AI-driven multimodal algorithm predicts immunotherapy and targeted therapy outcomes in clear cell renal cell carcinoma

Authors

Stupichev D

Miheecheva N

Postovalova E

Lyu Y

Ramachandran A

Galkin I

Khegai G

Perevoshchikova K

Love A

Menshikova S

Tarasov A

Svekolkin V

Bruttan M

Varlamova A

Kriukov K

Ataullakhanov R

Fowler N

Cheng E

Bagaev A

Hsieh JJ

Doi

10.1016/j.xcrm.2025.102299

PMID: 22138691 · DOI: 10.1016/j.xcrm.2025.102299 · Journal: Cell Reports Medicine (2025)

TL;DR

Stupichev et al. compiled the largest harmonized clear cell renal cell carcinoma (ccRCC) transcriptomic database (n = 3,621 samples from 14 cohorts) and identified five harmonized immune tumor microenvironment (HiTME) subtypes using density-based clustering of functional gene expression signatures. They developed gradient-boosting machine learning models integrating genomic, transcriptomic, and TME features to predict ICI and TKI therapy response, validated these on independent cohorts including a Washington University cohort (WU-RCC, n = 193), and proposed an integrated decision-tree model that classifies 56% of patients as ICI/ICI-combo-preferred, 41% as TKI-preferred, and 3% as non-responsive to either regimen.

Cohort & data

  • Meta-cohort: 3,621 ccRCC samples from 14 public cohorts (17 databases total, 4,187 samples before filtering), including TCGA-KIRC (n = 491), IMmotion150 (n = 163), IMmotion151 (n = 784), JAVELIN Renal 101 (n = 701), CheckMate 009/010/025, COMPARZ (n = 341), E-MTAB-3267 (n = 53), CPTAC, and Sato et al.
  • Independent validation cohort: WU-RCC (n = 193) from Washington University, with DNA-seq and RNA-seq (n = 157 with bulk RNA-seq).
  • Spatial proteomics validation: Multiplex immunofluorescence (MxIF) on a subset of WU-RCC patients (n = 34) with a 19-marker ccRCC antibody panel.
  • Single-cell RNA-seq: 259,441 cells from three public ccRCC cohorts for FGES validation.
  • Cancer type: clear cell renal cell carcinoma (CCRCC).
  • Assays: bulk RNA-seq, scRNA-seq, MxIF, DNA-seq (NGS).

Key findings

  • Five HiTME subtypes were defined: IE (immune-enriched, nonfibrotic), IE/M (immune-enriched, myeloid immunosuppressive), F (fibrotic-myeloid immunosuppressive), V (highly vascularized), and D (immune desert).
  • Fibrotic subtypes IE/M and F had the worst overall survival with standard of care in TCGA-KIRC and Sato et al. cohorts.
  • IE and IE/M subtypes had increased proportions of ICI responders (p = 4 x 10^-5), while IE and V subtypes had the highest TKI responder proportions (p = 0.001).
  • Subtype F had the largest proportion of ICI non-responders.
  • The ICI response model (gradient boosting, trained on n = 214 CR vs. PD samples) achieved ROC-AUC = 0.77 on training and ROC-AUC = 0.78 on JAVELIN validation, outperforming all published single-biomarker signatures.
  • The TKI response model (trained on n = 240 samples) achieved ROC-AUC = 0.74 on validation (n = 822), outperforming JAVELIN Angio, IMmotion150 Angio, proliferation, and macrophage signatures.
  • ICI responder scores had consistently significant hazard ratios across cohorts (IMmotion151 HR = -0.52, p = 0.003; JAVELIN HR = -0.55, p = 0.01).
  • Patients with ICI-high/TKI-low scores had significant PFS benefit from atezolizumab+bevacizumab vs. sunitinib (p = 0.0003).
  • A 3% subset of patients (ICI-low/TKI-low) was identified as unlikely to respond to either regimen, characterized by strong myogenesis signatures and protumor TME features.
  • 15 of 16 patients with MTOR-activating mutations were resistant to ICI regimens.
  • No patients with antigen presentation gene mutations exhibited complete response to ICI or ICI+TKI.

Genes & alterations

  • BAP1: Mutations associated with TKI non-response; more prevalent in TKI-low score group. Mutations enriched in fibrotic IE/M and F HiTME subtypes.
  • SETD2: Mutations associated with TKI non-response in WU-RCC.
  • MTOR: Activating mutations strongly associated with ICI and ICI+TKI non-response (15/16 patients resistant); enriched in ICI-low score group.
  • PBRM1: Mutations enriched in TKI-high score group; more prevalent in immune-enriched IE and IE/M HiTMEs.
  • KDM5C: Mutational frequency varied across ICI and TKI score groups.
  • NF2: Mutations more prevalent in F HiTME subtype.
  • CD274 (PDL1): Expression on immune cells (not tumor cells) was significantly higher in ICI responders; used as a feature in ICI response model. CXCL8 higher in IE/M subtype.
  • PDCD1 (PD1): Expression used as feature in ICI response model.

Clinical implications

  • The integrated decision-tree model classifies ccRCC patients into ICI/ICI-combo-preferred (56%), TKI-preferred (41%), or non-responsive (3%) categories, potentially guiding first-line therapy selection.
  • Patients classified as ICI/ICI-combo-preferred had significantly longer PFS with atezolizumab+bevacizumab vs. sunitinib (IMmotion151, p = 0.00005) and with avelumab+axitinib vs. sunitinib (JAVELIN, p = 0.000007).
  • TKI-preferred patients trended toward longer PFS with sunitinib alone (median 14 months) compared to bevacizumab+atezolizumab (median 10 months, p = 0.06).
  • MTOR-activating mutations and antigen presentation gene mutations may serve as molecular exclusion criteria for ICI therapy (termed “molecular ICI NRs”).
  • The 3% ICI-low/TKI-low subgroup may benefit from anti-proliferative and anti-fibrotic agents rather than standard AA and ICI therapies.

Limitations & open questions

  • The model cannot stratify patients for ICI monotherapy vs. anti-CTLA4+anti-PD1 combinations; data for lenvatinib+pembrolizumab and nivolumab+cabozantinib combinations are lacking.
  • Small sample sizes in some cohorts may limit statistical power and ability to demonstrate significance.
  • No prospective clinical validation datasets were available.
  • The harmonization approach (normalizing gene signatures across datasets) masks biological features specific to individual datasets, so HiTME subtypes may be most effective when comparing patients from cohorts with similar characteristics.
  • Spatial imaging and scRNA-seq were performed on limited patient subsets (n = 34 and n = 22,456 cells, respectively).
  • Training and test cohorts may be skewed toward certain immunotherapy and TKI types.
  • Open question: whether anti-proliferative or anti-fibrotic agents can improve outcomes for the 3% non-responsive subgroup.

Citations from this paper used in the wiki

  • “By unifying transcriptomic data from 14 cohorts (total n = 3,621), we present harmonized immune tumor microenvironment (HiTME) ccRCC subtypes validated with spatial proteomics.” (Summary)
  • “15 out of 16 patients with mTOR-activating mutations were resistant to ICI regimens.” (Results, Robust genomic and transcriptomic relationships with immunotherapy response)
  • “Our decision-tree model classified 56% of patients as ICI/ICI combo-preferred, 41% as TKI-preferred, and 3% as non-responsive to either.” (Results, Clinical applications)
  • “Patients with ICI-high/TKI-low scores exhibited a significant survival benefit from atezolizumab+bevacizumab compared to single-agent sunitinib (p = 0.0003).” (Results, Clinical applications)
  • “ICI/ICI combo-preferred patients had increased PFS when treated with atezolizumab plus bevacizumab compared to sunitinib (IMmotion151, p = 0.00005), and with avelumab plus axitinib compared to sunitinib (JAVELIN, p = 0.000007).” (Results, Clinical applications)

This page was processed by paper-compiler on 2026-05-06.