Atlas
Sign in
Models / single-cell / cpa / benchmarks

single-cell/cpa

Visibility: PUBLIC
Model cardFilesBenchmarks
Group by

Other Benchmarks drawn from no hub dataset you can see

Perturbation response (Pearson on Δ expression) single-cell/perturbation-response-delta-pearson

atlas-full-v1

Metric Value 95% CI n Rank Best in catalogue Source Overlap Board
Pearson Δ, all genes (trained) ↑ higher is better · supervised
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.373 [0.317, 0.425]43
Two-gene combination, both genes seen in training0.201 [0.137, 0.265]19
Single gene, unseen in training0.477 [0.408, 0.542]36
 0.396 [0.364, 0.429] 107 2 / 2 0.543 reproduced unknown Board
Pearson Δ, top-20 DE genes (trained) ↑ higher is better · supervised
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.604 [0.489, 0.703]43
Two-gene combination, both genes seen in training0.403 [0.221, 0.569]19
Single gene, unseen in training0.523 [0.374, 0.673]36
 0.559 [0.487, 0.629] 107 2 / 2 0.750 reproduced unknown Board
MSE, top-20 DE genes (trained) ↓ lower is better · log1p-normalised expression² · supervised
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.556 [0.434, 0.676]43
Two-gene combination, both genes seen in training0.523 [0.389, 0.681]19
Single gene, unseen in training0.270 [0.210, 0.331]36
 0.444 [0.389, 0.504] 107 2 / 2 0.214 reproduced unknown Board
Direction accuracy, top-20 DE genes (trained) ↑ higher is better · supervised
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.786 [0.745, 0.826]43
Two-gene combination, both genes seen in training0.726 [0.650, 0.800]19
Single gene, unseen in training0.758 [0.706, 0.808]36
 0.772 [0.744, 0.799] 107 2 / 2 0.849 reproduced unknown Board
MSE, all genes (trained) ↓ lower is better · log1p-normalised expression² · supervised · secondary
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.009 [0.006, 0.011]43
Two-gene combination, both genes seen in training0.010 [0.008, 0.012]19
Single gene, unseen in training0.003 [0.002, 0.004]36
 0.007 [0.006, 0.008] 107 — 0.006 reproduced unknown Board
Discrimination score (trained) ↑ higher is better · supervised · secondary
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.524 [0.442, 0.612]43
Two-gene combination, both genes seen in training0.521 [0.391, 0.645]19
Single gene, unseen in training0.514 [0.426, 0.607]36
 0.515 [0.464, 0.566] 107 — 0.764 reproduced unknown Board
DE-gene overlap, top 20, pseudobulk proxy (trained) ↑ higher is better · supervised · secondary
By subgroup: 4 levels
subgroupValue [95% CI]n
Two-gene combination, neither gene seen in training—fewer than 10 items9
Two-gene combination, one gene seen in training0.183 [0.144, 0.215]43
Two-gene combination, both genes seen in training0.153 [0.118, 0.187]19
Single gene, unseen in training0.247 [0.212, 0.282]36
 0.209 [0.188, 0.230] 107 — 0.238 reproduced unknown Board