AUC 1 = perfect, 0.5 = chanceCorrelation r 1 = perfect, 0 = nonep2singleEvery score is on cells, patients or experiments kept out of training.Click each box to see more detail on the model's predictions and the sequence features it uses.
Immunotherapy · skin cancer
Exhausted vs active killer T cells in patients on immunotherapy
Patients the model never saw · 2,326 test cells
sequence features (gene) · higher in
what it is
known in the literature
known variants in its sequences
TIGIT papersequence features higher in exhausted
Inhibitory checkpoint receptor on T and NK cells
97 SNPs 8 ClinVar 1 of 46 span a splice junction
CXCL13sequence features higher in exhausted
Chemokine that attracts B cells
82 SNPs 4 ClinVar 8 of 16 span a splice junction
IFNG / IFNG-AS1 papersequence features higher in active
Interferon gamma, the main effector cytokine of activated T and NK cells / Antisense RNA next to IFNG
107 SNPs 4 ClinVar 6 of 31 span a splice junction
known variants: dbSNP and ClinVar variant positions inside the model's selected sequences for that gene, and how many of those sequences span a splice junction (two exons)
paper also named in the original paper
GSE123813 · Yost et al. 2019 · updated 2026-10-28 12:32Click to view sequence features used by the modelfull results →
Lung cancer · PD-L1 protein
PD-L1 levels in lung tumour immune cells of a new patient
All 14,903 test cells are from patient 706, never seen in training · trained on patients 626, 630, 695
The published methods were tested on random cells from the same patients, an easier test.
RAB11FIP1sequence features higher in PD-L1 low cells
Recycling-endosome protein
16 SNPs 3 ClinVar 1 sequence placed
genes from the high vs low PD-L1 model; known variants: dbSNP and ClinVar variant positions inside the model's selected sequences for that gene
SRP251372 · Leader et al. 2021 · updated 2026-10-28 12:59Click to view sequence features used by the model
Bone marrow · surface protein
Predicting protein-defined cell types in bone marrow from RNA
Within one study · held-out cells from the same 2 donors · 4,339 test cells
sequence features (gene) · higher in
what it is
known in the literature
known variants in its sequences
CD74sequence features higher in B cells
MHC class II invariant chain
93 SNPs 0 ClinVar 0 of 19 span a splice junction
IGHMsequence features higher in naive B cells
IgM antibody heavy chain
18 SNPs 0 ClinVar 0 of 2 span a splice junction
S100A8sequence features higher in CD14 monocytes
Calprotectin subunit (with S100A9)
62 SNPs 4 ClinVar 2 of 10 span a splice junction
LYZsequence features higher in CD14 monocytes
Lysozyme, an antibacterial enzyme
47 SNPs 2 ClinVar 0 of 7 span a splice junction
known variants: dbSNP and ClinVar variant positions inside the model's selected sequences for that gene, and how many of those sequences span a splice junction (two exons)
GSE128639 · Stuart et al. 2019 · updated 2026-10-28 12:29Click to view sequence features used by the modelfull results →
CRISPR screen · PD-L1 protein
PD-L1 protein level predicted from RNA, tested on a separate experiment
Same CRISPR screen run 3 times: trained on runs 1 and 2 (13,937 cells), tested on run 3 (5,516 cells)
PD-L1's own RNA alone: 0.40 · high vs low PD-L1 cells: AUC 0.93
sequence features (gene) · higher in
what it is
known in the literature
known variants in its sequences
CTSLsequence features higher in cells with more PD-L1
Cathepsin L, a lysosomal protease
198 SNPs 9 ClinVar 34 sequences placed
B2Msequence features higher in cells with more PD-L1
Beta-2-microglobulin, part of MHC class I
208 SNPs 5 ClinVar 37 sequences placed
CALHM6-AS1sequence features higher in cells with more PD-L1
Antisense RNA next to CALHM6
235 SNPs 18 ClinVar 37 sequences placed
FTLsequence features higher in cells with more PD-L1
genes from the arterial model; known variants: dbSNP and ClinVar variant positions inside the model's selected sequences for that gene, and how many sequences were placed on the genome
HCA arterial vasculature (Zhao 2025) · HCA uterus (Ulrich 2024) · updated 2026-10-28 12:25Click to view sequence features used by the model
Human pancreas · cell types
Pancreas cell types: at or above every published method
Xin 2016 human pancreas dataset: Fluidigm C1 with SMARTer full-length sequencing (not 10x), 4 cell types: alpha (glucagon), beta (insulin), delta (somatostatin) and PP (pancreatic polypeptide) cells · each published result on the same dataset and metric as its paper
45 of 46 published results on this dataset score lower than p2single and 1 ties; all 20 deep learning models score lower.
Macro F1 · p2single and the 17 published methods
p2single · deep learning other published methods · 1 = perfect. Same result on the other 3 metrics: all 9 weighted F1 and all 12 accuracies lower; 7 of 8 median F1 lower, 1 tie (XGBoost, 1.000).
sequence features (gene) · higher in
what it is
known in the literature
known variants in its sequences
GCGsequence features higher in alpha cells
Glucagon
239 SNPs 2 ClinVar 1 of 51 span a splice junction
INSsequence features higher in beta cells
Insulin
340 SNPs 101 ClinVar (33 pathogenic) 4 of 58 span a splice junction
PPYsequence features higher in PP cells
Pancreatic polypeptide
265 SNPs 8 ClinVar 16 of 161 span a splice junction
SSTsequence features higher in delta cells
Somatostatin
289 SNPs 7 ClinVar 11 of 144 span a splice junction
known variants: dbSNP and ClinVar variant positions inside the model's selected sequences for that gene, and how many of those sequences span a splice junction (two exons)
top genes of the p2single model on the Xin dataset (P2G v0.12): these 4 genes make up 82% of the model's weight
Xin 2016 and Segerstolpe 2016 · updated 2026-10-28 13:46Click to view sequence features used by the model
Brain · across species
Brain cell types predicted across species
Human → mouse: long-range vs local neurons · other rows: spiny vs non-spiny neurons
p2single · each small dot is one donor pair, the large dot is the middle value · AUC (0.5 = chance, 1 = perfect)
sequence features (gene) · higher in
what it is
known in the literature
known variants in its sequences
PDE10Asequence features higher in spiny neurons
Phosphodiesterase that breaks down cAMP and cGMP
55 SNPs 0 ClinVar top sequence of 6 human donor models
PHACTR1sequence features higher in spiny neurons
Regulator of protein phosphatase 1 and actin
85 SNPs 0 ClinVar top sequence of 6 human donor models
SSTsequence features higher in non-spiny
Somatostatin
13 SNPs 1 ClinVar top sequence of 2 human donor models
known variants: dbSNP and ClinVar positions inside the top sequence of each human donor model (macaque sequences are on the macaque genome and not counted)
Allen Brain Atlas · human, macaque, mouse · updated 2026-10-28 13:29Click to view sequence features used by the model
CRISPR screens · gene response
How cells respond when 2 genes are changed at once
Gene pairs the model never saw · same test as the benchmark