs in parallel with the main LLM synthesis for faster response times.

Configuration

Create a .env file (or set as Hugging Face Secrets):

# Required for LLM research mode (Gemini 2.0 Flash is the default)
GOOGLE_API_KEY=your-google-api-key
# Optional: use Anthropic or OpenAI models instead
ANTHROPIC_API_KEY=your-anthropic-key
# Optional: use OpenAI models instead
OPENAI_API_KEY=your-openai-key
# Optional: better literature context 
SEMANTIC_SCHOLAR_API_KEY=your-s2-key

Supported LLM models:

  • gemini/gemini-2.0-flash (default, fast)
  • gemini/gemini-1.5-pro
  • claude-sonnet-4-20250514
  • claude-3-5-haiku-20241022
  • gpt-4o-mini, gpt-4o, gpt-4-turbo

Supported Variant Types

Currently supports:

  • Missense (e.g., V600E, L858R)
  • Nonsense (e.g., R248*)
  • Small indels (e.g., E746_A750del)
  • Frameshift (e.g., K132fs)

Variants can be provided as simple protein changes (V600E, p.V600E) or in HGVS notation; normalization is handled under the hood.

Planned: fusions, amplifications, and copy-number variants.

Data Sources

Clinical & Therapeutic

Source Data Type Access
CIViC Curated variant–drug associations API / dump
VICC MetaKB Aggregated knowledge bases API
ClinVar Clinical significance Via aggregation layer
COSMIC Somatic mutation catalog Via aggregation layer
CGI Biomarker annotations Local DB
FDA Drug approvals Public APIs / labels
ClinicalTrials.gov Active and historical trials Public API

Functional & Biological

Source Data Type Access
cBioPortal Co-mutation patterns, prevalence API
AlphaMissense Pathogenicity predictions Precomputed scores
gnomAD Population frequencies Via aggregation layer

Literature

Source Data Type Access
Semantic Scholar AI-powered literature search API
PubMed Biomedical literature E-utilities

Preclinical Research

Source Data Type Access
DepMap Gene essentiality (CRISPR), drug sensitivity (PRISM), cell line models API / downloads

Development

pytest tests/unit/ -v
pytest tests/unit/ --cov=src/oncomind --cov-report=html

mypy src/oncomind
ruff check src/oncomind
ruff format src/oncomind

Validation

Variant Tumor Type Source Expected Data OncoMind Match Error Description
BRAF V600E Melanoma FDA Labels Trametinib Mekinist, Trametinib + Dabrafenib Mekinist + Tafinlar, Encorafenib + Binimetinib Braftovi + Mektovi, Vemurafenib Zelboraf, Cobimetinib + Vemurafenib Cotellic + Zelboraf, Dabrafenib Tafinlar, Atezolizumab + Cobimetinib + Vemurafenib Tecentriq + Cotellic + Zelboraf, Atezolizumab and Hyaluronidase-tqjs + Cobimetinib + Vemurafenib Tecentriq Hybreza + Cotellic + Zelboraf ✓
BRAF V600E Melanoma CIViC Level A, 5 evidence items ✓
BRAF V600E Melanoma ClinVar Pathogenic ✓
BRAF V600E Melanoma cBioPortal Melanoma (MSK, Clin Cancer Res 2021) ✓
EGFR L858R NSCLC FDA Labels Gefitinib Iressa, Erlotinib Tarceva, Afatinib Gilotrif, Dacomitinib Vizimpro, Osimertinib Tagrisso, Osimertinib + pemetrexed and platinum-based chemotherapy Tagrisso, Amivantamab Rybrevant + lazertinib Lazcluze, Amivantamab Rybrevant + carboplatin + pemetrexed, Amivantamab and hyaluronidase-lpuj Rybrevant Faspro ✓
EGFR L858R NSCLC CIViC Level A, 14 evidence items ✓
EGFR L858R NSCLC DepMap Gene essentiality - EGFR is not essential ✓
KRAS G12C NSCLC FDA Labels Sotorasib, adagrasib ✓
KRAS G12C NSCLC ClinVar Pathogenic ✓
PIK3CA H1047R Breast FDA Labels Inavolisib + palbociclib + fulvestrant, Alpelisib + fulvestrant, Capivasertib + fulvestrant ✓
PIK3CA H1047R Breast CIViC 3 Level A variant specific items ✓
IDH1 R132H Glioma FDA Labels Vorasidenib (Voranigo) ✓
IDH1 R132H Glioma ClinVar Pathogenic ✓
IDH1 R132H Glioma Hotspots This variant is at known cancer hotspot ✓
ERBB2 S310F Breast CIViC Level B, 3 evidence items ✓

Summary:

  • Total checks: 15
  • Matches: 15 (100%)

License

MIT License – see LICENSE.

Acknowledgments

Built on the work of CIViC, VICC, MyVariant.info, DepMap, Semantic Scholar, cBioPortal, and the broader open-data oncology community.