Try VERDICT

Our Science

Virtual human models, measured against real experiments and real trials.

A composition of tissue microscopy, expression heatmaps, and dot grids

Benchmarks

By therapeutic area

Held-out performance across 16 therapeutic areas, measured by AUROC. AUROC is how often VERDICT ranks a trial that succeeded above one that failed: 0.5 is a coin flip and 1.0 is always right.

VERDICT AUROC and trial counts by therapeutic area, November 2026 evaluation
Therapeutic areaLater trialsUnseen drugs
Metabolic
0.86n = 184
0.93n = 149
Endocrinology
0.76n = 112
0.89*n = 42
Pulmonology
0.74n = 287
0.87n = 89
Urogenital
0.62n = 173
0.84n = 72
Musculoskeletal
0.72n = 87
0.83*n = 51
Oncology
0.73n = 409
0.81*n = 47
Dermatology
0.76n = 213
0.80n = 103
Cardiology
0.74n = 146
0.78*n = 56
Immunology
0.74n = 242
0.77n = 102
Hematology
0.76n = 118
0.77*n = 21
Infectious disease
0.74n = 239
0.76n = 75
Neurology
0.66n = 205
0.75n = 156
Congenital
0.79n = 131
0.73n = 72
Gastroenterology
0.77n = 169
0.69*n = 57
Ophthalmology
0.70*n = 55
0.66*n = 28
Psychiatry
0.65n = 162
0.59n = 117

Bars run from 0.5 (chance) to 1.0 (perfect). Areas overlap. * Fewer than 60 trials; estimates are less certain.

Comparison with previous methods

VERDICT outperforms SPOT and HINT across all five held-out tests, measured by AUROC. AUROC is how often a method ranks a trial that succeeded above one that failed: 0.5 is a coin flip and 1.0 is always right.

Most recent trials1,717 held-out trials
0.50.60.70.8VERDICT0.75SPOT0.60HINT0.59
Trials of new drugs806 held-out trials
0.50.60.70.8VERDICT0.81SPOT0.67HINT0.64
New mechanisms636 held-out trials
0.50.60.70.8VERDICT0.78SPOT0.63HINT0.61
New diseases1,573 held-out trials
0.50.60.70.8VERDICT0.75SPOT0.52HINT0.57
Random trials1,592 held-out trials
0.50.60.70.8VERDICT0.78SPOT0.62HINT0.65

Same-split comparison. Bars start at 0.5 (chance). Whiskers show reported intervals.

Biology (GEM)

89%

Of lab-replicate correlation on the strongest reference-conditioned holdouts.

0.83 gene-wise correlation, using an observed control sample, versus the pooled 0.93 lab-replicate benchmark.

Reference-conditioned gene-expression benchmarks
Lab replicates0.93Known contexts0.83New genetic perturbations0.83New drugs0.77New tissues and cells0.75
Read the GEM-1 preprint

Publications

Synthesize open-source software

Trained on human biology.

Our models learn directly from one of the world’s largest harmonized human RNA-seq collections, connecting biological measurements with curated experimental context and clinical trial outcomes.

bulk RNA-seq samples
800,000
single cells
400M
unique perturbation contexts
22,300
human tissues
50+
therapeutic areas
40+
clinical trials with labelled outcomes
12,418

Partner with us

Synthesize Bio enables teams to explore broader research spaces and make faster, more confident drug development decisions. We are actively engaging with biopharma partners to deploy these capabilities at scale.