Skip to content

Dataset specification, clinical annotations, and experimental design for the CPTAC pan-cancer & clinical biopsy projection study (#566) #577

Open

Description

@aarmey

Background & Connection to #566

Issue #566 ("Out-of-sample peptide fold-in and patient biopsy projection for pan-cancer reference atlases and longitudinal clinical trials", Milestone: CPTAC study with DDMC) proposes extending DDMC from a single-cohort transductive clustering algorithm into a reference-to-query transfer learning framework for clinical cancer phosphoproteomics.

Following the completion of core algorithmic prerequisites—e-means mini-batch EMA scaling (#568), out-of-sample peptide fold-in DDMC.fold_in() and bi-cross-validation (#569), expansion of the PSPL atlas to 404 human kinases (#570), exact observed-only diagonal Gaussian marginalization and simulate_dropout() (#571), and flanking PTM crosstalk tokens (#572)—this issue specifies the exact public datasets, data scales, file sizes, download endpoints, verified clinical/genomic/immune annotations, testable hypotheses, and proposed computational approach for the study.


Clinical Use Case & Central Hypothesis

Clinical Problem

In precision oncology, actionable signaling states (target kinase hyperactivation, adaptive bypass rewiring under targeted therapy, and immune-evasive tumor microenvironment states) are governed by post-translational phosphorylation rather than DNA mutations alone (e.g., "pseudo-HER2+" tumors that lack downstream ERBB2 signaling, or EGFR/KRAS-mutant tumors with divergent bypass pathway activation). However, clinical core-needle biopsies (14G–18G) and microscale/DIA phosphoproteomics workflows suffer from:

  1. Small cohort sizes ($n = 15\text{–}80$ patients) insufficient to de novo learn 30–100 kinase-substrate modules and position-specific scoring matrices (PSSMs) without overfitting.
  2. Severe stochastic and intensity-dependent peptide dropout (50–85% missingness) and batch-specific peptide identification, where single-site phosphopeptide biomarkers frequently drop out in individual patient biopsies.
  3. Unseen phosphosites detected in a query clinical trialplex that were not present in the discovery cohort's filtered feature intersection.

Central Hypothesis

  1. Conserved Kinase-Substrate Module Transfer: A reference DDMC model trained on a deep, harmonized pan-cancer or discovery atlas (>1,000 tumors, >50,000–110,000 phosphosites) learns biologically conserved kinase-substrate clusters defined jointly by sequence recognition motifs ($\mathcal{M}_c$) and co-regulated phosphorylation dynamics.
  2. Sequence-Informed Biopsy Projection: Projecting an out-of-sample patient biopsy onto fixed reference modules (DDMC.project_samples)—combining overlapping phosphosites (weighted by reference posterior responsibilities $\gamma_{ic}$ and site precisions $\tau_{ic}$) with novel biopsy phosphosites folded in via sequence motifs (DDMC.fold_in(..., cols=[]))—recovers accurate patient-level kinase module activities ($\hat{\mu}_{\text{new}, c}$) even when 50–85% of phosphosites are missing.
  3. Superior Clinical Biomarker & Pharmacodynamic Sensitivity: Reference-projected DDMC module signatures outperform (a) single-site phosphopeptide biomarkers, (b) sample-wise KSEA / PTM-SEA enrichment scores, and (c) expression-only dimensionality reduction (seq_weight = 0, PCA/NMF) in predicting somatic driver mutations, immune infiltration, longitudinal target inhibition, and pathological complete response (pCR) across independent cohorts and clinical trials.

Specific Datasets to Use (Verified Scale, File Sizes & Download Locations)

All datasets below are publicly available without embargo via the NCI Proteomic Data Commons (PDC), the CPTAC Pan-Cancer Portal, and the cptac Python API, with verified clinical, genomic, and treatment-outcome annotations.

Summary Table of Datasets

Tier & Role Dataset / Study Name Accession / File ID Scale (Patients / Samples $\times$ Phosphosites) Matrix Download Size (Total Study Size) Verified Clinical & Molecular Annotations Available
Tier 1A: Pan-Cancer Reference Atlas (Broad Spectrum Mill) CPTAC Pan-Cancer Harmonized Phosphoproteome (Geffen et al., Cell 2023) PDC Pan-Cancer UUID fc771375-c1d2-4533-86c8-3a617f21d331 (Phosphoproteome_Broad_Institute_harmonized_v1.tsv.gz) 1,119 patient samples $\times$ 110,274 phosphosites across 11 cohorts (BRCA, ccRCC, COAD, GBM, HGSC, HNSCC, LSCC, LUAD, PDAC, UCEC, MB) 95.29 MB .tsv.gz Paired with Clinical_meta_data_v1.zip (0.74 MB, 1,081 cases, 197–245 cols/cohort) & Mutation_Broad_WashU_union_v1.zip (131.38 MB compressed / 531.32 MB MAF)
Tier 1B: Pan-Cancer Reference Atlas (UMich / BCM / Sinai GENCODE v34) CPTAC Pan-Cancer Resource (Li et al., Cancer Cell 2023) PDC UUIDs 76378111-b53d-4435-9463-e5c416b9de5e (BCM), 0116b655-2b6a-4d18-a379-98d19be22f89 (UMich), f0c1a635-fedc-4cf2-996e-da8d027b15ee (Sinai) 1,081 cases (Tumor + Matched Normal) across 10 cohorts; ~25,000–55,000 single-site & multi-site phosphosites per cohort (>100,000 union sites) 275.67 MB (BCM .zip), 1,643.30 MB (UMich .zip), 1,372.99 MB (Sinai .zip) OS/PFS survival, stage, grade, driver mutations, CIBERSORT (22 cell types), xCell (38 scores), ESTIMATE, PROGENy (14 pathways), Hallmark, & pre-computed PTM-SEA
Tier 1C: Pan-Cancer Co-Measured PTMs (Crosstalk #572) CPTAC Pan-Cancer Acetylome & Ubiquitylome PDC UUIDs 5551b44b-a3c8-4a6c-86e5-48c9e66f74e2 (UMich Acetylome), a6d10e8e-531e-4a38-91fd-964ad968619c (Broad Acetylome), 1951cd08-7a85-4fbd-ba7b-753fe62a6926 (Ubiquitylome) 5–6 cohorts (BRCA, GBM, LSCC, LUAD, UCEC, MB) overlapping Pan-Cancer phosphoproteome patients 199.12 MB (UMich Acetyl), 13.31 MB (Broad Acetyl), 33.82 MB (CDAP Ubiquityl) Same harmonized patient IDs and clinical/mutation tables as Tier 1A/1B
Tier 2A: LUAD Discovery & Confirmatory Cross-Cohort Transfer CPTAC LUAD Discovery (PDC000149, Gillette et al., Cell 2020) & Confirmatory (PDC000490) PDC000149 (UUID 8855d018-bdc6-4ea7-a96a-f0c7db78f90c) & PDC000490 (UUID 4d2f2a6c-b2e9-4564-ae73-1868ab5b2f37) Discovery: 47,956 phosphosites $\times$ 225 columns (115 cases, 217 aliquots: 111 Tumor + 101 NAT). Confirmatory: 36,502 phosphosites $\times$ 245 aliquots (131 independent cases) Discovery TSVs: 199.04 MB (69.06 MB site TSV; 231.67 GB raw). Confirmatory TSVs: 225.03 MB (66.88 MB site TSV; 258.72 GB raw) EGFR, KRAS, STK11, KEAP1, TP53 mutations, ALK/ROS1/RET fusions, Tumor vs. NAT, Hot/Cold immune labels, stage, smoking, survival
Tier 2B: External Multi-Ethnic LUAD Cohorts (ICPC Taiwan) Academia Sinica Taiwan LUAD100 (PDC000220, Chen et al., Cell 2020) & ICPC-B (PDC000564) PDC000220 (UUID ebf7bd08-225c-48c5-8552-caad105168dd) & PDC000564 (UUID 73444b92-19ca-47ad-be93-d4f69818aaf0) Phase 1 (PDC000220): 30,866 phosphosites $\times$ 170 aliquots (103 clinical patients, 86 phospho cases). Phase 2 (PDC000564): 53,640 phosphosites $\times$ 251 aliquots (127 cases) PDC000220 TSVs: 200.76 MB (66.87 MB site TSV). PDC000564 TSVs: 364.89 MB (111.09 MB site TSV) Academia Sinica LUAD-100_Clinical.xlsx (0.03 MB, 103 patients $\times$ 22 cols): EGFR_mut (~85% prevalence, L858R/Exon19del), TP53_mut, Stage, APOBEC_group, smoking, differentiation
Tier 2C: Naturally Sparse Clinical LUAD Cohort (NCI APOLLO) APOLLO LUAD Phosphoproteome (PDC000435 TiO2 & PDC000436 FeNTA; Soltis et al., Cell Rep Med 2022) PDC000435 (UUID 3f670550-63c2-4b21-9467-ef9e46932f2b) & PDC000436 (UUID e8f465ac-92f5-4441-a16a-14e612cd4e25); PRIDE PXD036025 101 LUAD cases; sparse single-site rollup (949 FeNTA + 105 TiO2 sites; peptide TSVs 2.34 MB + 0.37 MB) Processed TSVs: 4.57 MB combined (123.40 GB raw per fraction) APOLLO_LUAD_clinical_treatment_manifest.tsv (0.01 MB, 84 rows $\times$ 32 treatment/regimen/outcome columns) + PDC clinical survival & stage
Tier 3A: Longitudinal Anti-HER2 Core-Needle Biopsy Trial Microscaled Proteogenomics DP1 ERBB2+ Breast Cancer Trial (Satpathy et al., Nat Commun 2020; NCT02267290) PDC000326 (Phosphosite UUID 4bd2ea8b-c8de-446a-9f99-f82ad0dc7721; Exp Design UUID de817ae3-0bb2-4979-8bec-6e710c615d8e) 13,148 phosphosites $\times$ 38 OCT-embedded 14G core-needle biopsy aliquots (14–17 patients, paired Baseline vs. 48–72h post-trastuzumab+pertuzumab On-Treatment biopsies) + 10 PDX aliquots (PDC000328/330) Processed TSVs: 27.27 MB (9.60 MB site TSV; 21.53 GB total study) Longitudinal timepoint (Pre-treatment vs. 48–72h On-treatment), pathological complete response (pCR vs. non-pCR), true vs. pseudo-ERBB2 amplification
Tier 3B: Phase III HER2+ Randomized Neoadjuvant Trial Biopsies Alliance CALGB 40601 HER2+ Trial (Jaehnig et al., Cell Rep Med 2025; NCT00770809) PDC000583 (Phosphosite UUID 1b7f9bac-ce73-4720-b2bf-221f385d6f05; Clinical Table UUID e8abe3fc-2559-4679-bf05-16ee4d15b937) 8,258 phosphosites $\times$ 80 frozen core-needle biopsy aliquots from 59 Stage II/III HER2+ patients Processed TSVs: 27.14 MB (7.77 MB site TSV; 27.77 GB total study) sample_table_CRMpaper.xlsx (0.02 MB, 75 biopsies $\times$ 13 cols): Treatment arm Tx (paclitaxel + trastuzumab TH, lapatinib TL, or dual THL), pCR, ER_status, PR_status, PAM50_updated, ERBB2_CNA_logratio, ERBB2_protein, Herceptest_IHC
Tier 3C: Longitudinal TNBC Neoadjuvant Trial Core Biopsies WashU Phase II TNBC Trial (Anurag et al., Cancer Discov 2022; NCT02124902 / NCT02547987) PDC000409 (Site UUID 12b4500b-7296-4e10-a4d9-cd4043e61fee; Metadata UUID 6a1bf6f2-1161-4b42-971c-1550008bd3b3) & PDC000580 PDC000409: 10,715 phosphosites $\times$ 74 core-needle biopsies across 58 TNBC patients (Baseline vs. Cycle 1 Day 3 C1D3 on docetaxel+carboplatin). PDC000580: 60 baseline TNBC cases PDC000409 TSVs: 34.93 MB (9.74 MB site TSV). PDC000580 TSVs: 85.35 MB (26.51 MB site TSV) TNBC_only_metadata_v6.3_forPDC_shankha_27thJune2022.xlsx (0.03 MB, 72 biopsies $\times$ 20 cols): CollectionEventlabelSimplified (Baseline vs. C1D3), pCR_response, PAM50.Mar2020, average.TC (tumor cellularity)
Tier 3D: Targeted Kinase Inhibitor Perturbation & Resistance Series Beat AML Clinical & FLT3-TKI Resistance (PDC000478, PDC000318–320) + decryptM Dose-Response Atlas (Zecha et al., Science 2023) PDC000478 (UUID c04bbb8b-ea8b-45bc-b372-90da3c4f469c), PDC000318/319/320; PRIDE PXD037285 / ProteomicsDB decryptM Beat AML: 25,680 phosphosites $\times$ 210 patient samples + 93 longitudinal Gilteritinib/Quizartinib resistance & time-course aliquots. decryptM: 31 cancer drugs (10 doses, time-resolved) Beat AML TSVs: 109.57 MB (39.67 MB site TSV). decryptM curves: via ProteomicsDB API / Zenodo 10.5281/zenodo.7647751 Ex vivo drug sensitivity (IC50/AUC for FLT3, MEK, BCL2, pan-TKIs), FLT3-ITD/NPM1/TP53 status, early vs. late TKI resistance states, dose-response $\text{EC}_{50}$

Detailed Download Instructions & Programmatic Access

1. Direct Programmatic Download from NCI PDC (Zero-Auth REST + GraphQL API)

Every processed matrix and clinical metadata file listed above can be downloaded directly in Python without authentication using its file_id UUID via the PDC signed-URL endpoint:

import json
import urllib.request

def download_pdc_file(file_id: str, out_path: str) -> None:
    """Resolve a fresh CloudFront signed URL from NCI PDC and download the file."""
    api_url = f"https://pdc.cancer.gov/pdcapi/file/signedURLFromUuid/{file_id}"
    req = urllib.request.Request(
        api_url,
        headers={
            "User-Agent": "Mozilla/5.0",
            "Referer": "https://pdc.cancer.gov/pdc/cptac-pancancer",
        },
    )
    with urllib.request.urlopen(req, timeout=15) as resp:
        signed_url = json.loads(resp.read().decode())["data"]
    urllib.request.urlretrieve(signed_url, out_path)
  • CPTAC Pan-Cancer Portal Page: https://pdc.cancer.gov/pdc/cptac-pancancer
  • NCI PDC GraphQL Endpoint (for study manifests & clinical tables): https://pdc.cancer.gov/graphql (queries: getAllUIPancancerFiles, filesPerStudy(pdc_study_id: "..."), clinicalPerStudy(pdc_study_id: "...")).

2. Python cptac Package (PayneLab/cptac)

For per-cancer cohort loading across brca, ccrcc, coad, gbm, hnscc, lscc, luad, ov, pdac, and ucec (supporting bcm, umich, broad, washu, mssm, and harmonized sources):

  • Repository: https://github.com/PayneLab/cptac (pip install cptac)
  • Key API calls: cohort.get_phosphoproteomics(source="bcm" | "umich" | "broad"), cohort.get_proteomics(), cohort.get_clinical(source="mssm"), cohort.get_somatic_mutation(source="harmonized").

3. Companion Analysis & Reference Mapping Repositories

  • Broad Getz Lab Pan-Cancer PTM Repository: https://github.com/getzlab/CPTAC_PanCan_PTM_2023 (includes data/reference/RefSeq.20180629_Human_ucsc_hg38_cpdbnr_mito_264contams_553smORFs.fasta and var_map_full_v4.tsv for mapping RefSeq/GENCODE v34 site IDs onto 11-mer flanking motifs).

Verification of Clinical Annotations

We inspected the clinical metadata archives directly to ensure all necessary clinical endpoints and covariates are present and non-empty:

  1. Clinical_meta_data_v1.zip (0.74 MB, 10 files: BRCA_meta.txt, ccRCC_meta.txt, COAD_meta.txt, GBM_meta.txt, HGSC_meta.txt, HNSCC_meta.txt, LSCC_meta.txt, LUAD_meta.txt, PDAC_meta.txt, UCEC_meta.txt):
    • 1,081 patients (123 BRCA, 104 ccRCC, 111 COAD, 100 GBM, 94 HGSC, 109 HNSCC, 109 LSCC, 111 LUAD, 106 PDAC, 114 UCEC).
    • Demographics, Pathology & Survival: idx (Patient ID), Tumor, Normal, Age, Sex, BMI, Tumor_Size_cm, Histologic_Grade, Tumor_necrosis, Path_Stage_pT, Path_Stage_pN, Stage, Tobacco_smoking_history, OS_days, OS_event, PFS_days, PFS_event.
    • Somatic Driver Mutations & Genomics: Binary mutation status per cohort (e.g., EGFR_mutation, KRAS_mutation, STK11_mutation, KEAP1_mutation, TP53_mutation in LUAD; VHL_mutation, PBRM1_mutation, KDM5C_mutation in ccRCC; KRAS_mutation, TP53_mutation in PDAC; MAP3K1_mutation, GATA3_mutation, PIK3CA_mutation in BRCA) + full MAF in PanCan_Union_Maf_Broad_WashU_v1.1.maf (531.32 MB uncompressed) + WES_purity, WES_ploidy, TMB, MSI_H, POLE, CNV_index_CIN_index, and COSMIC Mutation_signature_SBS*.
    • Immune Deconvolution & Microenvironment: ESTIMATE_StromalScore, ESTIMATE_ImmuneScore, ESTIMATE_ESTIMATEScore, 22 CIBERSORT_* immune cell fractions, and 38 xCell_* immune/stroma scores.
    • Pre-Computed Baseline Pathway & PTM-SEA Scores (for head-to-head benchmarking): 14 PROGENy_* pathway scores, 50 HALLMARK_* ssGSEA scores, and PTM_SEA_KINASE-PSP_* / PTM_SEA_PATH-NP_* / PTM_SEA_PERT-* scores.
  2. External LUAD Validation (PDC000220 Academia Sinica LUAD-100_Clinical.xlsx & PDC000435 APOLLO_LUAD_clinical_treatment_manifest.tsv):
    • Academia Sinica LUAD-100_Clinical.xlsx (Table S1A_clinical_103patient, 103 rows $\times$ 22 cols): ID, Proteome_Batch, Age, Gender, SK_SmokingStatus, Stage, StagingN, PathologyDifferentiation, VisceralPleuralStatus, EGFR_mut (including exact allele L858R, Exon19del, Other, WT), TP53_mut, APOBEC_enrichment, APOBEC_group, Carcinogen_group.
    • APOLLO_LUAD_clinical_treatment_manifest.tsv (84 rows $\times$ 32 cols): case_submitter_id, treatment_intent_type, treatment_type, therapeutic_agents, regimen_or_line_of_therapy, number_of_cycles, reason_treatment_ended.
  3. Clinical Trial Core-Needle Biopsies (PDC000326, PDC000583, PDC000409):
    • PDC000583 (sample_table_CRMpaper.xlsx, TableS2B sample data, 75 biopsies): Core_id, Patient_ID, ER_status, PR_status, Tx (TH, TL, THL), pCR, PAM50_updated, ERBB2_CNA_GISTIC_threshold, ERBB2_CNA_logratio, ERBB2_RNA, ERBB2_protein, Herceptest_IHC.
    • PDC000409 (TNBC_only_metadata_v6.3_forPDC_shankha_27thJune2022.xlsx, 72 biopsies): PPI (Patient ID), CollectionEventlabelSimplified (Baseline vs. Cycle1Day3), pCR_response, PAM50.Mar2020, average.TC, TC.binary.
    • PDC000326 (S051_Microscaled_Proteogenomics_Satpathy2020_exp_design.xlsx, 35 clinical cores): Core ID, Patient ID (BCN xxxx), Plex No, TMT channel, paired pre-treatment vs. 48–72h on-treatment core IDs (P suffix) and clinical pCR annotations from Satpathy et al., Nat Commun 2020.

Proposed Approach & What Will Be Tested

Phase 1: Reference Model Training & DDMC.project_samples Implementation

  1. Harmonized 11-mer Sequence Mapping Pipeline (ddmc/datasets.py):
    • Add dataset loaders for the CPTAC Pan-Cancer harmonized tables, Taiwan LUAD (PDC000220/PDC000564), and clinical trial biopsy cohorts (PDC000326, PDC000583, PDC000409) that map RefSeq/GENCODE v34 protein/site identifiers (or flanking 15-mers in phosphosite.tmt*.tsv) to DDMC's length-11 sequence motif representation (preserving flanking PTM tokens s, t, y, k where multi-site/acetylome data are available).
  2. Out-of-Sample Biopsy Projection (DDMC.project_samples(p_signal_new, return_var=False) in ddmc/clustering.py):
    • Given a fitted reference DDMC model with $K$ clusters, motif models $\mathcal{M}c$, reference peptide responsibilities $\gamma{ic}$ (for $i \in \mathcal{I}_{\text{ref}}$), and per-cluster/per-peptide residual variance estimates $\sigma_c^2$:
    • Partition query biopsy peptides $\mathcal{I}_{\text{new}}$ into:
      • Overlapping reference peptides ($\mathcal{I}{\text{overlap}} = \mathcal{I}{\text{new}} \cap \mathcal{I}_{\text{ref}}$): Use the reference posterior responsibilities $\gamma_{ic}$ (or optionally update $\gamma_{ic}$ when projecting a multi-sample cohort) and reference precision weights $\tau_{ic}$.
      • Novel biopsy peptides ($\mathcal{I}{\text{novel}} = \mathcal{I}{\text{new}} \setminus \mathcal{I}_{\text{ref}}$): Fold in via sequence motifs alone using self.fold_in(p_signal_novel, cols=[]), yielding sequence-prior responsibilities $\gamma^{\text{seq}}_{ic} \propto \pi_c \exp(w \cdot \log P(S_i \mid \mathcal{M}_c))$.
    • Estimate the query biopsy's module activity vector $\hat{\mu}{\text{new}, c}$ and posterior uncertainty $\operatorname{Var}(\hat{\mu}{\text{new}, c})$ for each sample $j$ and cluster $c$ via responsibility-weighted maximum likelihood over observed sites $i \in \operatorname{obs}(j)$:
      $$\hat{\mu}{\text{new}, cj} = \frac{\sum{i \in \operatorname{obs}(j)} \gamma_{ic} \tau_{ic} X^{\text{new}}{ij}}{\sum{i \in \operatorname{obs}(j)} \gamma_{ic} \tau_{ic}}, \qquad \operatorname{Var}(\hat{\mu}{\text{new}, cj}) = \left(\sum{i \in \operatorname{obs}(j)} \gamma_{ic} \tau_{ic}\right)^{-1}$$

Phase 2: Specific Hypotheses & Quantitative Tests

Test 1: Cross-Cohort Clinical Biomarker & Driver Mutation Transfer (Discovery $\to$ Confirmatory $\to$ ICPC Taiwan $\to$ APOLLO)

  • Hypothesis: Kinase modules learned on the CPTAC LUAD Discovery cohort (PDC000149, $n=111$ tumors) or the 1,119-tumor CPTAC Pan-Cancer atlas transfer zero-shot (via project_samples without refitting cluster assignments) to independent cohorts (PDC000490 Confirmatory LUAD, PDC000220/PDC000564 Taiwan LUAD, and PDC000435/436 APOLLO LUAD) and remain predictive of somatic driver genotypes (EGFR, KRAS, STK11, KEAP1, TP53, ALK), Tumor vs. NAT status, and immune infiltration (Hot_Cold / ESTIMATE_ImmuneScore) under severe biopsy dropout.
  • Experimental Protocol:
    1. Fit DDMC ($K \in {30, 50, 100}$, seq_weight selected via BCV Implement bi-cross-validation for joint DDMC hyperparameter selection (#552) #569) on PDC000149 (and on the leave-one-cohort-out Pan-Cancer reference).
    2. Train L1-penalized LogisticRegressionCV classifiers on reference module centers $\mu_{cj}$ (extending ddmc/figures/figureMS3.py).
    3. Project held-out patients from PDC000490, PDC000220, and PDC000435/436 using DDMC.project_samples(), both at native completeness and under simulated biopsy dropout from 0% to 85% via simulate_dropout(mechanism="mnar") and "block".
  • Quantitative Metrics:
    • Out-of-sample AUROC and AUPRC vs. dropout fraction (0%–85%), benchmarked against:
      1. Single-site phosphopeptide LASSO biomarkers,
      2. Sample-wise KSEA / PTM-SEA enrichment scores (using the 404-kinase PSPL atlas and the pre-computed PTM_SEA_* columns in Clinical_meta_data_v1.zip),
      3. Expression-only GMM/PCA/NMF projection (seq_weight = 0).

Test 2: Longitudinal Pharmacodynamic Rewiring & Pathological Complete Response (pCR) in Clinical Trial Core-Needle Biopsies

  • Hypothesis: Projecting sparse 14G core-needle biopsies (PDC000326 DP1 anti-HER2 trial, PDC000583 CALGB 40601 Phase III anti-HER2 trial, and PDC000409 WashU Phase II TNBC trial) onto the CPTAC BRCA / Pan-Cancer DDMC reference atlas accurately quantifies on-target kinase suppression (ERBB2/MAPK/PI3K-AKT-mTOR modules at 48–72h on-treatment) and separates pCR responders from non-responders (No pCR), including identifying "pseudo-HER2+" tumors and compensatory bypass modules (EMT/Notch/WNT/FAK/AURKB).
  • Experimental Protocol:
    1. Fit a reference DDMC model on the CPTAC BRCA (PDC000121) + Pan-Cancer reference atlas (110,274 sites) and annotate each cluster's upstream kinases via predict_upstream_kinases() across the 404-kinase PSPL atlas.
    2. Project the core-needle biopsies from PDC000326 ($13,148$ sites $\times 38$ aliquots), PDC000583 ($8,258$ sites $\times 80$ aliquots), and PDC000409 ($10,715$ sites $\times 74$ aliquots) onto the reference modules via project_samples().
    3. Compute paired within-patient pharmacodynamic shifts $\Delta \hat{\mu}{c} = \hat{\mu}{c, \text{On-Tx}} - \hat{\mu}_{c, \text{Baseline}}$ in PDC000326 (48–72h post-trastuzumab/pertuzumab) and PDC000409 (C1D3 vs. Baseline), and baseline module activities in PDC000583 across TH, TL, and THL arms.
  • Quantitative Metrics:
    • Target & bypass effect size: Paired Wilcoxon signed-rank $p$-value and Cohen's $d$ for suppression of ERBB2/EGFR/MAPK/AKT modules at 48–72h in pCR responders vs. non-responders.
    • pCR prediction accuracy: Cross-validated AUROC / AUPRC and Mann-Whitney U $p$-value for predicting pCR in PDC000583 ($n=59$) and PDC000409 ($n=58$) from projected DDMC module activities vs. single ERBB2/phosphosite levels and PTM-SEA.
    • Biopsy sample-efficiency curve: Minimum number of detected phosphosites per biopsy required for $\hat{\mu}_{\text{new}}$ to maintain $\ge 0.90$ cosine similarity with the full-biopsy module profile, quantifying the gain from folding in novel biopsy peptides via seq_weight > 0.

Test 3: Pan-Cancer PTM Crosstalk & Scalability Benchmark

  • Hypothesis: Incorporating co-measured multiply phosphorylated peptides and proximal lysine acetylation (Acetylome_UMich_GENCODE34_v1.zip / Acetylome_Broad_Institute_harmonized_v1.tsv.gz) using the expanded PTM alphabet (Support modified flanking residues (pS/pT/pY, acK) for PTM crosstalk (#565) #572) improves out-of-sample BCV predictive log-likelihood and separates canonical phospho-primed kinase modules (GSK3A/B, CSNK1/2, PLK1, CDK1/2) across the 1,119-patient Pan-Cancer atlas while fitting in minutes and projecting a new biopsy in $< 1\text{ second}$.
  • Quantitative Metrics:
    • BCV held-out MSE and predictive log-likelihood (score_held_out) on multiply modified vs. singly modified peptides across the Pan-Cancer atlas.
    • Wall-clock runtime and peak memory to fit DDMC on the 110,274 × 1,119 Pan-Cancer matrix ($K = 50\text{–}100$) and latency per projected biopsy (< 1 s).

Activity

  1. added this to the CPTAC study with DDMC milestone on Nov 14, 2026
  2. self-assigned this
    on Nov 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

enhancementNew feature or request

Type

No type

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions