Five candidate T-cell brakes. One genome-scale discovery engine.
Brakepoint is a genome-scale discovery engine for the next generation of cancer-immunotherapy drug targets. The best immunotherapies cut the brakes off a patient's T cells; Brakepoint reads the public Marson lab (Gladstone) and Pritchard lab (Stanford) screen in one pass — 2.6 million cells and 12,449 genes switched off — and tells a real drug target apart from the machinery a cell needs to survive. It rediscovered CBLB from raw data with zero prior hints — its inhibitors are already in early-phase trials — then delivered a five-candidate pipeline ready for functional testing.
Five candidate targets. Seven lines of evidence each.
Five candidate T-cell targets rise to the top, each backed by seven independent lines of evidence. A "brake" is a gene whose shutoff pushes a T cell toward a stronger fighting state — the same move checkpoint drugs make. This screen reads transcription, so functional assays are the next step; how ready each target is to drug ranges widely — CBLB inhibitors are already in trials, UBASH3A is undrugged, and SMAD3 is supported at the pathway level. Each candidate is scored across seven independent lines of evidence: how hard its shutoff hits the cell, which way it pushes the cell, consistency across donors, viability, druggability, immune genetics, and clinical precedent.
Effect, direction, donor consistency and viability are computed from the genome-scale CD4⁺ leaderboard; druggability, immune-genetics and clinical precedent are curated from Open Targets, ChEMBL and ClinicalTrials.gov. Reproduce with make figure.
CBLB rose from raw data to the top.
Brakepoint found CBLB with zero prior hints, straight from the raw screen. It also carries the strongest external evidence of any candidate — the screen, human genetics, and an active clinical program all point the same way. That convergence makes CBLB the lead.
Tell a real drug target from survival machinery.
One experiment runs 12,449 causal tests at once — but read it the usual way and statistical significance can't separate a candidate brake from a gene the cell can't live without. Brakepoint reads the same data in three moves.
Measure the hit, not the p-value
Rank every knockdown by how hard the gene's shutoff hits the cell — the size of the shift away from control, compared fairly across genes. The standard test lights up for 97.5% of the 11,438 tested knockdowns, so significance alone stops telling you anything.
Give every hit a direction
Ask which way each shutoff pushes the cell — toward a stronger fighter or a weaker one. All 2.6 million cells are scored in 38.9 seconds, and the map flips: the biggest raw effects reveal themselves as machinery the cell needs to survive, while the candidate brakes rise into view.
Demand seven lines of evidence
The shortlist is then ranked across seven independent lines of evidence — effect, direction, donor consistency, viability, druggability, immune genetics, clinical precedent. The result is a test-ready order built on convergence, not a one-score guess.
Step 2 in one figure: effect size alone (x) would nominate the T cell's own machinery — 14 of the 15 largest effects are essential machinery. Adding direction (y) reclassifies those top effects as required for survival and lifts the candidate brakes into the upper region.
Step 1, at scale: nearly every tested knockdown clears significance — 9,802 of the 11,438 pile up at the floor of the test. Only effect size — how hard each shutoff hits the cell — spreads the biology apart, which is why we rank by it.
The computational work behind it
Search all 11,438 tested knockdowns.
This isn't a picture of the map — it's the real leaderboard, rendered straight from the data. Every tested knockdown is placed by how hard its shutoff hits the cell (x) and which way it pushes the cell (y). Search a gene — best on a touchscreen — or hover to read its numbers: the survival machinery sits low, the candidate brakes sit high.
x = how hard each gene's shutoff hits the cell (energy distance) · y = which way it pushes the cell over 8 hours (toward a stronger fighter, up; a weaker one, down). Loaded from data/causal_map_points.json — 11,438 tested knockdowns, exactly what the ranking sees. Prefer a static view? The map above plots the same axes.
Find the brake, not just the change.
Every column below is a genuine tool, and Brakepoint uses several of them as evidence layers. But for the specific job — finding a druggable brake whose shutoff makes a T cell a stronger fighter — each leaves a blind spot. Brakepoint fills it by reading impact, direction, and converging evidence together.
Add the signed direction axis and every one of those machinery genes flags negative and drops out. The real candidates surface: 2,016 knockdowns push the cell toward a stronger fighter, and 1,286 of them do it consistently in both donors. Traditional pipelines can't even see this search space.
| Traditional approach | What it gives | Blind spot for finding brakes | Brakepoint's edge |
|---|---|---|---|
| Differential expression | correlated markers of a cell state | correlation ≠ causation — can't tell a driver from a passenger | measures what a knockdown does (CRISPRi vs control), genome-wide |
| Human genetics / GWAS | disease-linked loci, in humans | rarely tied to a cell type or a direction; mostly non-coding | shows which way a gene pushes the actual T cell — genetics still folded in as a layer |
| Literature / network mining | re-weights known biology | recapitulates the known; hub-biased; buries understudied genes | ranks by measured effect, independent of publication volume |
| Bulk CRISPR (viability) | one fitness phenotype | can't separate "kills the cell" from "makes a better effector" | a full transcriptional phenotype per cell; viability is only a gate |
| Significance- or DE-first Perturb-seq analysis | the closest competitor — same data class | at this scale the standard test lights up for almost everything (97.5% of tested knockdowns), and those rankings never say which way a gene pushes the cell | ranks by how hard each shutoff hits the cell, then adds the direction that splits a brake from the machinery |
In context: none of these is a knockout blow — Brakepoint leans on genetics and clinical precedent, and its direction score is an 8-hour transcriptional readout that still needs functional validation. The core statistic (energy distance, the scPerturb standard) isn't new; the win is the combination — a genome-scale ranking by how hard each shutoff hits the cell, a direction that says which way the cell is pushed, and an immuno-oncology druggability shortlist. The genome-wide map uses no publication-volume features, so understudied, genetics-led picks like UBASH3A can still surface.
Two donors establish the map. Four sharpen it.
Across the two donors analyzed so far, the machinery axis is unanimous — every essential-machinery gene scores negative in both. On the brake side, CD5 and DGKA hold up in both donors, while CBLB, SMAD3 and UBASH3A are driven by one donor (n=2). As a group, the 29 curated known brakes don't yet score higher on the direction axis than background (one-sided Mann–Whitney, p = 0.70) — so the positive region is not yet broadly brake-enriched. What the analysis does establish is clean: the machinery is flagged negative in both donors, and the known clinical brake CBLB is recovered in the positive region alongside the donor-consistent CD5 and DGKA. That p = 0.70 measures the power of a two-donor set to enrich a whole class — not the five individual calls. The shortlist is a prioritized set to test at the full four-donor scale, where the map sharpens.
From 2.6 million cells to five candidate targets.
A short walkthrough of the question, the map, and the target shortlist — with Claude Science provenance behind every number, from raw input to final list.
Every candidate traces back to the exact run.
One person built Brakepoint end to end in a week with Claude Science, running the genome-scale analysis on an NVIDIA DGX Spark. Each result is a versioned artifact a reviewer agent checks against what actually ran, and an adversarial self-critique caught and fixed a real bug in the effect-size math before it reached a figure. From a clean clone, make smoke re-runs the core checks and make figure regenerates every figure.
# dependency-free core tests
git clone https://github.com/duanchengchen-oss/brakepoint.git
cd brakepoint/pipeline && make smoke
# regenerate the target figures
make figure
# is the brake quadrant enriched? (the p=0.70 null)
python brake_enrichment.py