karkive learns from 988 real human cancer cell lines in the GDSC drug-screening project, then reads a tumor's genomic profile and ranks 369 drugs by predicted sensitivity, with the biomarkers behind every recommendation.
Two patients with the “same” cancer can need completely different drugs. karkive exists to make that difference visible, turning a tumor's genomics into a ranked, explained treatment shortlist in seconds.
karkive is under active development and subject to many changes. The model, the underlying data, the drug panel, and the design are all moving targets, expect it to look and behave differently as it grows. The drug panel now spans the full GDSC screen — 369 distinct compounds across 24 target pathways, now spanning two independent GDSC screens (GDSC1 + GDSC2). Feedback and rough edges are part of the process.
I keep coming back to one idea: two people can get the exact same diagnosis and still have cancers running on completely different wiring. So it never sat right with me that treatment so often follows the average patient instead of the tumor in front of you. I wanted to build toward the version where it follows the biology.
And it's worth being upfront that we're not there yet. A lot of treatment still starts from a standard playbook. The targeted matching we do have leans on a handful of biomarkers, the testing isn't evenly available, and plenty of care still comes down to trial and error. The promise is real; the day-to-day is still slow, patchy, and imprecise.
karkive is my small attempt at chipping away at that gap. The idea is simple: take a tumor's genomic profile, learn from real drug-response data, and suggest which drugs it's most likely to respond to, ranked, explained, and upfront about what it doesn't know. It's early and imperfect, but I'd rather build it in the open and keep improving than wait around for perfect. One sample in, the right treatment out.
Three reasons the molecular profile, not the organ, should steer treatment.
A breast tumor driven by HER2 behaves nothing like one driven by BRCA loss. The driver mutation often decides what works far more than where the tumor sits.
Every cycle on an ineffective regimen is time a tumor keeps growing. Matching therapy to biology up front is the entire promise of precision oncology.
Every karkive recommendation ships with the biomarkers that drove it and an honest uncertainty band. No mystery scores; you can see the reasoning.
Driver mutations, tissue and MSI describe the tumor.
The model scores every therapy's sensitivity.
Therapies ordered, each with an uncertainty band.
The biomarkers behind the top match, quantified.
When someone is diagnosed and needs chemotherapy, the drug they receive is usually chosen from a standard protocol for that cancer type — not from the specific biology of their tumor.
Treatment tends to follow the average patient for a given cancer type and stage. But two people with the “same” diagnosis can have tumors running on completely different genetic wiring — and respond to completely different drugs.
Physically screening a tumor sample against dozens of chemotherapies means growing cells and running assay after assay — costly lab work that can take weeks. All while the cancer keeps growing.
Even after all that testing, many tumors never show a strong response to anything on the shelf. Care then falls back to trial and error — cycle after cycle of “try it and see,” losing precious time.
karkive aims to skip the guesswork. From a tumor's genomic profile alone, it predicts how that tumor is likely to respond to 369 drugs computationally, in seconds — before any physical screening — and returns a ranked, explained shortlist of the therapies most worth trying first.
See how it works →A transparent pipeline built on a well-worn idea in pharmacogenomics: predict how a tumor responds to every drug, then rank.
A sample is described by the features GDSC records for every cell line: driver-gene mutations (TP53, KRAS, EGFR, BRAF, ALK, ERBB2, BRCA1/2, PIK3CA, PTEN, NRAS, APC, CDKN2A), ERBB2/HER2 amplification, MSI status, and the tissue of origin. Anything you leave out is treated as not-present and flagged.
karkive frames this as drug-response prediction and blends two gradient-boosted models: a per-drug model trained on each drug's own screened lines, and a multi-task model trained across all 369 drugs at once, so a drug tested on few lines borrows the pattern from similar drugs. Averaging the two beat either alone on a held-out-by-cell-line test set — the per-drug model calibrates each drug, the multi-task model sharpens the cross-drug ranking.
Drugs are ranked by predicted sensitivity, shown as a percentile ("more sensitive than X% of cell lines"), with an uncertainty band from the model's held-out residuals. A decision margin tells you how clearly the top pick beats the runner-up. A wide margin is a confident call; a narrow one says “these are close.”
For the recommended therapy, karkive reports a signed, quantified attribution: how many percentage points each biomarker added or subtracted, split into supporting factors and resistance cautions. The recommendation is a glass box, not a black one.
Drop a single tumor as CSV, TSV, or JSON. Wide, key-value, or one-record formats all parse.
No file? Pick a tissue and toggle the mutations and markers; the form is generated from the model's own schema.
Upload one tumor per row and get a ranked recommendation table for the whole group, exportable to CSV.
The therapy panel, the biomarker rules that ground the model, and how well it actually performs.
All 369 distinct compounds from two GDSC screens (GDSC1 + GDSC2), spanning 24 target pathways — from broad cytotoxics to precisely targeted kinase inhibitors.
We model 248 distinct compounds from GDSC1 (release 17, replicate screens merged) plus 121 additional compounds from the newer GDSC2 assay — 369 in total, each carrying an authoritative name and target pathway. The two screens share cell-line IDs, so GDSC2 slots straight onto the same genomic profiles.
These associations are measured directly in the GDSC cell lines (mutant vs wild-type IC50), not hand-written.
Honest numbers, held out on 20% of cell lines — split by cell line so no line appears in both train and test. Predicting drug response from a small biomarker panel is genuinely hard.
Honesty check: when we hold out whole tissue types (a much harder test), rank correlation collapses to ~0.07. The model leans heavily on tissue context and doesn't yet extrapolate to tissues it has never seen — a real limitation worth stating plainly.
Upload a sample, fill in a profile, or score a whole cohort, then read the ranked match and the reasoning.
Click an example to match it instantly:
One click runs the match, no download needed (right-click to download).
Upload a sample to see the ranked match.
| # | Cancer | Top-ranked therapy | Sensitivity pctile | Separation | Margin | Runner-up |
|---|
karkive is a one-person project. Here's who's building it — and why it matters to me.
Cancer reaching into my own family is what set this in motion — but karkive isn't really about me. It's about an idea worth chasing.
Picture a world where, the moment a tumor is sequenced, its own biology — not the average patient — decides the treatment. Where the therapy most likely to work is identified in seconds from data, instead of discovered through months of trial and error. Where "let's try this and see" becomes "the biology points here." That is the standard of care cancer patients deserve, and the direction medicine is genuinely moving.
That world isn't here yet, but it's buildable. karkive is one small step toward it: turning a tumor's genomic profile into a ranked, explained shortlist of the therapies worth trying first. It's early and imperfect — but the gap between how cancer is treated today and how precisely it could be is exactly what this is for.
This started with my own family. Every improvement to the model is, in a small way, for them and for anyone else still waiting on an answer that fits their cancer.
I'm a first-year student, so I'm learning the biology, the statistics, and the engineering as I go. Building in the open keeps me honest about what I do and don't understand yet.
karkive is grounded in real GDSC drug-response data and recovers real pharmacogenomics. It's a demo today, but the ambition is to keep pushing it toward something genuinely useful.