ChatGPT and Cancer: How a Tech Founder Rewrote His Treatment Plan

TL;DR: GitLab co-founder Sid Sijbrandij turned his terminal osteosarcoma into an engineering sprint—25 TB of multi-omic data, AI-driven literature agents, custom mRNA vaccines and a patient-built pharma pipeline—to stay alive and open-source the playbook. ## From ER Oddity to 6-cm Spinal Tumor In 2022, a persistent chest pain sent Sid Sijbrandij to the ER at 4 a.m. A precautionary CT found no cardiac trouble but revealed a 6-centimeter tumor wrapped around his thoracic spine. Within days he underwent spinal fusion, radiation, and high-dose chemotherapy—"medieval-level brutal," he recalls—requiring four transfusions to survive the first cycle. The diagnosis: conventional osteosarcoma, an ultra-rare bone sarcoma seen in fewer than 1,000 U.S. adults per year. Standard therapy bought a brief remission, but two years later scans lit up again: local relapse, no approved options left. ## Founder Mode Meets Oncology Sid stepped down from day-to-day duties at GitLab and re-allocated the same resources he once spent scaling DevOps toward a single user story—himself. Operating principles: - **Hoard data first, ask questions later.** Every emerging diagnostic modality—bulk RNA-seq, single-cell multi-ome, long-read genome, methylome, spatial transcriptomics, targeted radiomics—was run immediately. - **Parallelize experiments.** Patients die when sequential trials take months; multiple hypotheses run in overlapping tracks. - **Build the drug if it doesn’t exist.** FDA’s single-patient IND (compassionate use) route clears 99.7 % of applications, effectively allowing n-of-1 therapeutics. The cache now totals 25 TB and is publicly downloadable at [osteiosark.com](https://osteiosark.com) to accelerate community research. ## Mining 25 TB with ChatGPT Agents Jacob Stern, a geneticist formerly at 10x Genomics, orchestrates the data pipeline. He illustrates how modern GPTs collapse week-long literature dives into minutes: 1. Raw CSV of bulk tumor RNA counts uploaded to GPT-4—no prompt engineering. The model immediately flags B7-H3 (CD276) over-expression and notes an immune-excluded micro-environment, a hypothesis later corroborated by single-cell data. 2. A home-built multi-agent swarm turns natural-language questions ("Does Sid show CHIP risk post-chemo?") into automated PubMed search, marker selection, Jupyter-notebook generation, 600 k blood-cell re-analysis, and an interactive Plotly report—30 min, ~$20 in API spend. Pathologists verified the negative CHIP call the same afternoon. Sid credits these AI shortcuts for letting a small, non-clinical team "punch above their weight" in weekly tumor-board-style reviews with world-class oncologists. ## Targeting FAP: From Computer to Isolation Ward Single-cell maps showed cancer-associated fibroblasts uniquely over-expressing FAP (fibroblast-activation protein). A literature scan surfaced a German phase-I program coupling a FAP-binding ligand to Lutetium-177 beta-emitters. Sid flew to Cologne, received two cycles, and lived in a lead-lined room while radioactive. Result: 60 % necrosis, 20 % shrinkage, and surgical cleavage from the dura—enabling a margin-negative resection that had been deemed impossible. Post-therapy TCR sequencing revealed clonal T-cell expansion, suggesting FAP depletion converted an immune-"cold" tumor to "hot." ## Rescuing an Orphan MDM2 Inhibitor Osteosarcoma usually carries p53 pathway defects; Sid’s tumor instead showed massive MDM2 amplification (top 3 % of pan-cancer samples). Pharma giant Roche had shelved its MDM2 inhibitor idasanutlin after mixed trials in liposarcoma. With the company’s consent, Sid financed freezer maintenance and stability assays, aiming for a formal single-patient IND resupply. If successful, the program could salvage therapy for the subset of MDM2-amplified osteosarcoma patients historically excluded from p53-reactivating strategies. ## DIY Cancer mRNA Vaccine Parallel effort: a personalized mRNA neo-antigen vaccine. Workflow: 1. **Variant calling** from WGS/WES pinpoints 134 tumor-specific mutations. 2. **Epitope prediction** (NetMHCpan-4.1) ranks mutant peptides by HLA-A*02:01 binding affinity. 3. **ChatGPT + Python** automate in-silico cloning, inserting top 30 epitopes into the 5′-UTR optimized backbone used in Moderna’s Covid shots. 4. **Contract manufacturing** (GMP mRNA synthesis & LNP encapsulation) scheduled for 6-week turnaround. 5. **Combination plan:** vaccine priming → low-dose anti-PD-1 → FAP-radiation boost to keep tumor antigen supply high. FDA pre-IND feedback is pending; the team expects first-in-human dosing within four months. ## Click Chemistry for a Bespoke Small Molecule Jose, a YC alumnus and click-chemistry entrepreneur Sid once backed, is building a tetrazine-armed alkylator that "snaps" onto a trans-cyclooctene linker pre-conjugated to a tumor-homing peptide. The two-step protocol confines chemotherapy to the lesion, sparing marrow. Murine PK/PD data look clean; dog toxicity studies start Q3 2024 with Sid penciled as patient #1 if survival endpoints are met. ## Lessons for Patients & Builders 1. **Data is the new tissue.** Freeze everything: plasma, viable cells, FFPE blocks, cfDNA. Cheap storage beats repeat biopsies. 2. **AI shortens iteration cycles.** Use GPTs to summarize papers, script analyses, draft IRB protocols, and rehearse tough conversations with payers. 3. **Regulatory paths exist.** Single-patient INDs, Right-to-Try, and overseas compassionate-use programs convert desperation into legal drug development. 4. **Open-source the journey.** Publishing protocols, raw files, and code invites scrutiny and recruits talent that no cancer center can hire full-time. ## Toward a Scalable Model Sid and Jacob emphasize they are not advocating cowboy medicine; every experiment is reviewed by affiliated oncologists and cleared through institutional boards. Their broader mission is to template a "minimum-viable precision oncology" stack—cloud lab APIs, off-the-shelf sequencing, AI triage—that any hospital can replicate for rare tumors. GitLab’s dev-ops mindset (merge requests, issue tracking, CI/CD) now governs Sid’s treatment ledger: each therapy is a pull request, imaging and labs are unit tests, progression-free survival is the release milestone. ## Closing the Loop Scott McKinlay, himself battling refractory lymphoma, closed the session by noting that AGI’s greatest near-term impact may be "raising the medical floor," but stories like Sid’s show it can also "raise the ceiling" for those out of options. The same models that write code or draft marketing copy are now knitting together multi-omic threads into real-time treatment plans—one conversation, one cell, one patient at a time. Source: [OpenAI Forum – ChatGPT and Cancer: How a Tech Founder Rewrote His Treatment Plan](https://www.youtube.com/watch?v=OAlHiQLsYQM)