Rosa Bio · P2G · liquid biopsy

P2G provides robust liquid biopsy prediction across all analyte types.

Last updated 2026-11-11 09:21 PST

P2G builds a predictive model directly from the raw sequencing reads of each sample: alignment performed after model predictions, no gene list. The analysis is performed with the same modeling, no difference between RNA or DNA, between or base-conversion or native nucleic acid sequences. No study-specific optimization. All predictions are reported on data never seen during training.

AUC 1 is perfect, 0.5 is a coin flipP2G within 0.02 of the paper, or higherpaper highernot tested in the paper; predicted by usfewer than 20 test samplesHover or tap any dot or bar to see its test samples.

Main point 1

P2G finds all 16 fusions in the reference mix and predicts the tumour DNA percent of 38 held-out libraries (R² 0.93)

Question: does P2G find tumour signals in reference samples where the answer is known before the test?

These are reference samples, not patients: lab-made cell-line RNA or DNA with known fusions and mutations, mixed into normal RNA or DNA at set amounts. They check that P2G finds a known answer before the patient studies below.

Spiked fusions

16 of 16

fusions in the Seraseq Mix v3 reference found with no model training; 9 of them already have reads at a 1% spike-in; no fusion read in any of 277 control samples

Tumour DNA percent

R² 0.93

on 38 held-out SEQC2 libraries, from 5 million of a median 85 million reads per library

Spiked mutations

34%

of the model's weight is on sequences with the spiked mutations, among them KRAS G12A, BRAF V600E and TP53 R306*

Reference samples · fusion RNA · Seraseq Fusion RNA Mix v3 (SRP293279)

All 16 fusions in the reference mix are found with no model training, 9 already at a 1% spike-in, and none in 277 control samples

18 libraries: the Seraseq Fusion RNA Mix v3 reference diluted into normal RNA, 3 replicate libraries at each of 6 spike-in levels, 101 to 227 million read pairs per library (median 159 million). 277 control samples with no spiked fusion: 90 plasma RNA (7.3 billion reads) and 187 tumour DNA libraries (42.2 billion reads). No stage: lab-made reference material, not patients.

No training. No model was trained, and the method was not told which fusions are in the mix: it searched every read for 41 known fusion junctions and reports the ones it finds. It was also not told which libraries were spiked, or at what level.

Two fusions marked on the chart. KIF5B-RET and SLC45A3-BRAF are the only fusions found in all 3 replicate libraries at the 5% spike-in, and both are already found at 1%.

each box: how many of the 3 replicate libraries at that level have a read across the fusion junction; darker is moreoutlined and labelled: KIF5B-RET and SLC45A3-BRAF, the only fusions in all 3 replicate libraries at 5%; both found at 1%

SRP293279 · Seraseq Fusion RNA Mix v3 reference materialupdated 2026-11-09 12:36 PST

Reference samples · tumour DNA · SEQC2 (SRP296025)

P2G predicts the tumour DNA percent of 38 held-out libraries (R² 0.93)

38 libraries kept out of training: 8 at 0%, 4 at 0.8%, 17 at 4% and 9 at 20% tumour DNA. Tumour DNA percent: how much of a library's DNA comes from the SEQC2 reference tumour DNA, which the study diluted into normal DNA. P2G used 5 million of a median 85 million reads per library (6%). No stage: lab-made reference material, not patients.

one test librarydashed line: predicted equals true

34% of the model's weight is on sequences with the spiked mutations, among them KRAS G12A, BRAF V600E and TP53 R306*

42,317 variants were spiked into the reference tumour DNA. The model was not told where they are. 57 of them sit in 154 of its 601 selected sequences (26% of the sequences, 34% of the weight).

spiked variant · weight	what the gene is	ClinVar	known in the literature	libraries with variant reads
BRCA1 chr17:43092418 T>C; allele frequency 25.6% in the tumour reference 2.72% of the model's weight; higher in libraries with more tumour DNA	DNA repair gene; inherited variants raise breast and ovarian cancer risk	not recorded	152 papers name BRCA1 with circulating tumour DNA in title or abstract; most cited: Clinical Implications of Plasma-Based Genotyping With the Delivery of Personalized Therapy in Metastatic Non-Small Cell Lung Cancer. (2019, PMID 30325992)	32 of 43 libraries at 20%; 2 of 37 at 0%
KRAS G12A paper chr12:25245350 C>G; allele frequency 1.8% in the tumour reference 2.53% of the model's weight; higher in libraries with more tumour DNA	RAS GTPase; one of the most often mutated driver genes in colorectal, lung and pancreatic cancer	Pathogenic/Likely pathogenic	798 papers name KRAS with circulating tumour DNA in title or abstract; most cited: Detection of circulating tumor DNA in early- and late-stage human malignancies. (2014, PMID 24553385)	24 of 44 libraries at 20%; 3 of 44 at 0%
ALK G1125= chr2:29222592 G>T; allele frequency 17.9% in the tumour reference 2.46% of the model's weight; higher in libraries with more tumour DNA	receptor tyrosine kinase; fusions and mutations drive some lung cancers and neuroblastoma	Benign	268 papers name ALK with circulating tumour DNA in title or abstract; most cited: Liquid Biopsy for Advanced Non-Small Cell Lung Cancer (NSCLC): A Statement Paper from the IASLC. (2018, PMID 29885479)	43 of 44 libraries at 20%; 2 of 44 at 0%
BRAF V600E paper chr7:140753336 A>T; allele frequency 8.3% in the tumour reference 2.45% of the model's weight; higher in libraries with more tumour DNA	kinase in the MAPK signalling pathway; V600E is a common driver mutation in melanoma and colorectal cancer	Conflicting classifications of pathogenicity	628 papers name BRAF with circulating tumour DNA in title or abstract; most cited: Fragment Length of Circulating Tumor DNA. (2016, PMID 27428049)	27 of 44 libraries at 20%; 1 of 44 at 0%
also named in the paper
TP53 R306* paper chr17:7673704 G>A; allele frequency 5.2% in the tumour reference 1.78% of the model's weight; higher in libraries with more tumour DNA	tumour suppressor; the most often mutated gene in human cancer	Pathogenic	783 papers name TP53 with circulating tumour DNA in title or abstract; most cited: Circulating Tumor DNA Genomics Correlate with Resistance to Abiraterone and Enzalutamide in Prostate Cancer. (2018, PMID 29367197)	29 of 41 libraries at 20%; 1 of 37 at 0%

paper named in the paper textliterature: Europe PMC papers naming the gene with circulating tumour DNA in the title or abstract

SRP296025 · study: Deveson et al. 2021, Nature Biotechnology · data description: Gong et al. 2022, Scientific Data · spiked variants: Jones et al. 2021, Genome Biologyupdated 2026-11-09 14:06 PST

Main point 2

P2G is within 0.02 of the paper, or higher, in 11 of 28 comparisons, using 4% to 23% of the paper's reads

Question: on the same question and the same sample type, how does P2G compare with the model each paper published?

Pregnancy · maternal plasma RNA · SRP271730

Adverse pregnancy outcomes are told apart from uncomplicated pregnancies at AUC 1.000; the paper reports 0.91

Cases: preeclampsia, gestational hypertension, gestational diabetes or chronic hypertension. Controls: pregnancies with no complication. 26 test samples (17 patients, 9 in the comparison group) from women kept out of training. P2G used 2 million of a median 32.7 million reads per sample (6.1%). Not a cancer, so there is no stage.

The paper (Del Vecchio 2021) reports AUC 0.91 for adverse outcomes from first-trimester plasma RNA, by leave-one-out cross-validation on 23 women (15 with an adverse outcome, 8 normal).

Example genes · The model uses 990 selected sequences; 286 of them (29%) are placed on the human genome, in 88 genes.

example genes · sequences higher in	what it is	known in the literature
SERINC2sequences higher in pregnancy complicationRank 7 of 88 genes · 1.2% of the model's weight · 3 of 3 sequences higher in pregnancy complication	Serine incorporator 2
FLNAsequences higher in no complicationRank 8 of 88 genes · 1.2% of the model's weight · 16 of 16 sequences higher in no complication	Filamin A

SRP271730 · Del Vecchio et al. 2021, Epigeneticsupdated 2026-11-09 12:36 PST

Cancer · plasma cell-free DNA, bisulfite · SRP610281

Gastric cancer is found from plasma DNA at AUC 0.94; the paper reports 0.87

196 test samples (82 gastric cancers, 114 people with no disease) kept out of training. P2G used 3 million of a median 13.3 million reads per sample (22.6%).
By cancer stage: P2G finds gastric cancer at every stage, including early stage I.
Stage I: AUC 0.92, from 23 stage I patients in the test
Stage II: AUC 0.94, from 12 stage II patients in the test
Stage III: AUC 0.95, from 40 stage III patients in the test
Stage unknown: AUC 0.89, from 7 patients whose stage was not recorded

The paper's multimodal model scores 0.87 (95% CI 0.80 to 0.93) on its own validation cohort (Long 2026).

Example genes · The model uses 480 selected sequences; 135 of them (28%) are placed on the human genome, in 59 genes.

example genes · sequences higher in	what it is	known in the literature
ARID1Bsequences higher in gastric cancerRank 1 of 59 genes · 3.2% of the model's weight · 9 of 9 sequences higher in gastric cancer	AT-rich interaction domain 1B
PPTC7sequences higher in gastric cancerRank 5 of 59 genes · 0.8% of the model's weight · 1 of 1 sequence higher in gastric cancer	Protein phosphatase targeting COQ7

SRP610281 · Long et al. 2026, BMC Cancerupdated 2026-11-09 12:36 PST

Cancer · plasma vesicle small RNA · SRP345647

Colorectal cancer is found from vesicle RNA at AUC 0.96; the paper's miR-320c marker scores 0.66

58 test samples (40 cases, 18 healthy controls) kept out of training. P2G used 2 million of a median 20.8 million reads per sample (9.6%).
By cancer stage: P2G finds colorectal cancer at every stage, including early stage I.
Stage I: AUC 1.000, from 2 stage I patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage II: AUC 0.98, from 7 stage II patients in the test
Stage III: AUC 0.94, from 13 stage III patients in the test
Stage IV: AUC 0.96, from 18 stage IV patients in the test

The paper's single marker, miR-320c, scores 0.658 in its sequencing validation cohort of 102 cases and 90 controls (Yang 2022).

Example genes · The model uses 955 selected sequences; 52 of them (5%) are placed on the human genome, in 23 genes.

example genes · sequences higher in	what it is	known in the literature
UNC45Bsequences higher in healthyRank 1 of 23 genes · 0.4% of the model's weight · 3 of 4 sequences higher in healthy	Unc-45 myosin chaperone B
RNU2-39Psequences higher in colorectal cancerRank 2 of 23 genes · 0.4% of the model's weight · 3 of 3 sequences higher in colorectal cancer	RNA, U2 small nuclear 39, pseudogene

SRP345647 · Yang et al. 2022, Cellular Oncologyupdated 2026-11-09 12:36 PST

Cancer · plasma cell-free DNA, methylation (MRE-seq) · SRP432064

Lung cancer is found from 2 of 46 million reads per sample at AUC 0.97; the paper reports 0.96

44 test samples (19 lung cancers, 25 healthy controls) kept out of training. P2G used 2 million of a median 46.1 million reads per sample (4.3%).
By cancer stage: P2G finds lung cancer at every stage, including early stage I.
Stage I: AUC 0.93, from 7 stage I patients in the test
Stage II: AUC 1.000, from 2 stage II patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage III: AUC 1.000, from 4 stage III patients in the test (too few to be sure; 95% CI 0.75 to 1.00)
Stage IV: AUC 1.000, from 6 stage IV patients in the test (95% CI 0.81 to 1.00)

The paper's deep learning model scores 0.956 for lung cancer, using all reads (Kwon 2023).

Example genes · The model uses 509 selected sequences; 491 of them (96%) are placed on the human genome, in 339 genes.

example genes · sequences higher in	what it is	known in the literature
DRAXINsequences higher in lung cancerRank 1 of 339 genes · 2.7% of the model's weight · 7 of 7 sequences higher in lung cancer	Dorsal inhibitory axon guidance protein
ANKRD6sequences higher in healthyRank 2 of 339 genes · 1.2% of the model's weight · 2 of 2 sequences higher in healthy	Ankyrin repeat domain 6

SRP432064 · Kwon et al. 2023, Scientific Reportsupdated 2026-11-09 12:36 PST

Weight: each sequence counts by how much the model relies on it; a gene adds up its sequences. Higher in: the group whose score the gene's sequences raise. Mitochondrial genes and genes with no gene name are left out here; the panel lists them.

Key results · every highlighted question

P2G is within 0.02 of the paper's AUC, or higher, on 9 of 14 highlighted questions with a published AUC

One dot per question. x: the paper's AUC; y: P2G held-out AUC. Questions with no published AUC sit in the left strip. Dot size: number of test samples. The next chart shows all 28 comparisons with a published AUC for the same question. Hover a dot for its tumour stage.

within 0.02 of the paper, or higherpaper highernot tested in the paper; predicted by usno published AUC (paper reports another measure, or no paper)fewer than 20 test samples

All 29 questions with their test samples and tumour stage

1 Glioblastoma vs people without symptoms · platelet RNA · Sol 2020 (SRP278818): AUC 0.97; 122 test samples (49 patients, 73 in the comparison group); paper's AUC 0.97. Glioblastoma is not given a stage; all patients had grade 4 tumours.
2 Colon cancer vs healthy · plasma cfDNA methylation · Kwon 2023 (SRP432064): AUC 0.96; 44 test samples (19 patients, 25 in the comparison group); paper's AUC 0.98.
By cancer stage: P2G finds colon cancer at every stage, including early stage I.
Stage I: AUC 0.93, from 5 stage I patients in the test
Stage II: AUC 0.96, from 6 stage II patients in the test
Stage III: AUC 0.99, from 6 stage III patients in the test
Stage IV: AUC 0.98, from 2 stage IV patients in the test (too few to be sure)
3 Lung cancer vs healthy · plasma cfDNA methylation · Kwon 2023 (SRP432064): AUC 0.97; 44 test samples (19 patients, 25 in the comparison group); paper's AUC 0.96.
By cancer stage: P2G finds lung cancer at every stage, including early stage I.
Stage I: AUC 0.93, from 7 stage I patients in the test
Stage II: AUC 1.000, from 2 stage II patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage III: AUC 1.000, from 4 stage III patients in the test (too few to be sure; 95% CI 0.75 to 1.00)
Stage IV: AUC 1.000, from 6 stage IV patients in the test (95% CI 0.81 to 1.00)
4 Breast cancer vs healthy · plasma cfDNA methylation · Pham 2023 (SRP420167): AUC 0.95; 104 test samples (48 patients, 56 in the comparison group); paper's AUC 0.91.
By cancer stage: P2G finds breast cancer at every stage, including early stage I.
Stage I: AUC 0.98, from 9 stage I patients in the test
Stage II: AUC 0.91, from 17 stage II patients in the test
Stage III: AUC 0.94, from 8 stage III patients in the test
Stage unknown: AUC 0.97, from 14 patients whose stage was not recorded
5 Gastric cancer vs no disease · plasma cfDNA methylation · Long 2026 (SRP610281): AUC 0.94; 196 test samples (82 patients, 114 in the comparison group); paper's AUC 0.87.
By cancer stage: P2G finds gastric cancer at every stage, including early stage I.
Stage I: AUC 0.92, from 23 stage I patients in the test
Stage II: AUC 0.94, from 12 stage II patients in the test
Stage III: AUC 0.95, from 40 stage III patients in the test
Stage unknown: AUC 0.89, from 7 patients whose stage was not recorded
6 Breast cancer vs healthy · plasma cfDNA, whole genome · Yang 2021 (SRP302308): AUC 0.94; 65 test samples (42 patients, 23 in the comparison group); paper's AUC 0.83.
By cancer stage: P2G finds breast cancer at every stage, including early stage I.
Stage I: AUC 0.91, from 10 stage I patients in the test
Stage II: AUC 0.94, from 25 stage II patients in the test
Stage III: AUC 1.000, from 6 stage III patients in the test (95% CI 0.81 to 1.00)
Stage IV: AUC 0.83, from 1 stage IV patient in the test (too few to be sure)
7 Preeclampsia vs control pregnancies, later sequencing batch, same hospital · maternal plasma cfRNA · Moufarrej 2022 (PRJNA792450): AUC 0.81; 20 test samples (8 patients, 12 in the comparison group); paper's AUC 0.71. Not a cancer, so there is no stage.
8 Preeclampsia vs control pregnancies, other hospitals · maternal plasma cfRNA · Moufarrej 2022 (PRJNA792450): AUC 0.66; 89 test samples (28 patients, 61 in the comparison group); paper's AUC 0.72. Not a cancer, so there is no stage.
9 Colon cancer vs healthy · plasma cfDNA 5hmC · Li 2017 (SRP080977): AUC 0.78; 29 test samples (10 patients, 19 in the comparison group); paper's AUC 0.95.
By cancer stage:
Stage II: AUC 0.89, from 1 stage II patient in the test (too few to be sure)
Stage III: AUC 0.92, from 6 stage III patients in the test
Stage IV: AUC 0.47, from 3 stage IV patients in the test (too few to be sure)
10 Acute myeloid leukaemia vs control · plasma cfDNA 5hmC · Shao 2022 (SRP299294): AUC 0.83; 32 test samples (24 patients, 8 in the comparison group); paper's AUC 0.98. Leukaemia is not given a stage. The study gives risk groups for all 61 patients: favourable 7, intermediate 33, adverse 21.
11 Colorectal cancer vs healthy · plasma cfDNA 5hmC · Walker 2022, Cambridge Epigenetix (PRJNA755688): AUC 0.71; 209 test samples (125 patients, 84 in the comparison group); paper's AUC 0.90; 26% of cancers found at 90% specificity (32 of 125).
By cancer stage:
Stage I: AUC 0.74, from 28 stage I patients in the test
Stage II: AUC 0.73, from 50 stage II patients in the test
Stage III: AUC 0.62, from 32 stage III patients in the test
Stage IV: AUC 0.81, from 15 stage IV patients in the test
12 Lung cancer vs control · plasma cfDNA 5hmC · Shao 2022 (SRP375150): AUC 0.94; 55 test samples (11 patients, 44 in the comparison group); the paper reports sensitivity and specificity, no AUC (94.1% of lung cancers found at 96.2% specificity). The study does not say which patient had which stage. Across its 97 lung cancer patients: stage I 22, stage II 7, stage III 11, stage IV 57.
13 Colorectal or stomach cancer vs healthy · plasma cfRNA · Tao 2023 (SRP343186): AUC 0.93; 25 test samples (15 patients, 10 in the comparison group); not tested in the paper; predicted by us.
By cancer stage: P2G finds colorectal or stomach cancer at every stage, including early stage I.
Stage I: AUC 0.95, from 6 stage I patients in the test
Stage II: AUC 0.90, from 4 stage II patients in the test (too few to be sure)
Stage III: AUC 0.97, from 3 stage III patients in the test (too few to be sure)
Stage IV: AUC 0.90, from 2 stage IV patients in the test (too few to be sure)
14 Colorectal cancer vs healthy · plasma cfRNA · Gowrisankar 2025 (SRP618157): AUC 0.94; 54 test samples, which are sequencing runs (23 cancer runs from 13 people, 31 comparison runs from 16 people); paper's AUC 0.96; the paper also reports 94.1% of cancers found at 91% specificity.
By cancer stage: P2G finds colorectal cancer at stages II to IV. There was no stage I patient in the test samples.
Stage II: AUC 1.000, from 4 stage II runs in the test (too few to be sure; 95% CI 0.75 to 1.00)
Stage III: AUC 0.99, from 11 stage III runs in the test
Stage IV: AUC 0.98, from 6 stage IV runs in the test
Stage unknown: AUC 0.40, from 2 runs whose stage was not recorded (too few to be sure)
15 Melanoma vs healthy · plasma cfDNA methylation · Dumas 2025 (SRP576689): AUC 0.96; 38 test samples (22 patients, 16 in the comparison group); the paper reports sensitivity and specificity, no AUC: 60% of late-stage (9 of 15) and 27% of early-stage melanomas (4 of 15) found at 98.15% specificity; P2G finds 16 of 22 at 100% specificity.
By cancer stage: P2G finds melanoma at stages II to IV.
Stage II: AUC 1.000, from 1 stage II patient in the test (too few to be sure; 95% CI 0.44 to 1.00)
Stage III: AUC 0.98, from 10 stage III patients in the test
Stage IV: AUC 0.94, from 11 stage IV patients in the test
16 Death after immunotherapy (deceased vs alive) · bladder cancer, whole blood RNA · Chevrier 2026, IOPREDI study (SRP619729): AUC 0.88; 33 test samples (20 patients, 13 in the comparison group); not tested in the paper; predicted by us.
All patients had advanced bladder cancer and were treated with immunotherapy.
Stage III: AUC 1.000, from 3 stage III patients who died (too few to be sure; 95% CI 0.67 to 1.00)
Stage IV: AUC 0.86, from 17 stage IV patients who died
Patients who died are compared with the 13 patients who were still alive.
17 Adenoma vs healthy · plasma cfDNA, shallow whole genome · Van 2026 (SRP736589): AUC 0.74; 102 test samples (24 patients, 78 in the comparison group); no published AUC for this question. Adenoma is a pre-cancer growth, so it has no cancer stage.
18 Adenoma vs healthy · plasma cfDNA 5hmC · Walker 2022, Cambridge Epigenetix (PRJNA755688): AUC 0.61; 118 test samples (34 patients, 84 in the comparison group); 26% of adenomas found at 90% specificity (9 of 34); not tested in the paper; predicted by us. Adenoma is a pre-cancer growth, so it has no cancer stage.
19 Head and neck cancer vs healthy · plasma cfRNA · Tosevska 2022 (GSE196038): AUC 0.81; 12 test samples (6 patients, 6 in the comparison group); small test set; not tested in the paper; predicted by us. The study does not report cancer stage.
20 Colorectal cancer vs healthy · plasma cfDNA methylation · Li 2019 (PRJNA529690): AUC 0.85; 54 test samples (28 patients, 26 in the comparison group); not tested in the paper; predicted by us.
By cancer stage:
Stage I: AUC 0.76, from 7 stage I patients in the test
Stage II: AUC 0.91, from 15 stage II patients in the test
Stage III: AUC 0.78, from 6 stage III patients in the test
21 Liver cancer vs hepatitis B and healthy · plasma cfDNA 5hmC · Song 2017 (SRP074577): AUC 1.000 (95% CI 0.69 to 1.00); 10 test samples (5 patients, 5 in the comparison group); small test set; not tested in the paper; predicted by us; controls are hepatitis B carriers and healthy people. Tumour size stage of the 5 liver cancer patients in the test: T1 (1 patient), T2 (1), unknown (2), and 1 sample taken after surgery.
22 Preterm labour vs control pregnancies · plasma exosome RNA · Fallen 2018 (SRP121661): AUC 0.88; 22 test samples (8 patients, 14 in the comparison group); not tested in the paper; predicted by us. Not a cancer, so there is no stage.
23 Melanoma vs healthy · plasma microRNA · Schneegans 2020 (SRP239782): AUC 1.000 (95% CI 0.90 to 1.00); 46 test samples (32 patients, 14 in the comparison group); not tested in the paper; predicted by us. All 32 melanoma patients in the test had stage IV (spread) melanoma, so the AUC of 1.000 is the stage IV result.
24 Pregnancy complications vs normal pregnancy · plasma cfRNA · Manders 2020 (SRP267167): AUC 0.77; 29 test samples (7 patients, 22 in the comparison group); not tested in the paper; predicted by us. Not a cancer, so there is no stage.
25 Pancreatic cancer vs control · plasma cfDNA methylation · no published paper (SRP289488): AUC 0.88; 15 test samples (7 patients, 8 in the comparison group); small test set; no published paper, so no published AUC. The study does not report cancer stage.
26 Colorectal cancer, blood taken before diagnosis, vs control · plasma cfDNA 5hmC · Gao 2023 (SRP433251): AUC 0.74; 24 test samples (12 patients, 12 in the comparison group); not tested in the paper; predicted by us. No stage: blood was taken before diagnosis, in a screening trial, from people who were diagnosed with colorectal cancer within the next year.
27 Lung cancer, response vs no response to sintilimab immunotherapy · plasma exosome RNA · Cheng 2024 (SRP493158): AUC 1.000 (95% CI 0.69 to 1.00); 10 test samples (5 patients, 5 in the comparison group); small test set; not tested in the paper; predicted by us. All patients had early-stage lung cancer. The study does not say which patient had which stage. Across the 36 trial patients: stage 0 (Tis) 1, stage I 31, stage II 2, stage III 2.
28 Tuberculosis vs no tuberculosis, test group · plasma cfRNA · Chang 2024 (SRP498212): AUC 0.94; 12 test samples (7 patients, 5 in the comparison group); small test set; paper's AUC 0.95. Not a cancer, so there is no stage.
29 Diffuse large B-cell lymphoma vs follicular lymphoma · plasma cfDNA 5hmC · Chiu 2021 (SRP273769): AUC 0.94; 15 test samples (10 patients, 5 in the comparison group); small test set; paper's AUC 0.92. The study does not say which patient had which stage. Across its 48 large B-cell lymphoma patients: stage I or II 17, stage III or IV 20, not recorded 11. Across its 25 follicular lymphoma patients: stage I or II 5, stage III or IV 17, not recorded 3.

Paper AUCs: each study's publicationupdated 2026-11-09 15:02 PST

Same question as the paper · reads used

P2G is within 0.02 of the paper, or higher, in 11 of 28 comparisons, using 4% to 23% of the paper's reads

One dot per question with a published AUC for the same question (28 comparisons, 25 studies). x: reads P2G used per sample, as a percent of the reads the paper used (log scale). y: P2G AUC minus the paper's AUC. Each question has one P2G result here. Hover a dot for its test samples, reads and tumour stage.

within 0.02 of the paper, or higherpaper higher

The two dots at 100% are Alzheimer's and Parkinson's disease from blood DNA (SRP722164, Pollard 2026). The paper first estimates how much of the cell-free DNA comes from each type of neuron, using the methylation of every read and a reference map of neuron methylation made outside the study. It then scores those amounts. It reports no result from the sequences alone. P2G used only the sequences, and the same reads as the paper.

All 28 comparisons with tumour stage

1 colorectal cancer (SRP345647): P2G 0.956 vs paper 0.658; 58 test samples (40 patients, 18 in the comparison group).
By cancer stage: P2G finds colorectal cancer at every stage, including early stage I.
Stage I: AUC 1.000, from 2 stage I patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage II: AUC 0.98, from 7 stage II patients in the test
Stage III: AUC 0.94, from 13 stage III patients in the test
Stage IV: AUC 0.96, from 18 stage IV patients in the test
2 gastric cancer vs benign (SRP468146): P2G 1.000 vs paper 0.870; 20 test samples (10 patients, 10 in the comparison group).
By cancer stage: P2G finds gastric cancer at every stage, including early stage I.
Stage I: AUC 1.000, from 2 stage I patients in the test (too few to be sure; 95% CI 0.59 to 1.00)
Stage II: AUC 1.000, from 2 stage II patients in the test (too few to be sure; 95% CI 0.59 to 1.00)
Stage III: AUC 1.000, from 6 stage III patients in the test (95% CI 0.77 to 1.00)
3 breast cancer (SRP302308): P2G 0.941 vs paper 0.834; 65 test samples (42 patients, 23 in the comparison group).
By cancer stage: P2G finds breast cancer at every stage, including early stage I.
Stage I: AUC 0.91, from 10 stage I patients in the test
Stage II: AUC 0.94, from 25 stage II patients in the test
Stage III: AUC 1.000, from 6 stage III patients in the test (95% CI 0.81 to 1.00)
Stage IV: AUC 0.83, from 1 stage IV patient in the test (too few to be sure)
4 adverse pregnancy outcome: preeclampsia, gestational hypertension, gestational diabetes or chronic hypertension vs no complication (SRP271730): P2G 1.000 (95% CI 0.84 to 1.00) vs paper 0.910; 26 test samples (17 patients, 9 in the comparison group). Not a cancer, so there is no stage.
5 gastric cancer (SRP610281): P2G 0.937 vs paper 0.870; 196 test samples (82 patients, 114 in the comparison group).
By cancer stage: P2G finds gastric cancer at every stage, including early stage I.
Stage I: AUC 0.92, from 23 stage I patients in the test
Stage II: AUC 0.94, from 12 stage II patients in the test
Stage III: AUC 0.95, from 40 stage III patients in the test
Stage unknown: AUC 0.89, from 7 patients whose stage was not recorded
6 Diffuse large B-cell lymphoma vs follicular lymphoma (SRP273769): P2G 0.940 vs paper 0.917; 15 test samples (10 patients, 5 in the comparison group). The study does not say which patient had which stage. Across its 48 large B-cell lymphoma patients: stage I or II 17, stage III or IV 20, not recorded 11. Across its 25 follicular lymphoma patients: stage I or II 5, stage III or IV 17, not recorded 3.
7 lung cancer (SRP432064): P2G 0.975 vs paper 0.956; 44 test samples (19 patients, 25 in the comparison group).
By cancer stage: P2G finds lung cancer at every stage, including early stage I.
Stage I: AUC 0.93, from 7 stage I patients in the test
Stage II: AUC 1.000, from 2 stage II patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage III: AUC 1.000, from 4 stage III patients in the test (too few to be sure; 95% CI 0.75 to 1.00)
Stage IV: AUC 1.000, from 6 stage IV patients in the test (95% CI 0.81 to 1.00)
8 breast cancer (SRP420167): P2G 0.946 vs paper 0.910; 104 test samples (48 patients, 56 in the comparison group).
By cancer stage: P2G finds breast cancer at every stage, including early stage I.
Stage I: AUC 0.98, from 9 stage I patients in the test
Stage II: AUC 0.91, from 17 stage II patients in the test
Stage III: AUC 0.94, from 8 stage III patients in the test
Stage unknown: AUC 0.97, from 14 patients whose stage was not recorded
9 glioblastoma (SRP278818): P2G 0.974 vs paper 0.970; 122 test samples (49 patients, 73 in the comparison group). Glioblastoma is not given a stage; all patients had grade 4 tumours.
10 tuberculosis (SRP498212): P2G 0.943 vs paper 0.947; 12 test samples (7 patients, 5 in the comparison group). Not a cancer, so there is no stage.
11 colon cancer (SRP432064): P2G 0.962 vs paper 0.978; 44 test samples (19 patients, 25 in the comparison group).
By cancer stage: P2G finds colon cancer at every stage, including early stage I.
Stage I: AUC 0.93, from 5 stage I patients in the test
Stage II: AUC 0.96, from 6 stage II patients in the test
Stage III: AUC 0.99, from 6 stage III patients in the test
Stage IV: AUC 0.98, from 2 stage IV patients in the test (too few to be sure)
12 advanced breast cancer (PRJNA494975): P2G 0.929 vs paper 0.971; 22 test samples (14 patients, 8 in the comparison group). All 14 breast cancer patients in the test had stage IV (advanced) breast cancer, so its AUC is the stage IV result.
13 preeclampsia (SRP546783): P2G 0.822 vs paper 0.868; 74 test samples (20 patients, 54 in the comparison group). Not a cancer, so there is no stage.
14 gastric cancer (SRP080977): P2G 0.921 vs paper 0.970; 30 test samples (12 patients, 18 in the comparison group).
By cancer stage:
Stage I: AUC 1.000, from 2 stage I patients in the test (too few to be sure; 95% CI 0.60 to 1.00)
Stage II: AUC 0.86, from 2 stage II patients in the test (too few to be sure)
Stage III: AUC 0.88, from 5 stage III patients in the test
Stage IV: AUC 0.97, from 2 stage IV patients in the test (too few to be sure)
Stage unknown: AUC 1.000, from 1 patient whose stage was not recorded (too few to be sure; 95% CI 0.44 to 1.00)
15 colorectal cancer (SRP341621): P2G 0.857 vs paper 0.930; 17 test samples (10 patients, 7 in the comparison group).
By cancer stage:
Stage I: AUC 1.000, from 2 stage I patients in the test (too few to be sure; 95% CI 0.58 to 1.00)
Stage II: AUC 0.79, from 4 stage II patients in the test (too few to be sure)
Stage III: AUC 1.000, from 2 stage III patients in the test (too few to be sure; 95% CI 0.58 to 1.00)
Stage IV: AUC 0.71, from 2 stage IV patients in the test (too few to be sure)
16 tuberculosis (SRP498188): P2G 0.830 vs paper 0.921; 26 test samples (15 patients, 11 in the comparison group). Not a cancer, so there is no stage.
17 Parkinson's disease (SRP722164): P2G 0.923 vs paper 1.000; 24 test samples (13 patients, 11 in the comparison group). Not a cancer, so there is no stage.
18 pancreatic cancer vs benign (SRP519537): P2G 0.795 vs paper 0.945; 22 test samples (8 patients, 14 in the comparison group).
By cancer stage:
Stage I: AUC 0.93, from 1 stage I patient in the test (too few to be sure)
Stage III: AUC 0.88, from 4 stage III patients in the test (too few to be sure)
Stage IV: AUC 0.64, from 3 stage IV patients in the test (too few to be sure)
19 acute myeloid leukemia (SRP299294): P2G 0.828 vs paper 0.981; 32 test samples (24 patients, 8 in the comparison group). Leukaemia is not given a stage. The study gives risk groups for all 61 patients: favourable 7, intermediate 33, adverse 21.
20 colon cancer (SRP080977): P2G 0.784 vs paper 0.950; 29 test samples (10 patients, 19 in the comparison group).
By cancer stage:
Stage II: AUC 0.89, from 1 stage II patient in the test (too few to be sure)
Stage III: AUC 0.92, from 6 stage III patients in the test
Stage IV: AUC 0.47, from 3 stage IV patients in the test (too few to be sure)
21 stage I colorectal cancer (SRP636578): P2G 0.767 vs paper 0.950; 11 test samples (5 patients, 6 in the comparison group). All 5 colorectal cancer patients in the test had stage I cancer, so its AUC is the stage I result.
22 case vs control (SRP576513): P2G 0.621 vs paper 0.810; 34 test samples (19 patients, 15 in the comparison group). Not a cancer, so there is no stage.
23 cancer vs no cancer (SRP502369): P2G 0.702 vs paper 0.980; 26 test samples (14 patients, 12 in the comparison group). The study does not say which patient had which stage. Across the 27 colorectal cancer patients in its discovery group: stage 0 5, stage I 6, stage II 8, stage III 8.
24 peritoneal metastasis (SRP566788): P2G 0.518 vs paper 0.827; 22 test samples (14 patients, 8 in the comparison group). The study does not report cancer stage. Patients are grouped by whether the cancer had spread to the lining of the belly.
25 lung cancer, controlled vs not (SRP435125): P2G 0.675 vs paper 0.998; 60 test samples (48 patients, 12 in the comparison group). All patients had lung cancer that had spread to other organs. The study does not say which patient had which stage.
26 tuberculosis (SRP498234): P2G 0.514 vs paper 0.946; 12 test samples (7 patients, 5 in the comparison group). Not a cancer, so there is no stage.
27 Alzheimer's disease (SRP722164): P2G 0.707 vs paper 0.999; 20 test samples (9 patients, 11 in the comparison group). Not a cancer, so there is no stage.
28 colorectal cancer (SRP618157): P2G 0.940 vs paper 0.960; 54 test samples, which are sequencing runs (23 cancer runs from 13 people, 31 comparison runs from 16 people).
By cancer stage: P2G finds colorectal cancer at stages II to IV. There was no stage I patient in the test samples.
Stage II: AUC 1.000, from 4 stage II runs in the test (too few to be sure; 95% CI 0.75 to 1.00)
Stage III: AUC 0.99, from 11 stage III runs in the test
Stage IV: AUC 0.98, from 6 stage IV runs in the test
Stage unknown: AUC 0.40, from 2 runs whose stage was not recorded (too few to be sure)

Reads in the paper: median reads sequenced per sample (public sequencing records)updated 2026-11-09 15:02 PST

Same question · the paper's own measure

On the papers' own sensitivity and specificity, P2G is higher in 2 of 4, about the same in 1 and lower in 1

These 4 papers tested the same question but reported the percent of cases found at a set specificity, not an AUC. P2G is scored the same way: the percent of its held-out cases found at the paper's specificity, from its held-out test samples only.

question · sample · study	paper	P2G, held-out test samples	result	stage
Melanoma vs healthy plasma cfDNA methylation · Dumas 2025 (SRP576689)	60% of late-stage (9 of 15) and 27% of early-stage melanomas (4 of 15) at 98.15% specificity	73% of melanomas (16 of 22) at 100% specificity; 38 test samples (22 patients, 16 in the comparison group)	P2G higher	P2G finds melanoma at stages II to IV (stage II, AUC 1.000).
Preeclampsia vs control pregnancies serum small RNA · UCSD triage cohort · Morey 2023	3-microRNA ratio: 91% sensitivity and 57% specificity on a separate group of 18 (11 cases, 7 controls)	93% (13 of 14) at 88% specificity (7 of 8), and 100% (14 of 14) at 62% specificity; 22 test samples (14 patients, 8 with normal blood pressure)	P2G higher	Not a cancer, so there is no stage. The paper's controls mix normal and high blood pressure; the P2G test group has normal blood pressure.
Adenoma vs healthy plasma cfDNA, shallow whole genome · Van 2026 (SRP736589)	40% of precancerous lesions at 92.4% specificity from fragment patterns, the same kind of data P2G used (73.3% at 91.0% with a separate methylation PCR test, not in the public data)	38% (9 of 24) at 91% specificity (Wilson 95% CI 21% to 57%); 102 test samples (24 patients, 78 healthy controls)	about the same	Adenoma is a pre-cancer growth, so it has no cancer stage.
Lung cancer vs control plasma cfDNA 5hmC · Shao 2022 (SRP375150)	94.1% of lung cancers at 96.2% specificity on its validation set	45% (5 of 11) at 96.2% specificity or more; 55 test samples (11 patients, 44 in the comparison group)	paper higher	The study does not say which patient had which stage; across its 97 lung cancer patients: stage I 22, stage II 7, stage III 11, stage IV 57.

Held-out samples of each study; paper values from each paperupdated 2026-11-11 09:21 PST

Sensitivity at fixed specificity

At 95% specificity, P2G achieves >=80% specificity

Questions with a held-out AUC of 0.85 or more on 20 or more test samples (18 questions); shown: the 15 where P2G finds half of the cases or more at 95% specificity. Bar: percent of cases found when 95% of controls are called negative; ticks: the same at 90% and 99% specificity. Hover a bar for the tumour stage.

liquid biopsy│ thin tick: 90% specificity · thick tick: 99% specificitywhisker: 95% CI of the percent of patients found at 95% specificity (wider when there are few cases)

Test samples of each studyupdated 2026-11-09 12:36 PST

Main point 3

Models trained in one study work on a study from another lab (AUC 0.69 to 1.000)

Question: does a model fitted in one study keep working, unchanged, on samples from a different lab?

Cross-study · applied unchanged

Table 2Cross-study: five models score 0.69 to 1.000 on a study from another lab

Each model was trained in one study and applied unchanged to every sample of the other study. None of the pairs is from the same lab.

sample	trained on	scored on	stage	AUC	test samples	check
Plasma RNA	Stanford (PRJNA792450): preeclampsia vs control	UCSD discovery: severe preeclampsia vs control	not a cancer	0.69	187 test samples (51 patients, 136 in the comparison group)	Same model on its own Stanford test samples: 0.81, 77 test samples (17 patients, 60 in the comparison group)
Plasma RNA	UCSD discovery: preeclampsia vs control	Stanford (PRJNA792450): severe preeclampsia vs control	not a cancer	0.70	344 test samples (57 patients, 287 in the comparison group)	232 of the model's 241 sequences (96%) are found in the other cohort
Plasma RNA	UCSD triage: preeclampsia vs control	Stanford (PRJNA792450): severe preeclampsia vs control	not a cancer	0.73	344 test samples (57 patients, 287 in the comparison group)	Same model on its own UCSD triage test samples: 0.61, 24 test samples (14 patients, 10 in the comparison group); 35 of its 401 sequences (9%) are found in the other cohort
cfDNA	SRP080977 (5hmC cfDNA, Li 2017): colorectal cancer vs healthy	SRP636578 (5mC and 5hmC cfDNA, Puddu 2026): stage I colorectal cancer vs control	Scored on stage I patients only. Training study (Li 2017) colon cancer samples: stage I 6, stage II 5, stage III 36, stage IV 18, not recorded 12.	0.69	58 test samples (26 patients, 32 in the comparison group)	Same model on its own held-out samples: 0.73, 29 test samples (10 patients, 19 in the comparison group); 39 of its 498 sequences (8%) are found in the other study
cfDNA	SRP080977 (5hmC cfDNA, Li 2017): colorectal cancer vs healthy	SRP636578: stage I or IV colorectal cancer vs control	Scored on stage I and stage IV patients together. Training study (Li 2017) colon cancer samples: stage I 6, stage II 5, stage III 36, stage IV 18, not recorded 12.	0.72	69 test samples (37 patients, 32 in the comparison group)
cfDNA	SRP074577 (5hmC cfDNA, Song 2017): liver cancer vs hepatitis B or healthy	SRP690785 (serum cfDNA, 2026): liver cancer vs healthy	Scored study (Sun 2026): the study does not say which patient had which stage. Across its 23 liver cancer patients: early stage 11, advanced stage 12. Training study (Song 2017): tumour size stage T1 to T3 before surgery, plus samples taken after surgery or at recurrence.	1.000 95% CI 0.93 to 1.00	56 test samples (23 patients, 33 in the comparison group)	Same model on its own held-out samples: 0.84, 10 test samples (5 patients, 5 in the comparison group); 509 of its 733 sequences are found in the other study

Example genes the models use on the other study (two per model, leaving out mitochondrial genes; only sequences found in both studies count).

Stanford to UCSD discovery · AUC 0.69 on 187 samples of the other study (51 patients, 136 in the comparison group) · 56 of the model's 669 sequences (8%) found in the other study, 2 of them in named genes (9.4% of the matched weight)

example genes · higher in	what it is	known in the literature
ALPK2sequences higher in preeclampsiaRank 1 of 2 genes among the 56 matched sequences · 5.8% of the matched weight · 1 of 1 sequence higher in preeclampsia	Alpha kinase 2
RN7SL396Psequences higher in controlRank 2 of 2 genes among the 56 matched sequences · 3.6% of the matched weight · 1 of 1 sequence higher in control	RNA, 7SL, cytoplasmic 396, pseudogene

UCSD discovery to Stanford · AUC 0.70 on 344 samples of the other study (57 patients, 287 in the comparison group) · 232 of the model's 241 sequences (96%) found in the other study, 10 of them in named genes (3.7% of the matched weight)

example genes · higher in	what it is	known in the literature
LRRTM3sequences higher in controlRank 3 of 7 genes among the 232 matched sequences · 0.6% of the matched weight · 1 of 1 sequence higher in control	Leucine rich repeat transmembrane neuronal 3
RNA5SP202sequences higher in controlRank 4 of 7 genes among the 232 matched sequences · 0.3% of the matched weight · 1 of 1 sequence higher in control	RNA, 5S ribosomal pseudogene 202

SRP080977 to SRP636578 · AUC 0.72 on stage I or IV patients (37 patients, 32 in the comparison group) and 0.69 on stage I patients only (26 patients, 32 in the comparison group) · 39 of the model's 498 sequences (8%) found in the other study, 24 of them in named genes (62.5% of the matched weight)

example genes · higher in	what it is	known in the literature
LUCAT1sequences higher in healthyRank 1 of 24 genes among the 39 matched sequences · 16.1% of the matched weight · 1 of 1 sequence higher in healthy	Lung cancer associated transcript 1
USP37sequences higher in healthyRank 2 of 24 genes among the 39 matched sequences · 8.0% of the matched weight · 1 of 1 sequence higher in healthy	Ubiquitin specific peptidase 37

Moufarrej 2022; Srinivasan 2020; Li 2017; Puddu 2026; Song 2017; Sun 2026updated 2026-11-09 14:45 PST

Cross-study · what the models that work on another study use

31 cross-study scores between different labs reach 0.65 or more; the cancer models rely mostly on sequences in genes (57%)

Question: what do the models that score 0.65 or more on a study from another lab have in common? Counted over 31 cross-study scores of 0.65 or more between studies of different labs, from 24 models (highest 0.77). Pairs of studies from the same lab or possibly the same lab are left out: 129 scores. Shares are of the model's sequences that are also found in the other study.

which transfers	in genes incl. mitochondrial and rRNA genes	between genes	not placed on the genome	higher in the same group in the scored study
Cancer studies 23 scores, 18 models	57%	30%	13%	57% (2,889 of 5,030)
DNA (5hmC cfDNA) 14 scores, 11 models	62%	35%	3%	58% (1,799 of 3,121)
Preeclampsia, Stanford and UCSD 8 scores, 6 models	2%	1%	97%	not recorded

Genes among the top 10 sequences of 2 or more of these models (number of models):

RPPH1 · 2 AGBL4 · 2 C2CD6 · 2 RN7SL1 · 2 RPS29 · 2

Cross-study scores between different labsupdated 2026-11-09 12:36 PST

Within each study · all questions

Within their own study, 49 of 68 predictions reach a held-out AUC of 0.80 or more

One dot per study question (94 questions), the best model of each, scored on its own held-out samples. Colour: disease type. Count in the title: questions with 20 or more test samples. Hollow dots: fewer than 20. Hover a cancer dot for its tumour stage.

cancerpregnancyneurologicalinfection and otherhollow: fewer than 20 test samplesdashed line: AUC 0.80

Held-out samples of each studyupdated 2026-11-09 12:36 PST

Cross-study · preeclampsia · Stanford and UCSD

10 models trained on one center score 0.60 to 0.73 when tested on the other (Stanford and UCSD)

Each row is one model: trained on the first question (one cohort), then applied unchanged to the second question (the other cohort). Shown: every train and test combination that scores 0.60 or more with 20 or more of the model's sequences found in the other cohort; the best model of each combination. Not a cancer, so there is no stage.

example gene · weight · trained on, tested on	what it is	known with preeclampsia in the literature
ALPK2 5.8% of the weight · 1 sequence, 1 higher in preeclampsia trained on Stanford: preeclampsia vs control, tested on UCSD discovery: severe preeclampsia vs control (0.69)	alpha-kinase 2, a protein kinase	none found (Europe PMC, preeclampsia in title or abstract)
RN7SL396P 3.6% of the weight · 1 sequence, 0 higher in preeclampsia trained on Stanford: preeclampsia vs control, tested on UCSD discovery: severe preeclampsia vs control (0.69)	7SL RNA pseudogene	none found (Europe PMC, preeclampsia in title or abstract)
SNORD3B-1 0.8% of the weight · 1 sequence, 1 higher in preeclampsia trained on UCSD discovery: severe preeclampsia vs control, tested on Stanford: severe preeclampsia vs control (0.68)	small nucleolar RNA U3	none found (Europe PMC, preeclampsia in title or abstract)

Moufarrej 2022 (Stanford); Srinivasan 2020 and Morey 2023 (UCSD)updated 2026-11-09 14:06 PST

Main point 4

Broader insights: on questions the papers did not test, 5 of 11 reach an AUC of 0.80 or more on 20 or more test samples

Question: what does P2G find for groups the published papers did not compare?

Broader insights · questions the papers did not test

5 questions the papers did not score reach an AUC of 0.80 or more on 20 or more test samples

Each question was answered from the study's own samples; the paper did not report a score for it. Held-out test samples only. The chart at the end of this section shows the same questions.

question · sample · study	held-out AUC	test samples	stage
Melanoma vs healthy plasma microRNA · Schneegans 2020 (SRP239782)	1.000 (95% CI 0.90 to 1.00)	46 test samples (32 patients, 14 in the comparison group)	Stage IV for all 32 melanoma patients in the test, so the AUC of 1.000 is the stage IV result.
Colorectal or stomach cancer vs healthy plasma cfRNA · Tao 2023 (SRP343186)	0.93	25 test samples (15 patients, 10 in the comparison group)	Stage I, AUC 0.95 (6 patients); stage II, AUC 0.90 (4 patients, too few to be sure); stage III, AUC 0.97 (3 patients, too few to be sure); stage IV, AUC 0.90 (2 patients, too few to be sure).
Death after immunotherapy (deceased vs alive), bladder cancer whole blood RNA · Chevrier 2026, IOPREDI study (SRP619729)	0.88	33 test samples (20 patients, 13 in the comparison group)	Stage III, AUC 1.000 (3 patients who died, too few to be sure; 95% CI 0.67 to 1.00); stage IV, AUC 0.86 (17 patients who died).
Preterm labour vs control pregnancies plasma exosome RNA · Fallen 2018 (SRP121661)	0.88	22 test samples (8 patients, 14 in the comparison group)	Not a cancer, so there is no stage.
Colorectal cancer vs healthy plasma cfDNA methylation · Li 2019 (PRJNA529690)	0.85	54 test samples (28 patients, 26 in the comparison group)	Stage I, AUC 0.76 (7 patients); stage II, AUC 0.91 (15 patients); stage III, AUC 0.78 (6 patients).
fewer than 20 test samples (hollow dots on the chart)
Liver cancer vs hepatitis B and healthy plasma cfDNA 5hmC · Song 2017 (SRP074577)	1.000 (95% CI 0.69 to 1.00)	10 test samples (5 patients, 5 in the comparison group)	Unspecified stage; tumour size stage of the 5 patients: T1 (1), T2 (1), unknown (2), and 1 sample taken after surgery.
Lung cancer, response vs no response to sintilimab immunotherapy plasma exosome RNA · Cheng 2024 (SRP493158)	1.000 (95% CI 0.69 to 1.00)	10 test samples (5 patients, 5 in the comparison group)	Early stage lung cancer; the study does not say which patient had which stage (across the 36 trial patients: stage 0 1, stage I 31, stage II 2, stage III 2).
Head and neck cancer vs healthy plasma cfRNA · Tosevska 2022 (GSE196038)	0.81	12 test samples (6 patients, 6 in the comparison group)	Unspecified stage (the study does not report stage).

Two more questions are in the cards below: preeclampsia vs normal blood pressure (AUC 0.96) and lung cancer from 5hmC (AUC 0.94). Their papers reported sensitivity and specificity, not an AUC.

Held-out samples of each studyupdated 2026-11-11 07:39 PST

Pregnancy · serum small RNA · UCSD triage cohort

Preeclampsia is told apart from normal blood pressure at AUC 0.96, a score the paper did not report

22 test samples (14 patients, 8 in the comparison group) from women kept out of training. P2G used every read. Not a cancer, so there is no stage.

The paper reports sensitivity and specificity of a 3-microRNA ratio, not an AUC (Morey 2023).

Example genes · The model uses 992 selected sequences; 50 of them (5%) are placed on the human genome, in 30 genes.

example genes · sequences higher in	what it is	known in the literature
SNORD87sequences higher in preeclampsiaRank 1 of 30 genes · 0.5% of the model's weight · 6 of 6 sequences higher in preeclampsia	Small nucleolar RNA, C/D box 87
SH3GL3sequences higher in preeclampsiaRank 2 of 30 genes · 0.3% of the model's weight · 2 of 2 sequences higher in preeclampsia	SH3 domain containing GRB2 like 3, endophilin A3

UCSD triage cohort, no public accession recorded · Morey et al. 2023, Science Advancesupdated 2026-11-09 12:36 PST

Cancer · plasma cell-free DNA, 5hmC · SRP375150

Lung cancer is found from 5hmC in plasma DNA at AUC 0.94, with 9 of 11 cancers found at 95% specificity

55 test samples (11 patients, 44 in the comparison group) kept out of training. The study does not say which patient had which stage. Across its 97 lung cancer patients: stage I 22, stage II 7, stage III 11, stage IV 57.

The paper reports 94.1% of lung cancers found at 96.2% specificity on its own validation set, and no AUC (Shao 2022); the square marks it. At 96.2% specificity or more, P2G finds 5 of 11 in its test samples.

Example genes · The model uses 474 selected sequences; 411 of them (87%) are placed on the human genome, in 264 genes.

example genes · sequences higher in	what it is	known in the literature
DTNBsequences higher in controlRank 1 of 264 genes · 3.2% of the model's weight · 1 of 1 sequence higher in control	Dystrobrevin beta
MBOAT2sequences higher in controlRank 2 of 264 genes · 3.0% of the model's weight · 2 of 3 sequences higher in control	Membrane bound glycerophospholipid O-acyltransferase 2

SRP375150 · Shao et al. 2022, Clinical and Translational Discoveryupdated 2026-11-09 12:59 PST

Not tested in the paper; predicted by us

Untested predictions exceed 0.80

Shown: every untested question at 0.80 or more, highest first. Hollow: fewer than 20 test samples. Every 1.000 shows its 95% CI; Hover a dot for its tumour stage.

Held-out samples of each studyupdated 2026-11-09 13:16 PST

New insights · genes not yet linked to the disease

The 2 genes with the most model weight per study that no paper links to the disease yet

Studies with 20 or more test samples. For each model: the two named genes with the largest share of the model's weight that no PubMed paper names with the disease in its title or abstract. Mitochondrial genes and genes with no gene name are left out.

gene · sequences higher in	what it is	known in the literature
Melanoma vs healthy (SRP239782)
TAF12-DTsequences higher in melanoma0.75% of the model's weight · 1 of 2 sequences higher in melanoma	RNA next to the TAF12 gene (divergent transcript)
HJVsequences higher in melanoma0.69% of the model's weight · 1 of 1 sequence higher in melanoma	Hemojuvelin, a regulator of iron levels
Colorectal or stomach cancer vs healthy (SRP343186)
ALAS2sequences higher in cancer4.2% of the model's weight · 14 of 14 sequences higher in cancer	Aminolevulinate synthase 2, the red blood cell heme enzyme
HSH2D-AS1sequences higher in healthy1.8% of the model's weight · 3 of 3 sequences higher in healthy	Antisense RNA near HSH2D
Death after immunotherapy, bladder cancer (SRP619729)
SCARNA2sequences higher in deceased0.95% of the model's weight · 17 of 17 sequences higher in deceased	Small Cajal body RNA 2, guides chemical changes on other RNAs
MYZAPsequences higher in deceased0.83% of the model's weight · 10 of 10 sequences higher in deceased	Myocardial zonula adherens protein
Preterm labour vs control pregnancies (SRP121661)
MIR335sequences higher in control0.81% of the model's weight · 1 of 1 sequence higher in control	MicroRNA 335
MIR191sequences higher in preterm labour0.35% of the model's weight · 1 of 1 sequence higher in preterm labour	MicroRNA 191
Colorectal cancer vs healthy, methylation (PRJNA529690)
TRIM5sequences higher in colorectal cancer1.55% of the model's weight · 8 of 8 sequences higher in colorectal cancer	Tripartite motif 5, an antiviral protein
DLGAP2sequences higher in colorectal cancer0.82% of the model's weight · 2 of 2 sequences higher in colorectal cancer	DLG associated protein 2, a synapse scaffold
Preeclampsia vs normal blood pressure (UCSD triage cohort)
SNORD87sequences higher in preeclampsia0.5% of the model's weight · 6 of 6 sequences higher in preeclampsia	Small nucleolar RNA, C/D box 87
SH3GL3sequences higher in preeclampsia0.3% of the model's weight · 2 of 2 sequences higher in preeclampsia	SH3 domain containing GRB2 like 3, endophilin A3
Lung cancer vs control, 5hmC (SRP375150)
FAM20Bsequences higher in control1.74% of the model's weight · 1 of 1 sequence higher in control	FAM20B glycosaminoglycan xylosylkinase, starts sugar chains on proteins
ARMH1sequences higher in control1.66% of the model's weight · 1 of 1 sequence higher in control	Armadillo like helical domain containing 1

Weight: each sequence counts by how much the model relies on it; a gene adds up its sequences. Higher in: the group whose score the gene's sequences raise. Literature: PubMed search of the gene symbol with the disease words in the title or abstract.

Models behind each AUC aboveupdated 2026-11-11 07:39 PST

Definitions

Terms used on this page

AUC: How well a score separates two groups: 1 is perfect, 0.5 is no better than a coin flip.
95% CI next to 1.000: The range the AUC would likely fall in with more samples of the same kind (Newcombe 2006).
Test samples: Samples kept out of training and scored only once, at the end. Written as N test samples (cases, controls).
Sensitivity at 95% specificity: The percent of cases found when the cutoff is set so that 95% of controls are called negative.
Cross-study: A model fitted in one study and applied unchanged to all samples of another study.
Selected sequence and model weight: A short read sequence the model uses; its weight is how much the model relies on it, as a percent of the whole model.
Tumour DNA percent: How much of a library's DNA comes from the SEQC2 reference tumour DNA (0%, 0.8%, 4% or 20%).
R²: How much of the spread in the true values the predictions explain: 1 is all of it.
Spike-in: A known amount of reference material mixed into a sample, so the right answer is known before the test.
Sensitivity CI: The range the percent of cases found would likely fall in with more cases (Wilson).