Founder Mode Oncology

Founder Mode Oncology

Systematic framework for navigating personalized cancer treatment, generalized from Sid Sijbrandij's osteosarcoma case (2022-2026). Transforms the ad-hoc "billionaire with a team" approach into a reproducible methodology using open-source tools and structured decision-making.

The Three-Pillar Framework

Pillar 1: Maximal Diagnostics

Run every available diagnostic modality to build a complete molecular picture. Standard clinical panels miss non-obvious targets.

Minimum diagnostic stack (in priority order):

1. Genomics: WGS + WES (tumor/normal paired) — somatic mutations, CNV, structural variants 2. Transcriptomics: Bulk RNA-seq + scRNA-seq (tumor + PBMCs) — gene expression, immune landscape, non-obvious targets 3. Liquid biopsy: ctDNA (tumor-informed, e.g. Signatera) + methylation-based (e.g. Northstar) — real-time monitoring 4. Functional testing: Organoid drug testing + mass response assays — empirical drug sensitivity 5. Imaging: Standard (CT/MRI) + novel PET tracers (68Ga-FAP, 68Ga-B7H3) — target validation 6. Flow cytometry: B/T cell subsets — immune status tracking

Critical insight: scRNA-seq reveals targets that standard panels miss. In the reference case, scRNA-seq identified FAP overexpression in osteosarcoma — invisible to gene panels and WES — enabling the breakthrough radioligand therapy.

Tissue handling: Always request cryopreserved (flash-frozen) samples alongside FFPE. FFPE destroys RNA quality needed for transcriptomics.

See references/diagnostics-pipeline.md for the complete open-source bioinformatics pipeline.

Pillar 2: Personalized Therapeutic Development

Use diagnostic findings to design patient-specific treatments. Access experimental drugs via FDA expanded access.

Treatment categories (layer compatible modalities):

Regulatory pathway: FDA Form 3926 (Individual Patient Expanded Access IND) — typically approved within 48 hours. The FDA is faster than hospital IRBs.

See references/treatment-categories.md for detailed treatment logic. See references/regulatory-access.md for expanded access navigation.

Pillar 3: Parallel Treatment

Run compatible therapies simultaneously. Monitor with ctDNA and serial scRNA-seq to measure what works.

Combination logic:

Checkpoint inhibitors  → Remove immune brakes (foundation layer)
  + Neoantigen vaccines → Train recognition (synergizes with checkpoint)
  + Oncolytic virus     → Kill + release antigens (synergizes with vaccines)
  + Cell therapy        → Innate killing (independent mechanism)
  + Radioligand         → Targeted kill to specific marker (independent)
  + Immune modulators   → Amplify all of the above

Monitoring cadence:

• ctDNA: every 2-4 weeks (real-time response measurement)
• scRNA-seq PBMCs: monthly (immune evolution tracking)
• Imaging (PET/CT/MRI): every 2-3 months
• Flow cytometry: monthly

Success metric: Immune infiltration shift (cold → hot tumor). Reference case: 19% → 89% T cells in tumor microenvironment.

See references/mrd-monitoring.md for liquid biopsy interpretation.

Decision Workflow

1. DIAGNOSE COMPREHENSIVELY
   ├── Order WGS + WES + RNA-seq + scRNA-seq
   ├── Establish ctDNA baseline (multiple platforms)
   ├── Request cryopreserved tissue (not just FFPE)
   └── Run functional drug testing (organoids if available)

2. IDENTIFY TARGETS
   ├── Standard: Known driver mutations → approved targeted therapies
   ├── Non-obvious: scRNA-seq → overexpressed surface proteins (FAP, B7H3, EphA2)
   ├── Validate: PET imaging with target-specific tracers (theranostic confirmation)
   └── Predict: Neoantigen candidates via pVACseq + MHCflurry

3. DESIGN TREATMENT COMBINATION
   ├── Foundation: Checkpoint inhibitor (if not contraindicated)
   ├── Layer: Neoantigen vaccine (peptide or mRNA)
   ├── Layer: One or more of: oncolytic virus, cell therapy, radioligand
   ├── Support: Immune modulators, bone protection, etc.
   └── Access: FDA Form 3926 for experimental agents

4. MONITOR AND ADAPT
   ├── ctDNA every 2-4 weeks → detect response or progression early
   ├── Serial scRNA-seq → track immune landscape evolution
   ├── Imaging every 2-3 months → structural assessment
   └── Adjust: Add/remove therapies based on molecular response

5. MAINTAIN REMISSION
   ├── Preventive vaccines (mRNA neoantigen, ongoing)
   ├── Continued monitoring (ctDNA, imaging)
   └── Backup: Engineered cell therapies with logic gates (if needed)

Team Structure

Cost Reality

Structural Biology Layer (AlphaFold + Protein Design)

Structure prediction adds 3D validation on top of the sequence-based pipeline. Four integration points:

1. Neoantigen Vaccine Validation

After pVACseq + MHCflurry rank candidates by sequence, validate top 20-50 with AlphaFold Multimer (peptide + HLA chain). Filter by ipTM >0.5, PAE_interface <10. Re-rank. This catches peptides that score well in 1D but don't physically fit the MHC groove.

2. Radioligand Target Modeling

Retrieve target structure from AlphaFold DB (FAP: AF-Q12884-F1, B7H3: AF-Q5ZPR3-F1). Model ligand binding with Chai/Boltz. Validate that diagnostic (68Ga) and therapeutic (177Lu/225Ac) versions maintain equivalent binding.

3. De Novo Therapeutic Binder Design

When no existing drug fits the target: RFdiffusion (2.8K stars) generates backbone geometries → ProteinMPNN (1.7K stars) designs sequences → ESMFold/AlphaFold2 validates structures. Tier 1 candidates: pLDDT >85, pTM >0.8.

4. Mutation Impact Analysis

For each somatic mutation: predict wildtype vs mutant structures. Surface-altering mutations on expressed proteins → neoantigen candidates. Destabilizing mutations → misfolded protein → immune recognition. ESM-2 embeddings for fast batch screening.

See references/structural-biology.md for the complete pipeline, tool stack, quality thresholds, and key target UniProt IDs.

---

Key Reference Case Data

• Source: osteosarc.com — 25TB open data (Google Cloud)
• Article: centuryofbio.com/p/sid — "Going Founder Mode on Cancer" by Elliot Hershberg
• Venture fund: evenone.ventures — scaling personalized oncology
• Research repo: ~/broomva/research/founder-mode-cancer/ — complete local analysis

References

• Diagnostics Pipeline — Open-source bioinformatics toolchain (BWA → GATK → ASCAT → Scanpy → pVACtools)
• Treatment Categories — Detailed treatment modalities, mechanisms, and combination rationale
• Regulatory Access — FDA expanded access, IRB navigation, tissue access strategies
• MRD Monitoring — Liquid biopsy platforms, interpretation, and cross-platform comparison
• Structural Biology — AlphaFold, RFdiffusion, ProteinMPNN for neoantigen validation and de novo binder design
• Open-Source Tools — GitHub repos, neoantigen vaccine pipelines, analysis tools

Contributing to Founder Mode Oncology

Thank you for your interest in improving this skill. Contributions from clinicians, researchers, bioinformaticians, patients, and caregivers are especially valued.

How to Contribute

Improving References

The references/ directory contains the detailed knowledge base. To improve an existing reference:

1. Fork this repository. 2. Edit the relevant file under references/. 3. Cite your sources -- peer-reviewed publications, FDA documents, or tool documentation. 4. Open a pull request with a clear description of what changed and why.

Do not modify SKILL.md directly unless the change is structural (e.g., adding a new pillar or section). SKILL.md summarizes the references; update the reference first, then propose a corresponding SKILL.md change if needed.

Adding New Treatment Modalities

If a treatment category is missing (e.g., a new class of therapy, a novel diagnostic modality):

1. Add detailed content to the appropriate reference file, or create a new file under references/ if no existing file fits. 2. Include: mechanism of action, clinical evidence, access pathway, combination compatibility, and monitoring approach. 3. Add links to relevant GitHub repositories, clinical trial registries (clinicaltrials.gov), or FDA databases where applicable. 4. Update references/open-source-tools.md if the modality involves new computational tools.

Submitting Corrections

Medical knowledge evolves. If you find outdated information, incorrect dosing, withdrawn drugs, or factual errors:

1. Open an issue describing the error, citing the correct information and its source. 2. If you can fix it yourself, open a pull request instead. 3. Label corrections with the correction label if possible.

For urgent safety-related corrections (e.g., a listed drug combination is now known to be dangerous), open an issue with [SAFETY] in the title.

General Guidelines

• Be specific: Include citations, tool versions, and links.
• Keep the tone clinical: This skill is used by AI agents assisting real decisions. Precision matters.
• Preserve structure: Follow the existing format of whichever file you edit.
• One concern per PR: Keep pull requests focused on a single topic.

Issue Templates

When opening an issue, use one of these formats:

Correction:

• File affected: references/_____.md
• Current text (quote)
• Correct information
• Source / citation

New content proposal:

• Topic / modality
• Why it belongs in this skill
• Key references (2--3 links)

Tool update:

• Tool name and repository link
• What changed (new version, deprecated, superseded)
• Impact on the pipeline described in this skill

Code of Conduct

Be respectful. This project exists to help people navigate cancer treatment. Contributions should prioritize accuracy, clarity, and patient benefit.

License

By contributing, you agree that your contributions will be licensed under the MIT License.

Founder Mode Oncology

Personalized cancer treatment navigation — maximal diagnostics, parallel therapy, therapeutic development, structure-based protein design.

 

What This Skill Does

An agent skill that encodes a systematic framework for navigating personalized cancer treatment, generalized from Sid Sijbrandij's osteosarcoma case (2022--2026). It transforms the ad-hoc "billionaire with a team" approach into a reproducible methodology using open-source tools and structured decision-making.

For full skill content and clinical detail, see SKILL.md.

Quick Install

npx skills add broomva/founder-mode-oncology

Supported Agents

This skill works with any agent that supports the skills.sh format:

• Claude Code (Anthropic)
• Cursor
• Codex (OpenAI)
• Gemini CLI (Google)
• Windsurf
• Amp
• Any agent compatible with SKILL.md conventions

Skill Structure

founder-mode-oncology/
├── SKILL.md                              # Main skill file (agent-readable)
├── references/
│   ├── diagnostics-pipeline.md           # Open-source bioinformatics toolchain
│   ├── treatment-categories.md           # Treatment modalities and combination rationale
│   ├── regulatory-access.md              # FDA expanded access and IRB navigation
│   ├── mrd-monitoring.md                 # Liquid biopsy interpretation
│   ├── structural-biology.md             # AlphaFold, RFdiffusion, ProteinMPNN pipelines
│   └── open-source-tools.md              # GitHub repos and analysis tools
├── README.md
├── LICENSE
└── CONTRIBUTING.md

The Three Pillars

1. Maximal Diagnostics

Run every available diagnostic modality -- WGS, WES, RNA-seq, scRNA-seq, liquid biopsy, organoid drug testing, novel PET tracers -- to build a complete molecular picture. Standard clinical panels miss non-obvious targets; scRNA-seq in the reference case revealed FAP overexpression invisible to gene panels.

2. Personalized Therapeutic Development

Use diagnostic findings to design patient-specific treatments spanning checkpoint inhibitors, neoantigen vaccines, oncolytic viruses, cell therapies, radioligand therapy, and immune modulators. Access experimental drugs via FDA expanded access (Form 3926, typically approved within 48 hours).

3. Parallel Treatment

Run compatible therapies simultaneously rather than sequentially. Monitor with ctDNA every 2--4 weeks and serial scRNA-seq monthly to measure response in real time and adapt. The reference case achieved a T-cell infiltration shift from 19% to 89% in the tumor microenvironment.

Key Links

• Open patient data (25 TB): osteosarc.com
• Source article: centuryofbio.com/p/sid -- "Going Founder Mode on Cancer" by Elliot Hershberg
• Venture fund scaling personalized oncology: evenone.ventures
• Research repository: github.com/broomva/founder-mode-cancer

Contributing

Contributions are welcome -- especially corrections from clinicians, researchers, and patients. See CONTRIBUTING.md for guidelines on improving references, adding treatment modalities, and submitting corrections.

License

MIT -- Carlos D. Escobar-Valbuena (@broomva), 2026.

Diagnostics Pipeline — Open-Source Bioinformatics Toolchain

Table of Contents

• End-to-End Pipeline
• Step 1: Alignment
• Step 2: Variant Calling
• Step 3: Copy Number Analysis
• Step 4: Single-Cell Analysis
• Step 5: Gene Set Enrichment
• Step 6: Neoantigen Prediction
• Step 7: Vaccine Design
• Step 8: Visualization

---

End-to-End Pipeline

FASTQ (WGS/WES/RNA-seq/scRNA-seq)
  │
  ├─ DNA ──→ BWA align ──→ GATK process ──→ Mutect2 + Strelka2 (variants)
  │                                      └──→ ASCAT (copy number)
  │
  ├─ RNA ──→ STAR align ──→ featureCounts/HTSeq (quantification)
  │                      └──→ Limma/DESeq2 (differential expression + GSEA)
  │
  ├─ scRNA ─→ Cell Ranger ──→ Scanpy/Seurat (clustering, DE, target discovery)
  │
  └─ Variants + HLA ──→ pVACseq (neoantigen prediction)
                      └──→ MHCflurry (MHC binding)
                      └──→ Vaxrank (candidate ranking)
                      └──→ pVACvector (peptide vaccine design)
                      └──→ LinearDesign (mRNA optimization)

Step 1: Alignment

DNA alignment — BWA

• Repo: https://github.com/lh3/bwa
• Command: bwa mem -t 16 ref.fa tumor_R1.fq.gz tumor_R2.fq.gz | samtools sort -o tumor.bam
• Reference: GRCh38 (or GRCh37+decoy for OpenVax pipeline compatibility)

RNA alignment — STAR

• Repo: https://github.com/alexdobin/STAR
• Two-pass alignment recommended for novel junction discovery
• Command: STAR --genomeDir ref --readFilesIn R1.fq.gz R2.fq.gz --outSAMtype BAM SortedByCoordinate

Step 2: Variant Calling

GATK Mutect2 (primary somatic caller)

• Repo: https://github.com/broadinstitute/gatk
• Pipeline: MarkDuplicates → BaseRecalibration → Mutect2 (tumor/normal paired)
• Output: VCF with somatic SNVs and indels
• Filter: FilterMutectCalls for quality control

Strelka2 (validation caller)

• Repo: https://github.com/Illumina/strelka
• Run as secondary caller; intersect with Mutect2 for high-confidence calls

Step 3: Copy Number Analysis

ASCAT (Allele-Specific Copy number Analysis of Tumours)

• Repo: https://github.com/VanLoo-lab/ascat
• Convenience wrapper: https://github.com/CompEpigen/ezASCAT
• Input: Tumor/normal BAMs from WGS
• Output: Allele-specific copy number segments, ploidy, purity estimates
• Use at multiple timepoints to track clonal evolution

Step 4: Single-Cell Analysis

Cell Ranger (10x Genomics processing)

• Download: https://www.10xgenomics.com/support/software/cell-ranger
• Processes 10x Chromium scRNA-seq FASTQ → gene expression matrix
• Output: Filtered feature-barcode matrix (H5/MTX)

Scanpy (Python) or Seurat (R) — downstream analysis

• Scanpy: https://github.com/scverse/scanpy
• Seurat: https://github.com/satijalab/seurat

Standard workflow: 1. Quality control (filter doublets, dead cells) 2. Normalization + highly variable gene selection 3. PCA → UMAP/t-SNE dimensionality reduction 4. Clustering (Leiden/Louvain) 5. Differential expression per cluster 6. Cell type annotation 7. Target discovery: Identify surface proteins overexpressed on tumor clusters but not normal tissue (FAP, B7H3, EphA2, etc.)

Critical: This is where non-obvious targets are discovered. Standard genomics misses transcriptomic targets.

Step 5: Gene Set Enrichment

Limma (R/Bioconductor)

• Linear models for differential expression
• limma::camera() for competitive GSEA

DESeq2 (R/Bioconductor)

• Negative binomial model for count data
• Use with fgsea for preranked GSEA

Gene set databases: MSigDB (Hallmark, C2, C5), Reactome, KEGG

Step 6: Neoantigen Prediction

pVACtools (Griffith Lab, Washington University)

• Repo: https://github.com/griffithlab/pVACtools
• Components:
• pVACseq: Predict neoantigens from somatic mutations + HLA type
• pVACbind: Predict binding from peptide sequences
• pVACvector: Design optimal vaccine peptide sequences

MHCflurry (OpenVax, Mount Sinai)

• Repo: https://github.com/openvax/mhcflurry
• Neural network MHC-I peptide binding prediction
• More accurate than older methods (NetMHC)

Vaxrank (OpenVax)

• Repo: https://github.com/openvax/vaxrank
• Integrates variant calls + RNA expression + MHC binding → ranked vaccine candidates
• Prioritizes by: binding affinity, expression level, variant allele frequency

Input requirements:

• Somatic VCF (from Mutect2/Strelka2)
• RNA-seq BAM (for expression verification)
• HLA typing (Class I alleles) — from OptiType or clinical HLA typing

Step 7: Vaccine Design

Peptide vaccines:

• pVACvector arranges selected neoantigens into optimal peptide sequence
• Minimizes junctional epitopes (false neoepitopes at peptide junctions)

mRNA vaccines:

• LinearDesign (Stanford): Optimize mRNA sequence for stable secondary structure
• Codon optimization for translation efficiency
• Add structural elements: 5' cap, 5' UTR, signal peptide, poly-A tail

Physical production (from OpenVaxx guide): 1. DNA synthesis (BioXp, ~$600) 2. In vitro transcription (T7 polymerase, ~$2K) 3. LNP formulation (microfluidic mixing, ~$500) 4. QC (DLS particle sizing, ~$100)

• Total: ~$4.2K in-house per patient

Step 8: Visualization

IGV (Integrative Genomics Viewer)

• Repo: https://github.com/igvteam/igv
• Web version: https://github.com/igvteam/igv.js
• Browse WGS, WES, RNA-seq, scRNA-seq BAMs in context

UCSC Genome Browser — public track hubs for comparison

---

Hardware Requirements

Total pipeline: ~24-48h on a 16-core machine with 64GB RAM.

MRD Monitoring — Liquid Biopsy Interpretation

Table of Contents

• Overview
• Platform Comparison
• Interpretation Guide
• Monitoring Cadence
• Actionable Thresholds
• Complementary Monitoring

---

Overview

Minimum Residual Disease (MRD) monitoring via circulating tumor DNA (ctDNA) enables real-time treatment response measurement — weeks before imaging shows changes. This is the primary feedback loop for the parallel treatment strategy.

Principle: Tumor cells shed DNA fragments into the bloodstream. Detecting and quantifying these fragments tells you: 1. Whether cancer is present (even below imaging threshold) 2. Whether treatment is working (declining ctDNA = response) 3. Whether cancer is recurring (rising ctDNA = progression)

---

Platform Comparison

Tumor-Informed Assays

Custom panel designed from the patient's specific tumor mutations.

Signatera (Natera) — MTM/mL (mean tumor molecules per mL)

• Requires: Prior tumor sequencing to design personalized panel (16 variants tracked)
• Sensitivity: Can detect down to 0.01 MTM/mL
• Specificity: Very high (custom panel minimizes false positives)
• Turnaround: 7-10 business days
• Cost: ~$3,000-$5,000 per test
• Best for: Definitive clearance confirmation, recurrence detection

Personalis NeXT Personal — PPM (parts per million)

• WGS-based ctDNA quantification
• Ultra-sensitive (detects at single-digit PPM)
• Logarithmic scale provides dynamic response range
• Best for: Early response measurement (large dynamic range)

Tumor-Agnostic Assays

Detect cancer signal without prior tumor information.

Northstar — TMS (Tumor Methylation Signal)

• Methylation-based detection (epigenetic, not mutation-based)
• Does not require prior tumor sequencing
• Never reaches zero (background methylation signal)
• Range observed: 10-26 in reference case
• Best for: Trend monitoring, complementary to mutation-based assays
• Limitation: Cannot confirm true clearance (non-zero floor)

Comparison Matrix

---

Interpretation Guide

Declining ctDNA

Pattern: Progressive decrease over 2-4 readings Interpretation: Treatment is working Action: Continue current regimen. Layer additional therapies if decline is slow.

Rapidly Declining ctDNA

Pattern: >90% drop within 2-4 weeks (e.g., Personalis 963 → 21 PPM) Interpretation: Strong treatment response Action: Continue. Consider adding maintenance therapy (vaccine) to sustain.

Stable Low ctDNA

Pattern: Hovering at low but detectable levels Interpretation: Residual disease persists, treatment containing but not eliminating Action: Add new modality (radioligand, oncolytic virus, or updated vaccine).

Rising ctDNA

Pattern: Two consecutive increases Interpretation: Treatment resistance or progression Action: Urgent re-evaluation. New biopsy for updated sequencing. Switch or add therapies.

Undetectable ctDNA

Pattern: Zero on tumor-informed assay (Signatera = 0) Interpretation: No detectable molecular residual disease Action: Continue preventive vaccines. Extend monitoring intervals to monthly, then quarterly.

Blip (Transient Spike)

Pattern: Single elevated reading followed by return to baseline Interpretation: May reflect: tumor cell death from treatment (positive), inflammation, or assay noise Action: Repeat in 2 weeks before changing therapy. Correlate with clinical context (recent treatment, infection, surgery).

Discordance Between Platforms

Pattern: One platform shows signal, another doesn't Interpretation: Different assays have different sensitivities and specificities

Resolution:

• Signatera positive, Northstar negative: Trust Signatera (tumor-informed > agnostic)
• Northstar elevated, Signatera negative: Likely non-tumor methylation (inflammation, aging)
• Both positive: High confidence of disease
• Both negative: High confidence of clearance

---

Monitoring Cadence

---

Actionable Thresholds

ND = Not Detected

---

Complementary Monitoring

ctDNA alone is insufficient. Combine with:

Flow Cytometry

• Track B/T cell subsets (CD4, CD8, NK, MAIT cells)
• Monitor immune reconstitution during immunotherapy
• Declining T cells may indicate immunosuppression or disease progression

Serial scRNA-seq (PBMCs)

• Monthly during active treatment
• Reveals immune cell composition changes at single-cell resolution
• Can identify: T cell exhaustion markers, new immune populations, clonal expansion
• More informative than flow cytometry but more expensive

Imaging (PET/CT/MRI)

• Every 2-3 months during active treatment
• ctDNA detects molecular changes weeks before imaging shows structural changes
• Imaging confirms structural response after ctDNA response
• Novel PET tracers (68Ga-FAP, 68Ga-B7H3) for target-specific imaging

Tumor Markers (if applicable)

• ALP (alkaline phosphatase) for bone cancers
• PSA for prostate
• CA-125 for ovarian
• CEA for colorectal
• Less specific than ctDNA but cheap and widely available

ELISPOT

• Measures T cell reactivity to specific peptides (vaccine antigens)
• Confirms vaccine is generating immune response
• Run after each vaccine dose series

Open-Source Tools for Personalized Oncology

Table of Contents

• Neoantigen Vaccine Pipelines
• Genomics & Variant Calling
• Single-Cell Analysis
• Copy Number & Structural Variants
• MHC Binding Prediction
• RNA-seq & Expression Analysis
• Visualization & Browsers
• mRNA Vaccine Synthesis Guide
• Osteosarcoma-Specific Repos

---

Neoantigen Vaccine Pipelines

openvax/neoantigen-vaccine-pipeline

• URL: https://github.com/openvax/neoantigen-vaccine-pipeline
• Stars: 90 | Language: Python
• Status: Used in 2 Phase I clinical trials (NCT02721043, NCT03223103)
• What it does: End-to-end pipeline from FASTQ to ranked vaccine peptide candidates
• Pipeline: BWA align → GATK process → Mutect2/Strelka variant call → Vaxrank ranking
• Requirements: 16+ cores, 32GB RAM, ~500GB disk, tumor/normal WES + RNA-seq + HLA typing
• Docker available: Yes (recommended deployment)

griffithlab/pVACtools

• URL: https://github.com/griffithlab/pVACtools
• Stars: 300+ | Language: Python
• What it does: Suite of tools for personalized variant antigen prediction
• Components:
• pVACseq: Predict neoantigens from somatic mutations
• pVACbind: Predict binding from peptide sequences
• pVACvector: Design optimal vaccine peptide order (minimizes junctional epitopes)
• pVACview: Visualization of neoantigen candidates
• Input: Annotated VCF + HLA alleles
• Documentation: https://pvactools.readthedocs.io/

openvax/vaxrank

• URL: https://github.com/openvax/vaxrank
• Language: Python
• What it does: Ranks vaccine peptide candidates by integrating variant calls + RNA expression + MHC binding
• Prioritization criteria: Binding affinity, expression level, variant allele frequency

---

Genomics & Variant Calling

broadinstitute/gatk

• URL: https://github.com/broadinstitute/gatk
• What it does: Industry-standard variant calling toolkit
• Key tools: Mutect2 (somatic), HaplotypeCaller (germline), MarkDuplicates, BQSR
• Best practices: https://gatk.broadinstitute.org/hc/en-us/sections/360007226651-Best-Practices-Workflows

Illumina/strelka

• URL: https://github.com/Illumina/strelka
• What it does: Fast somatic SNV/indel caller (validation against Mutect2)

lh3/bwa

• URL: https://github.com/lh3/bwa
• What it does: DNA read alignment to reference genome (GRCh38)

samtools/samtools

• URL: https://github.com/samtools/samtools
• What it does: BAM/CRAM file manipulation, sorting, indexing, statistics

broadinstitute/picard

• URL: https://github.com/broadinstitute/picard
• What it does: BAM utilities (duplicate marking, metrics collection)

---

Single-Cell Analysis

10x Genomics Cell Ranger

• URL: https://www.10xgenomics.com/support/software/cell-ranger
• License: Free download (proprietary)
• What it does: Process 10x Chromium scRNA-seq raw data → gene expression matrix

scverse/scanpy

• URL: https://github.com/scverse/scanpy
• Stars: 2,000+ | Language: Python
• What it does: Full scRNA-seq analysis: QC, normalization, clustering, DE, visualization

satijalab/seurat

• URL: https://github.com/satijalab/seurat
• Stars: 2,500+ | Language: R
• What it does: Same as Scanpy but in R ecosystem. Industry standard.

scverse/anndata

• URL: https://github.com/scverse/anndata
• What it does: Data structure for single-cell data (the .h5ad format)

scverse/scvi-tools

• URL: https://github.com/scverse/scvi-tools
• What it does: Deep learning for single-cell analysis (batch correction, imputation)

---

Copy Number & Structural Variants

VanLoo-lab/ascat

• URL: https://github.com/VanLoo-lab/ascat
• What it does: Allele-specific copy number analysis from WGS/WES
• Output: Segments with allele-specific copy number, tumor purity, ploidy

CompEpigen/ezASCAT

• URL: https://github.com/CompEpigen/ezASCAT
• Stars: 12 | Language: R
• What it does: Convenient ASCAT wrapper for BAM input

dellytools/delly

• URL: https://github.com/dellytools/delly
• What it does: Structural variant discovery (deletions, duplications, inversions, translocations)

---

MHC Binding Prediction

openvax/mhcflurry

• URL: https://github.com/openvax/mhcflurry
• Stars: 200+ | Language: Python
• What it does: Neural network MHC-I peptide binding prediction
• Advantage: Open-source alternative to NetMHC (which has license restrictions)

IEDB Analysis Tools

• URL: http://tools.iedb.org/
• What it does: Comprehensive immune epitope prediction (free web interface)
• Includes: MHC-I binding, MHC-II binding, T cell epitope prediction, B cell epitope prediction

OptiType

• URL: https://github.com/FRED-2/OptiType
• What it does: HLA typing from sequencing data (needed as input for neoantigen prediction)

---

RNA-seq & Expression Analysis

alexdobin/STAR

• URL: https://github.com/alexdobin/STAR
• What it does: RNA-seq alignment (splicing-aware)

subread/featureCounts

• URL: https://github.com/ShiLab-Bioinformatics/subread
• What it does: Read quantification against gene annotations

DESeq2 (Bioconductor)

• URL: https://bioconductor.org/packages/DESeq2/
• What it does: Differential expression analysis + GSEA

Limma (Bioconductor)

• URL: https://bioconductor.org/packages/limma/
• What it does: Linear models for microarray/RNA-seq differential expression

subinoy/fgsea

• URL: https://github.com/ctlab/fgsea
• What it does: Fast preranked gene set enrichment analysis

---

Visualization & Browsers

igvteam/igv

• URL: https://github.com/igvteam/igv
• What it does: Desktop genome browser for BAM/VCF/BED visualization

igvteam/igv.js

• URL: https://github.com/igvteam/igv.js
• What it does: Web-embedded genome browser (used on osteosarc.com)

---

mRNA Vaccine Synthesis Guide

philfung/openvaxx

• URL: https://github.com/philfung/openvaxx
• Stars: 67 | Language: JavaScript (interactive guide)
• What it does: Complete open-source guide: sequencing → mutation detection → AI target selection → mRNA synthesis → LNP formulation → QC
• Interactive guide: https://philfung.github.io/openvaxx/

8-Step Pipeline Summary: 1. Genomic sequencing ($1K-$2.5K) 2. Mutation detection (GATK Mutect2) 3. AI target selection (pVACseq + MHCflurry) 4. Sequence optimization (pVACvector + LinearDesign) 5. DNA synthesis (BioXp, $600) 6. mRNA transcription ($2K) 7. LNP formulation ($500) 8. Quality assurance ($100)

Total: ~$4.2K in-house, ~$13.4K outsourced per patient Timeline: 4-6 weeks biopsy to vial Equipment capital: $500K-$800K for in-house lab

---

Osteosarcoma-Specific Repos

---

Open Data Reference

Sid Sijbrandij's dataset: 25TB on Google Cloud (publicly readable)

• Portal: https://osteosarc.com
• Contents: WGS, WES, RNA-seq, scRNA-seq, ONT long-read, spatial transcriptomics, H&E, IHC, HLA typing
• Contact: cancer@sytse.com

Regulatory Access and Navigation

Table of Contents

• FDA Expanded Access
• Hospital IRB Navigation
• Tissue Access Strategies
• Data Portability
• International Access

---

FDA Expanded Access

Individual Patient IND (Form 3926)

The fastest path to experimental drugs outside clinical trials.

What it is: An FDA mechanism allowing a single patient to access an investigational drug when: 1. The patient has a serious or life-threatening condition 2. No comparable alternative therapy is available 3. The potential benefit justifies the potential risk 4. Access will not interfere with ongoing clinical trials

Process: 1. Identify the experimental drug and its manufacturer/sponsor 2. Physician submits FDA Form 3926 (simplified IND application) 3. FDA reviews and responds — typically within 48 hours (emergency: 24 hours by phone) 4. Drug manufacturer must also agree to supply the drug 5. Local IRB must approve (this is often the slower step)

Key insight from reference case: The FDA was never the bottleneck. Every Form 3926 was approved within 48 hours. Hospital IRBs were the real friction point.

Form 3926 requirements:

• Patient diagnosis and treatment history
• Rationale for the specific drug
• Proposed dosing schedule
• Physician's assessment of risk/benefit
• IRB approval (can be concurrent)

Resources:

• FDA guidance: https://www.fda.gov/drugs/investigational-new-drug-ind-application/expanded-access
• Reagan-Udall Foundation: https://navigator.reaganudall.org/ (expanded access navigator)

Emergency IND

For immediately life-threatening situations:

• FDA can authorize by phone within 24 hours
• Written submission (Form 3926) follows within 15 working days
• No prior IRB approval required (must notify IRB within 5 days)

Right to Try Act (2018)

Alternative to expanded access:

• Patient has exhausted approved options
• Drug has completed Phase I trial
• No FDA application required (direct patient-manufacturer)
• Less regulatory oversight, fewer data collection requirements
• Many manufacturers prefer the Form 3926 route regardless

---

Hospital IRB Navigation

The Problem

Hospital IRBs (Institutional Review Boards) operate as independent gatekeepers. Unlike the FDA:

• No standardized timelines
• No standardized criteria for expanded access
• Single members can block access (vetocracy)
• Each hospital has different procedures
• Many IRBs are unfamiliar with expanded access

Strategies

1. Choose hospitals with expanded access experience:

• Major cancer centers (MSKCC, MD Anderson, Dana-Farber) have streamlined processes
• Academic medical centers with active clinical trial programs
• Ask upfront: "What is your IRB's typical turnaround for expanded access?"

2. Parallel submission:

• Submit FDA Form 3926 and IRB application simultaneously
• Do not wait for FDA approval before starting IRB process

3. Use IRB chair direct communication:

• Request a meeting with the IRB chair to explain the case
• Provide a concise summary: diagnosis, failed treatments, proposed drug, rationale
• Frame as individual patient compassionate use, not research

4. Escalation path:

• If IRB blocks: request written reasons
• Consider transferring care to a different institution with faster IRB
• Patient advocacy organizations can sometimes intervene

5. Central/commercial IRBs:

• Some expanded access can use commercial IRBs (WCG, Advarra)
• Faster than hospital-specific IRBs
• Confirm with the treating physician and hospital

---

Tissue Access Strategies

The FFPE Problem

Standard hospital pathology: biopsy → formalin fixation → paraffin embedding (FFPE). This:

• Preserves tissue morphology for pathology slides
• Destroys RNA quality (critical for scRNA-seq and RNA-seq)
• Cross-links proteins (limits proteomics)
• Is the only method most hospital pathology labs support

What to Request

Before any biopsy or surgery, communicate in writing: 1. Request that tissue be split: part FFPE (for clinical pathology), part flash-frozen (for research) 2. Provide cryopreservation instructions and shipping containers 3. Identify the receiving lab (with MTA in place) 4. Confirm with the surgeon AND pathology department (both must agree)

Cryopreservation protocol (basic):

• Tissue placed in cryovial within 30 minutes of excision
• Snap-freeze in liquid nitrogen or isopentane on dry ice
• Store at -80C or liquid nitrogen
• Ship on dry ice with temperature monitoring

Material Transfer Agreement (MTA)

Required for sending patient tissue between institutions:

• Initiate MTA process weeks before the planned procedure
• Required parties: sending hospital, receiving lab, patient consent
• Many hospitals have standard MTA templates
• Allow 2-4 weeks for processing

Patient Data Access

Tissue samples are the patient's property. Key rights:

• HIPAA Right of Access: patients can request copies of medical records including pathology
• 21st Century Cures Act: prohibits information blocking
• Direct-to-patient sequencing: services like Tempus, Foundation Medicine can work directly with patients

---

Data Portability

Getting Raw Sequencing Data

Clinical sequencing companies (Tempus, Foundation Medicine, etc.) typically provide:

• Summary report: Mutations, treatment recommendations (this is default)
• Raw data (FASTQ/BAM): Must be specifically requested

How to request raw data: 1. Contact the sequencing company's patient data access team 2. Submit a formal data release request (usually requires patient signature) 3. Specify format: FASTQ (preferred for reanalysis) or BAM 4. Specify delivery: cloud transfer (preferred for large files) or physical media 5. Typical turnaround: 2-4 weeks

Building a Personal Health Record

Aggregate all data in one place:

• Sequencing reports and raw data
• Imaging (DICOM files from radiology)
• Pathology reports and images
• Lab results (CBC, metabolic panels, tumor markers)
• Treatment records
• Clinical notes

Tools: Apple Health Records (for basic labs), patient portal exports, manual aggregation

---

International Access

Radioligand Therapy

FAP-targeted radioligand therapy (177Lu-FAPi, 225Ac-FAPi) is more accessible in:

• Germany: Multiple centers (University Hospital Heidelberg, LMU Munich)
• Australia: Theranostics Australia
• India: Several nuclear medicine centers

Process: 1. Obtain diagnostic PET scan (68Ga-FAP) to confirm target expression 2. Contact international center with PET images and medical records 3. Arrange travel and accommodation 4. Typical course: 1-3 treatments, 4-8 weeks apart 5. Follow-up imaging and monitoring can be done locally

Neoantigen Vaccines

• Germany: CeGaT (peptide vaccines)
• US: Multiple academic centers (MSKCC, Dana-Farber, MD Anderson)
• Self-manufacture: OpenVaxx guide (research/educational, not clinical)

Clinical Trial Search

• ClinicalTrials.gov: https://clinicaltrials.gov/
• WHO ICTRP: https://trialsearch.who.int/
• EU Clinical Trials Register: https://www.clinicaltrialsregister.eu/
• Filter by: cancer type + "expanded access" or "compassionate use"

Structural Biology for Personalized Oncology

Table of Contents

• Overview
• Neoantigen Vaccine Enhancement
• Radioligand Target Modeling
• De Novo Therapeutic Protein Design
• Mutation Impact Analysis
• Key Target Structures
• Tool Stack
• Quality Thresholds

---

Overview

Structure prediction adds a 3D validation layer to the sequence-based pipeline. Instead of predicting binding from sequence alone (MHCflurry), validate that the peptide physically fits the MHC groove (AlphaFold Multimer). Instead of confirming a target via PET alone, model the ligand-protein binding interface.

Where it fits in the three pillars:

• Pillar 1 (Diagnostics): Mutation impact analysis, target structure characterization
• Pillar 2 (Therapeutics): Vaccine validation, radioligand optimization, de novo binder design
• Pillar 3 (Parallel): Structural prioritization of which modalities to combine

---

Neoantigen Vaccine Enhancement

Enhance the pVACseq → MHCflurry → Vaxrank pipeline with 3D structure validation.

Enhanced pipeline:

Mutations → pVACseq → MHCflurry (1D binding) → TOP 20-50 CANDIDATES
  → AlphaFold Multimer (peptide + HLA chain) → 3D binding validation
  → Filter: ipTM >0.5, PAE_interface <10, pLDDT >85
  → Re-rank by structural confidence
  → Select top 10-20 for vaccine

Why this helps:

• MHCflurry predicts binding affinity from sequence — fast but approximate
• AlphaFold Multimer predicts actual 3D complex — reveals physical binding geometry
• Peptides scoring high on both = strongest candidates
• Structure reveals which residues are solvent-exposed (available for TCR recognition)
• The neoantigen-specific residue must face outward for T cells to distinguish it from wildtype

Implementation:

# Create FASTA with peptide + HLA-A*02:01 heavy chain
# Run ColabFold multimer
modal run modal_colabfold.py \
  --input-faa peptide_mhc_pairs.fasta \
  --out-dir af_neoantigen_validation/

Evaluation (using AlphaFold result files):

• Extract ipTM and pLDDT from result pkl files
• Filter: ipTM >0.5 and pLDDT >85 → PASS
• Re-rank passing candidates by ipTM descending

---

Radioligand Target Modeling

Model the binding interface between tumor targets and targeting ligands.

Workflow:

1. Retrieve target structure from AlphaFold DB
   → FAP: AF-Q12884-F1
   → B7H3: AF-Q5ZPR3-F1
   → EphA2: AF-P29317-F1

2. Identify surface-exposed druggable pockets
   → Catalytic domain (for FAP: dipeptidyl peptidase activity)
   → Extracellular domains (for B7H3, EphA2)

3. Model ligand binding
   → Existing ligands: FAPI-04, FAPI-46 (for FAP)
   → Use Chai or Boltz for protein-ligand complex prediction
   → Assess binding pose and contact residues

4. Optimize
   → If binding is suboptimal, modify ligand chemistry
   → Validate with re-docking
   → Same ligand scaffold carries 68Ga (imaging) or 177Lu/225Ac (therapy)

Theranostic validation: Structural modeling can predict whether the diagnostic (68Ga-labeled) and therapeutic (177Lu/225Ac-labeled) versions maintain equivalent binding — critical for the theranostic principle.

---

De Novo Therapeutic Protein Design

When no existing drug targets the identified surface, design custom proteins.

Full pipeline:

Phase 1: Target Characterization
  → AlphaFold DB or AlphaFold2 prediction of target
  → Identify binding epitope (tumor-specific, surface-exposed)
  → InterPro domain analysis

Phase 2: Backbone Generation
  → RFdiffusion: generate ≥5 backbone geometries complementary to epitope
  → Filter by geometry and contact surface area
  → Select top 3-5 backbones

Phase 3: Sequence Design
  → ProteinMPNN: design ≥8 sequences per backbone
  → Sample at temperature 0.1-0.3 for diversity
  → MPNN score >0.6

Phase 4: Structure Validation
  → ESMFold (fast): screen all candidates
  → AlphaFold2 (accurate): validate top candidates
  → Criteria: pLDDT >85, pTM >0.8

Phase 5: Developability
  → Aggregation propensity (low is better)
  → Isoelectric point (neutral range preferred)
  → Expression prediction
  → Immunogenicity assessment

Phase 6: Output
  → Ranked candidates with FASTA sequences
  → Structural models (.pdb)
  → Experimental recommendations

Tools:

Use cases in oncology:

• Design binder against FAP extracellular domain → fusion with radionuclide carrier
• Design binder against B7H3 → CAR construct or bispecific antibody
• Design binder against neoantigen-MHC complex → synthetic TCR mimic

---

Mutation Impact Analysis

Understand structural consequences of each somatic mutation.

For each somatic mutation from WES/WGS:
  1. Get wildtype protein sequence (UniProt)
  2. Create mutant sequence (apply SNV)
  3. Predict both structures (ESMFold for speed, AF2 for accuracy)
  4. Compare:
     - RMSD (structural deviation)
     - pLDDT change (confidence shift → stability impact)
     - Surface exposure change (buried → exposed = potential neoantigen)
  5. Classify:
     - Surface-altering on expressed protein → neoantigen candidate
     - Destabilizing (large RMSD, pLDDT drop) → misfolded protein → immune recognition
     - Neutral → deprioritize for vaccine

ESM-2 embeddings can also predict mutation effects:
  → Compute log-likelihood ratio (wildtype vs mutant)
  → Large negative ratio = destabilizing mutation
  → Faster than full structure prediction for screening

---

Key Target Structures

Retrieve from AlphaFold DB (alphafold.ebi.ac.uk):

---

Tool Stack

---

Quality Thresholds

Tier grading:

• T1 (best): pLDDT >85, pTM >0.8, low aggregation — proceed to synthesis
• T2: pLDDT >75, pTM >0.7 — consider with additional validation
• T3: pLDDT >70, pTM >0.65 — redesign recommended
• T4: Below thresholds — reject

Treatment Categories and Combination Logic

Table of Contents

• Checkpoint Inhibitors
• Neoantigen Vaccines
• Oncolytic Viruses
• Cell Therapies
• Radioligand Therapy
• Immune Modulators
• Targeted Therapy
• Combination Rationale
• Contraindications and Monitoring

---

Checkpoint Inhibitors

Mechanism: Block inhibitory receptors on T cells, removing the "brakes" on anti-tumor immunity.

Role in combination: Foundation layer. Required for other immunotherapies to work — if the brakes are on, no amount of immune activation matters.

Common combination: PD-1 + CTLA-4 dual blockade (e.g., Dostarlimab + Ipilimumab).

Monitoring: irAEs (immune-related adverse events) — thyroiditis, colitis, hepatitis, pneumonitis. Thyroid dysfunction is actually a positive prognostic sign.

---

Neoantigen Vaccines

Mechanism: Train the immune system to recognize tumor-specific mutations (neoantigens) that are absent from normal tissue.

Pipeline: Tumor sequencing → HLA typing → neoantigen prediction (pVACseq) → MHC binding prediction (MHCflurry) → candidate ranking (Vaxrank) → manufacturing

Iterative versions: As the tumor evolves and new sequencing is done, update the vaccine (JLFv1 → v2 → v3 in reference case).

Adjuvants: GM-CSF (granulocyte-macrophage colony-stimulating factor) enhances antigen presentation. Often co-administered.

Monitoring: ELISPOT assays measure T cell reactivity to vaccine peptides. Variant allele frequency tracking confirms target persistence.

---

Oncolytic Viruses

Mechanism: Engineered viruses that selectively infect and lyse cancer cells. Tumor lysis releases antigens, creating an in situ vaccine effect.

Routes: Intratumoral (IT) for accessible tumors, subcutaneous (SQ) for systemic priming.

Synergy: Complements checkpoint inhibitors and vaccines. Virus kills cells → releases neoantigens → vaccines have primed T cells against those antigens → checkpoint inhibitors ensure T cells can act.

---

Cell Therapies

Advanced: Genetic logic gates in cell therapies — engineered circuits that activate killing only when multiple tumor-specific signals are detected. Reduces off-target toxicity.

---

Radioligand Therapy

Mechanism: Conjugate a radioactive isotope to a molecule that targets a tumor-specific surface protein. Delivers radiation directly to cancer cells.

Theranostic principle: Same targeting molecule used for: 1. Diagnosis: 68Ga (gallium-68, PET imaging) — "does the tumor express the target?" 2. Therapy: 177Lu (lutetium-177, beta emitter) or 225Ac (actinium-225, alpha emitter) — "deliver radiation to target-expressing cells"

Known targets with existing ligands:

• PSMA (prostate cancer — Pluvicto, FDA approved)
• FAP (fibroblast activation protein — many solid tumors)
• SSTR (neuroendocrine tumors — Lutathera, FDA approved)
• B7H3, EphA2 (experimental)

Key insight from reference case: FAP was discovered via scRNA-seq, not standard tests. Target must be validated by PET imaging before therapy.

Access: Available in Germany and select US centers. May require international travel or expanded access.

---

Immune Modulators

---

Targeted Therapy

---

Combination Rationale

The goal is converting a "cold" tumor (immune-excluded) to a "hot" tumor (immune-infiltrated). Each modality attacks a different barrier:

BARRIER: Tumor is invisible to immune system
  → SOLUTION: Neoantigen vaccines train recognition

BARRIER: Immune cells are inhibited by tumor
  → SOLUTION: Checkpoint inhibitors remove brakes

BARRIER: Tumor microenvironment is immunosuppressive
  → SOLUTION: Oncolytic virus (TGF-beta trap), immune modulators

BARRIER: Not enough immune cells activated
  → SOLUTION: IL-15, GM-CSF amplify immune response

BARRIER: Tumor has a protective stromal shield
  → SOLUTION: FAP-targeted radioligand destroys stroma

BARRIER: Tumor escapes via clonal evolution
  → SOLUTION: Multi-antigen vaccines, updated per sequencing round

Evidence from reference case: This multi-modal parallel approach shifted immune infiltration from 19% → 89% T cells — the transformation from cold to hot.

---

Contraindications and Monitoring

Checkpoint inhibitor irAEs:

• Thyroiditis (10-20%): Monitor TSH. Treat with methimazole or levothyroxine.
• Colitis (5-15%): Monitor for diarrhea. May require steroids.
• Hepatitis (5-10%): Monitor LFTs.
• Pneumonitis (1-5%): Monitor for dyspnea.
• Positive prognostic sign: irAEs correlate with better tumor response.

Steroid risks:

• Repeated dexamethasone → avascular necrosis risk (hip, jaw)
• Minimize steroid exposure when possible

Radioligand risks:

• Bone marrow suppression (monitor CBC)
• Kidney toxicity (225Ac: monitor renal function)
• Salivary gland damage (for some targets)

Surgical site risks (post-spinal surgery):

• Seroma/infection: culture-guided antibiotics, may require hardware removal
• CSF leak: epidural blood patch