Patient-Driven Genomic Analysis

Obtain Your Raw Genomic Data

The single most important step. You have a legal right to request your genomic data files under HIPAA in the US and GDPR in Europe. Don't accept only a PDF report — request the underlying VCF, BAM/CRAM, or counts files.

Assemble Your Complete Clinical Records

Collect all pathology reports, operative notes, radiology reports, lab results, and genetics consult notes. Most patients dramatically underestimate how much data they have scattered across institutions.

• Use patient portal apps (MyChart, Epic) to download structured data
• Submit written records requests to each institution you've been seen at
• Organize chronologically — date of diagnosis is your anchor point
• Flag discrepancies between institutions (imaging reports often disagree on measurements)

Create a Structured Patient Profile Document

A structured document (plain text, JSON, or Word) that puts your entire medical story in one place. This serves two purposes: (1) gives AI systems the context they need to reason correctly about your case, and (2) becomes your permanent research record, independent of any EHR.

What to include:

• Demographics and your role as a patient-researcher
• Primary diagnoses with staging, dates, institutions, and current status
• Comorbidities with ICD codes and current management
• Treatment history with outcomes
• Germline genetics findings (variants, zygosity, population frequency, clinical significance)
• Somatic tumor findings (if available)
• Labs and imaging summaries
• Open clinical questions you want answered

The file bill_paseman_patient_profile2.json from this project is a working template you can adapt.

Add an LLM Orientation Prompt at the Very Top

This is the most important design lesson from this project. When LLMs receive large medical documents without a grounding frame, they hallucinate — misidentifying conditions, confusing comorbidities with primary diagnoses, and suggesting treatments already ruled out by the data.

Your orientation prompt should explicitly state:

• Who you are and what your actual diagnoses are (with current status — NED, active, surveillance)
• What the document IS and IS NOT
• What has been definitively ruled out (syndromes, treatments)
• How to navigate the document for different types of questions

Annotate Every Finding with Clinical Significance

Don't just list data — annotate each finding with what it means. For every variant, lab result, or genomic finding, include:

• A clinical_significance field in plain English (HIGH / MEDIUM / LOW)
• Evidence quality label (confirmed, plausible, ruled_out, unknown)
• Cross-references to related findings in other sections
• What it means for treatment decisions

This transforms raw data into a reasoning document that any clinician or AI can act on.

Identify Disease-Relevant Germline Genes

For your specific condition, research which hereditary syndromes are associated with it and which genes to check. Resources:

• ClinVar (clinvar.ncbi.nlm.nih.gov) — searchable database of variants and clinical significance
• OMIM (omim.org) — genetic disease reference encyclopedia
• NCCN Guidelines — gene panels recommended for specific cancer types (free registration)
• Your condition's patient advocacy organization — many maintain curated gene lists

Query Your Germline VCF for Target Genes

If you have a VCF file, you or a bioinformatician can filter it for your target genes. In this project we used command-line tools (zcat + grep) and Python scripts on annotated CSV files from BGI.

For each variant found, assess:

• Population frequency — variants >1% frequency are almost never clinically significant
• SIFT and PolyPhen2 scores — functional impact predictors
• ClinVar status — is it classified as Pathogenic, Likely Pathogenic, VUS, or Benign?
• Variant type — truncating (frameshift, stop-gain, splice) variants are higher concern than missense

Interpret Germline Findings with Care

The single most important judgment call in germline analysis is distinguishing a pathogenic variant from a benign polymorphism. Common polymorphisms (>1–5% population frequency) with benign functional predictions are almost never clinically significant, even in disease-associated genes.

Identify Differentially Expressed Genes (RNA-seq)

If you have RNA-seq from tumor vs. matched normal tissue, calculate fold-changes for each gene. Prioritize:

• Genes strongly UP in tumor — potential drivers or therapeutic targets
• Genes strongly DOWN in tumor — tumor suppressors lost
• Genes completely absent in normal but present in tumor — potential fusions or aberrant activation

Tools: DESeq2, edgeR (bioinformatics software). Or ask an LLM to analyze a processed counts table if you can provide one.

Cross-Reference RNA Findings with DNA

The most powerful signals are convergent — the same gene showing abnormality at multiple levels simultaneously. This is how you separate noise from signal:

Perform Somatic Copy Number Variant (CNV) Analysis

Compare tumor copy number against a germline baseline (ideally matched blood/normal tissue). Always confirm CNV is somatic — absent from germline — before calling it a tumor finding.

• Copy ratio >1.5x = amplification (potential driver or target)
• Copy ratio <0.8x = deletion (potential tumor suppressor loss)
• Copy ratio ~1.0 = normal diploid

Run Mutational Signature Analysis (SBS)

Mutational signatures reveal the mechanism that created the tumor's mutations — aging, APOBEC activity, MMR deficiency, tobacco exposure, UV, etc. This has direct clinical implications:

Tools: SigProfilerExtractor (Python), COSMIC Mutational Signatures database (cancer.sanger.ac.uk/signatures/).

Calculate Microsatellite Instability (MSI)

Count somatic indels per megabase from your somatic indel VCF. This determines whether checkpoint inhibitor monotherapy is appropriate:

Use Multiple AI Systems — The Surowiecki Principle

No single AI model has the full picture. Send your patient profile to multiple LLMs and compare their analyses. Diversity of AI opinion catches errors and reveals where the evidence is genuinely uncertain vs. where there is consensus.

The Wisdom of Crowds (Surowiecki) requires four conditions: diversity of sources, independence of models, decentralization (no single authority), and aggregation of outputs. Apply all four to your AI tumor board.

AI systems to consider: Claude, GPT-4o, Gemini, Perplexity (for real-time literature search).

Conduct Structured AI Tumor Board Sessions

Frame your questions specifically rather than asking "what do I have?" or "what should I do?":

• "Given the somatic findings in the genetics section, what would be the first-line treatment at recurrence?"
• "Are there clinical trials for [condition] that match my molecular profile?"
• "What does the co-occurrence of [condition A] and [condition B] suggest about underlying mechanisms?"
• "What are the limitations of this analysis and what data would change the conclusions?"

Always ask the AI to label each claim with an evidence grade and include citations (PMID or DOI format).

Challenge the AI's Conclusions

Ask: "What alternative interpretations exist?" and "What would need to be true for your conclusion to be wrong?"

Get CLIA-Certified Validation of Key Findings

Research-grade data (hackathons, direct-to-consumer sequencing) cannot directly drive clinical decisions. To get your findings into the medical record and usable by treating physicians, the most practical path is to order a CLIA-certified clinical genomic panel on archived tumor tissue or circulating tumor DNA.

Share Your Patient Profile with Your Care Team

Bring a printed or PDF version of your patient profile to appointments. Frame it as: "I've organized my medical history here — it may help you understand my full picture."

Physicians who receive organized, cited patient research respond far better than to unstructured verbal summaries. The HTML report format (bill_paseman_genomic_report.html) generated in this project is designed to be shareable with care teams and readable without technical expertise.

Connect with Patient Advocacy and Research Communities

• Your condition's patient advocacy organization (often has research grants, biobank programs, and clinical trial connections)
• Research hackathons — the model used in Bill's case: patient-organized computational analysis of personal genomic data with bioinformatics volunteers
• Academic medical centers with rare disease or precision oncology programs
• ClinicalTrials.gov — search your condition + molecular target (e.g., "papillary RCC PDGFRA")
• MatchMiner, TrialSpark, Massive Bio — AI-assisted clinical trial matching services