2026-09-19
MSK-Tox uses a RAG-LLM on 2.77M notes from 55,406 patients. Event-specific forests beat pooled models; DRB1*15 homozygotes on CTLA-4 hit 35% one-year adrenal insufficiency.
Precision oncology has a mature toolkit for who will benefit from a drug and almost nothing equivalent for who will be harmed. Trial toxicity data come from selected populations with short follow-up. In the clinic, onset, attribution, and grade live in free-text notes that do not scale to manual review.
Memorial Sloan Kettering turned that gap into MSK-Tox: a temporally resolved resource of six systemic-therapy toxicities extracted from 2.77 million oncology notes across 55,406 patients, then linked to labs, tumor sequencing, and germline genotypes. The claim is that safety can be studied with the same clinical and genomic machinery already used for efficacy.
Tox-Annotator is a retrieval-augmented LLM. Notes are chunked, embedded with BAAI/bge-large-en, retrieved by toxicity-specific queries, batched into 90-day windows, and scored by Llama 4 Maverick. Retrieval cuts the text volume enough to run at institution scale. The six endpoints are pneumonitis, adrenal insufficiency, liver toxicity, colitis, hyperthyroidism, and hypothyroidism.
Validation used four checks: CTCAE labels from 6,279 trial patients; manual adjudication on 300 patients; grade agreement; and time-to-onset. The gold standard itself is noisy: of 197 disagreements in 1,800 comparisons, 44.2% were gold-standard errors. After adjudication, AUCs ran from 1.00 (adrenal insufficiency) and 0.97 (pneumonitis) down to 0.77 (hyperthyroidism) and 0.81 (liver). F1 ranged from 0.40 to 0.90. Exact grade match was 63.1%, 91.7% within one grade. Median onset error was −10 days.
Risk models are random survival forests using only pretreatment variables: age, sex, cancer type, prior lines, and planned regimen. Time-to-event GWAS was replicated in 6,112 Dana-Farber ICI patients and checked in FinnGen and UK Biobank.
Immune-related toxicity is not a uniform good-prognosis signal. Among 6,407 first-line ICI patients, grade ≥3 pneumonitis had OS HR 2.79 and grade ≥3 liver toxicity HR 2.64; most of the six events had no significant survival association. Line 4+ versus first line raised adrenal insufficiency HR to 6.96 and hypothyroidism to 3.52.
A pooled any-toxicity model reached only c-index 0.615 for grade 1+ events on ICI. Event-specific models moved that to 0.640 for liver toxicity and 0.758 for adrenal insufficiency. High versus low RSF tertiles split one-year liver toxicity 24.6% vs 7.1% and hypothyroidism 29.1% vs 8.0%.
No SNP reached genome-wide significance for a pan-toxicity endpoint. Split by organ:
| Locus | Event | Scope |
| FOXE1 rs10983761 | hypothyroidism | ICI and non-ICI; MSK HR 1.25, replicated at DFCI |
| FOXP2-region rs12705907 | pneumonitis | mainly ICI; HR 3.67, not replicated |
| HLA-DRB115 | adrenal insufficiency | ICI only; 12-month incidence 35.0% in CTLA-4 homozygotes vs 6.7% in non-carriers |
The DRB1 risk residues sit in the peptide-binding groove at positions 11, 13, and 71, the same axis implicated in multiple sclerosis and rheumatoid arthritis. 31.1% of genotyped patients carry at least one copy. Inside the clinical high-risk tertile, carriers had 12.0% one-year incidence versus 6.5% in non-carriers.
For medical NLP this is a scale test that ties a RAG-LLM to clinical endpoints and germline data, not another note-classification AUC. For oncology and pharmacogenomics the scientific claim is sharper: inherited toxicity risk is organ- and drug-conditional, so pooled “any irAE” GWAS is the wrong endpoint. HLA typing is already routine in transplant; the dosage curve is a concrete pretreatment stratification signal.
The survival forests are incremental. Labs and tumor genomics added little. The new pieces are the phenotype scale and the DRB115 gene-dosage effect.
Only six toxicities were phenotyped; dermatologic and other endocrine events are missing, so results do not generalize to all irAEs. The trial gold standard under-reports, and the 5,279-patient holdout was not re-adjudicated. The FOXP2 pneumonitis locus failed at Dana-Farber. GWAS is European-ancestry heavy. This is a single-center retrospective cohort; the 35% homozygote figure comes from 33 patients, with a 14.2–56.8% interval. LLM F1 bottoms out at 0.40, so both false negatives and false positives flow into downstream association tests.