Skip to content

Mechanisms of resistance to immune checkpoint inhibitors

A living literature review of ICI resistance across PD-1, PD-L1, CTLA-4, LAG-3, TIGIT, TIM-3, and combination regimens. Regenerated on each scheduled update — see changelog.

established replicated, field consensus  ·  contested landmark weakened under replication  ·  suspected mechanistic, not yet clinical  ·  emerging new, worth tracking

Supporting evidence: human patient, trial, or primary tissue  ·  mouse in-vivo mouse work  ·  other animal non-mouse, non-human in vivo  ·  in vitro cell-line / primary-cell / organoid

What has held up

  • established human Pre-existing CD8+ T-cell infiltrate at the invasive margin predicts pembrolizumab response in melanoma (Tumeh 2014)1. The inflamed / excluded / desert taxonomy has been validated pan-tumor and is the backbone of most subsequent biomarker frameworks (Chen & Mellman 2017)2.
  • established human TMB predicts anti-PD-1 benefit in NSCLC (Rizvi 2015)3 and anti-CTLA-4 benefit in melanoma (Snyder 2014)4. The association survives, but see "contradictions" for the pan-tumor threshold debate.
  • established human dMMR/MSI-H is the single most reliable ICI biomarker across histologies (Le 2017)5; basis of the first tissue-agnostic FDA approval. Confirmed in KEYNOTE-177 for 1L dMMR mCRC6 — but not uniformly predictive (see contradictions).
  • established human mouse JAK1/JAK2 and B2M LOF cause acquired anti-PD-1 resistance in melanoma (Zaretsky 2016)7; rare (<5%) but biochemically definitive. JAK1/2 LOF also drives a subset of primary resistance (Shin 2017)8. PTPN2 loss in tumor cells is the inverse — a sensitizer that amplifies IFN-γ response (Manguso 2017)9, now in clinical development.
  • established human KRAS-mutant NSCLC with STK11/LKB1 co-mutation is primary ICI-refractory despite TMB-intermediate/high status (Skoulidis 2018)10. Now used in NSCLC trial stratification.
  • established human PD-1 + CTLA-4 combination rescues a meaningful fraction of single-agent PD-1 non-responders in melanoma, with the largest relative benefit in PD-L1-negative tumors (CheckMate 067)11 — indicating CTLA-4 and PD-1 address partly-complementary resistance mechanisms. ~55% grade 3/4 TRAE cost.
  • established human PD-L1 IHC (TPS ≥50%) prospectively identifies the NSCLC sub-population for whom single-agent PD-1 blockade is sufficient (KEYNOTE-024)12. The PD-L1-low population defines the monotherapy resistance group; they now receive chemo-IO by default.
  • established human mouse Response depends on a stem-like Tcf1+PD-1+ CD8+ progenitor pool, not reinvigoration of terminally exhausted effectors (Siddiqui 2019, Sade-Feldman 2018, Miller 2019)1314. Explains why a heavily PD-1+ infiltrate is not enough if it lacks the stem-like subset.
  • established mouse TOX is the master transcription factor of the exhausted chromatin state (Alfei 2019 and concurrent work)15. Epigenetic fixation of exhaustion is a durable barrier to reinvigoration — and TOX ablation impairs both exhaustion and persistence, so it is not a simple therapeutic brake.
  • established human mouse TGF-β signaling in peritumoral fibroblasts drives T-cell exclusion (Mariathasan 2018 in urothelial16; Tauriello 2018 in MSS CRC GEMM17). But clinical translation has failed — see contradictions.
  • established human Canonical resistance taxonomy — primary / adaptive / acquired (Sharma 2017)18 — remains the organizing framework.

Human-study evidence

Study N Feature Effect 95% CI / p Method
Tumeh 2014 n=46 (metastatic melanoma on pembrolizumab; discovery n=46 + validation n=15) pre-existing CD8⁺ T cells at the invasive tumor margin response prediction (AUC) AUC ~0.9 in discovery cohort — IHC + multiplex IF + TCR-seq
Rizvi 2015 NSCLC discovery + validation cohorts nonsynonymous tumor mutational burden DCB / PFS enrichment cutoff ~178 nonsynonymous mutations separates DCB from no-benefit — WES
Snyder 2014 n=64 (discovery n=25 + validation n=39 melanoma on anti-CTLA-4) somatic mutational load (and shared-neoantigen tetrapeptide signature) clinical benefit association mutational load associated with benefit P=0.01 WES
Le 2017 n=86 (phase 2 expansion across 12 tumor types, MMR-deficient tumors) MMR deficiency ORR 53% (CR 21%) — MMR IHC / MSI testing + pembro phase 2
André 2020 n=307 (KEYNOTE-177 phase 3, treatment-naive dMMR/MSI-H mCRC) 1L pembrolizumab vs 5-FU-based chemotherapy PFS HR mPFS 16.5 vs 8.2 mo; HR 0.60 95% CI 0.45–0.80, p=0.0002 phase 3 RCT
Zaretsky 2016 n=4 (melanoma patients with late relapse on pembrolizumab) acquired JAK1/JAK2 LOF or B2M truncation acquired resistance mechanism 3 of 4 relapse tumors harbored a candidate lesion — paired WES + functional IFN-γ / MHC-I assays
Shin 2016 n=39 (23 melanoma + 16 MMR-deficient CRC non-responders) biallelic JAK1/JAK2 LOF primary-resistance prevalence ~4% melanoma, ~6% MMR-d CRC non-responders — WES + cell-line IFN-γ / ISG induction assays
Skoulidis 2018 SU2C + CheckMate-057 KRAS-mutant NSCLC ICI cohorts STK11/LKB1 co-mutation (KL subgroup) ORR KL 7.4% vs KP 35.7% (SU2C); 0% vs 57.1% (CheckMate-057) P<0.001 targeted/WES + clinical correlation
Larkin 2015 n=945 (CheckMate-067 phase 3, untreated metastatic melanoma) nivo+ipi vs nivo vs ipi monotherapy PFS HR (combo vs ipi) mPFS 11.5 vs 6.9 vs 2.9 mo; HR 0.42 combo vs ipi (HR 0.57 nivo vs ipi) — phase 3 RCT
Reck 2016 n=305 (KEYNOTE-024 phase 3, 1L advanced NSCLC, PD-L1 TPS ≥50%, EGFR/ALK-wt) PD-L1 TPS ≥50% (pembrolizumab vs platinum doublet) PFS HR mPFS 10.3 vs 6.0 mo; HR 0.50 95% CI 0.37–0.68, p<0.001 PD-L1 IHC 22C3
Sade-Feldman 2018 n=48 (checkpoint-blockade-treated melanoma; 16,291 immune cells sequenced) TCF7⁺ stem-like vs dysfunctional CD8⁺ T-cell state response direction TCF7⁺ CD8 frequency associated with response (validation cohort) — scRNA-seq + IF validation
Mariathasan 2018 n=298 (IMvigor210 atezolizumab-treated metastatic urothelial carcinoma) fibroblast TGF-β response signature (F-TBRS) response direction non-responders enriched for F-TBRS and T-cell-excluded phenotype — bulk RNA-seq

Where the field has contradicted itself (the surprises)

Known initial findings that later weakened under replication or didn't translate. Useful to avoid over-weighting any single-study narrative.

  • contested human HLA-I LOH is a selected immune-escape event (~40% of NSCLCs, McGranahan 201719) but is not deterministic for ICI response. Subsequent pan-cancer ICI cohorts show some HLA-LOH patients still respond. What matters is which allele is lost (loss of the allele presenting a dominant neoantigen), tumor type, and whether residual heterozygous HLA alleles can still present key neoantigens. So: "this patient has HLA LOH" does not imply "this patient will fail anti-PD-1."
  • contested human β2M loss does not uniformly predict resistance either. Documented case-level resistance (Zaretsky 20167) and rare B2M-low/null tumors that still respond suggest NK-mediated or non-classical HLA-dependent mechanisms can partially compensate. The acquired-resistance rate attributable to β2M/JAK is a small fraction of overall failures.
  • contested human mouse Microbiome responder signatures do not replicate cleanly across geographies. Gopalakrishnan 2018 (Faecalibacterium/Ruminococcaceae)20, Routy 2018 (Akkermansia)21, Matson 2018 (Bifidobacterium longum), and Sivan 201522 / Vétizou 201523 mouse work each nominated different beneficial taxa. Meta-analyses (Lee 2022, McCulloch 2022 Nat Med) found no single taxon reliably predicts anti-PD-1 response across cohorts. For Akkermansia specifically, Derosa 2022 Nat Med showed a non-monotonic relationship — very high levels were worse than intermediate. Bacteroidales sign-flipped: beneficial in Vétizou's anti-CTLA-4 mouse work, associated with non-response in Gopalakrishnan's anti-PD-1 cohort. The antibiotic-harm signal is the only microbiome finding that has replicated consistently.
  • contested human FMT rescues anti-PD-1-refractory melanoma (Baruch 202124: 3/10; Davar 202125: 6/15 benefit including ORR ~20%). This was unexpected because refractoriness had been widely assumed to reflect tumor-intrinsic escape (B2M/JAK loss), not a reversible host-ecosystem state. The convergent independent replication at Sheba and NCI/Pittsburgh using different donors strengthens the signal.
  • contested human TGF-β is a biologically robust exclusion driver that has clinically disappointed. Bintrafusp alfa (TGFβRII trap / anti-PD-L1 bifunctional) failed pivotal trials — INTR@PID lung 037 discontinued 2021, biliary tract halted. Galunisertib + durvalumab, NIS793 combinations in MSS CRC — modest or negative. A reproducibility gap between GEMM and clinic; likely reflects redundancy, dose-limiting cardiotoxicity of TGF-β antagonism, or context-specific effects.
  • contested human bTMB failed prospective validation. Gandara 201826 retrospectively showed bTMB-high enriches atezolizumab benefit in POPLAR/OAK; prospective B-F1RST did not confirm a first-line predictive effect. bTMB is not broadly adopted.
  • contested human PBRM1 LOF as ICI sensitizer in ccRCC has not replicated. Miao 201827 reported enrichment in responders in two cohorts; subsequent CheckMate-009/010/025 (Braun 2020) and IMmotion analyses did not confirm.
  • contested human IPRES transcriptomic signature has underperformed as an independent predictor. Hugo 201628 nominated an EMT/wound-healing / angiogenesis program in innately resistant melanoma. Validation has been mixed — IPRES overlaps heavily with TGF-β and stromal signatures and adds little above CD8, PD-L1, TMB, and pan-fibroblast measures.
  • contested human dMMR is not a universal ICI predictor even in 1L mCRC. KEYNOTE-1776 PFS was biphasic — pembrolizumab underperformed chemotherapy in the first ~6 months (~29% primary progression), and the final OS analysis did not reach statistical significance (HR ~0.74, crossover confounded).
  • contested mouse WNT/β-catenin activation as clinical exclusion driver is weaker than Spranger 201529 mouse biology suggested. Subsequent clinical cohorts show context-dependent correlations; WNT inhibition + ICI limited by toxicity.
  • contested human The specific "shared neoantigen tetrapeptide signature" from Snyder 2014 did not replicate in larger cohorts; the TMB association survives but the signature itself is widely considered a small-sample artifact.

Human-study evidence

Study N Feature Effect 95% CI / p Method
McGranahan 2017 n=100 (TRACERx 100 NSCLC) allele-specific HLA-I LOH frequency / selection ~40% HLA-I LOH; subclonal, enriched at metastatic sites — LOHHLA (allele-specific WES)
Zaretsky 2016 n=4 (melanoma patients with late relapse on pembrolizumab) acquired JAK1/JAK2 LOF or B2M truncation acquired resistance mechanism 3 of 4 relapse tumors harbored a candidate lesion — paired WES + functional IFN-γ / MHC-I assays
Gopalakrishnan 2017 n=112 (melanoma on anti-PD-1; shotgun-metagenomics subset n=43 (30 R, 13 NR)) gut alpha diversity / Faecalibacterium enrichment response association higher diversity & Faecalibacterium in responders P<0.01 16S / shotgun metagenomics + responder→GF-mouse FMT
Routy 2017 n=249 (NSCLC, RCC, urothelial on anti-PD-1/PD-L1) peri-ICI antibiotic exposure / A. muciniphila abundance OS HR antibiotics → shorter PFS/OS (NSCLC mOS ~8 vs ~20 mo; HR ~3.5) — clinical-record audit + shotgun metagenomics
Baruch 2020 n=10 (phase I FMT in anti-PD-1-refractory metastatic melanoma) FMT from anti-PD-1-responder donors + anti-PD-1 reinduction ORR 3/10 (1 CR, 2 PR) — phase I FMT trial
Davar 2021 n=15 (single-arm phase II FMT in anti-PD-1-refractory melanoma) responder-FMT + pembrolizumab clinical benefit / ORR 6/15 benefit; ORR ~20% — phase II single-arm trial
Gandara 2018 POPLAR (test) + OAK (validation) NSCLC atezolizumab blood TMB ≥16 mut/Mb (FoundationACT ctDNA) PFS HR (atezo vs docetaxel) HR ~0.65 at bTMB ≥16 in OAK — ctDNA (FoundationACT)
Miao 2018 n=98 (discovery n=35 + validation n=63 ccRCC on anti-PD-1 ± anti-CTLA-4) PBRM1 LOF clinical-benefit enrichment enriched among responders P=0.012 (discovery); P=0.0071 (validation) WES
Hugo 2016 n=38 (pretreatment melanoma biopsies on anti-PD-1) IPRES transcriptional signature (EMT / ECM / angiogenesis / wound-healing) response direction IPRES-high enriched among non-responders — bulk RNA-seq + WES
André 2020 n=307 (KEYNOTE-177 phase 3, treatment-naive dMMR/MSI-H mCRC) 1L pembrolizumab vs 5-FU-based chemotherapy PFS HR mPFS 16.5 vs 8.2 mo; HR 0.60 95% CI 0.45–0.80, p=0.0002 phase 3 RCT

Suspected but unconfirmed (mechanistically coherent, not yet clinically validated)

  • suspected human in vitro PD-1 partially protects clonally expanding T cells from restimulation-induced cell death30. In vitro human primary T cells only. If replicated in vivo, would complicate the "release the brakes" framing and explain some non-durability of response.
  • suspected mouse AARS1-mediated PD-L1 K280 lactylation stabilizes PD-L1 against HUWE1 ubiquitination31. Counterintuitively, exogenous lactate enhances anti-PD-L1 efficacy preclinically — this paradox must be reconciled before clinical translation.
  • suspected mouse HILPDA-driven lipogenesis palmitoylates PD-L1 at Cys27232. Fenretinide (TRIM21 engager) degrades HILPDA and restores anti-PD-1 efficacy in breast cancer models. Single tumor type; preclinical only; but fenretinide has existing clinical exposure.
  • suspected human deltaHED (germline + somatic HLA-I evolutionary divergence) predicts worse PD-1 outcomes despite higher TMB/neoantigen load across three cohorts33. Inverts a naïve HED prediction; needs prospective single-agent PD-1 validation.
  • suspected mouse Hypoalbuminemia causally drives ICI resistance via macrophage arginine biosynthesis impairment (LLC mice)34. Would recast a routine prognostic lab as a reversible driver; dietary arginine rescue would need clinical study.
  • suspected human Hyperprogression on anti-PD-1 (Champiat 2017, Kato 2017 class of reports) remains contested; definitions vary and natural-history confounding is substantial. Not listed as established.

Human-study evidence

Study N Feature Effect 95% CI / p Method
Xu 2026 n=729 (3 cohorts: 164 NPC (POLARIS-02), 88 melanoma, 477 ESCC (JUPITER-06)) deltaHED (germline + somatic HLA-I evolutionary divergence) PFS/OS direction worse on PD-1 blockade despite higher TMB/neoantigen load — HLA typing + WES (deltaHED metric)

New directions worth watching (emerging this period)

  • emerging human ATOMIC phase 3 (NEJM 2026): adjuvant atezolizumab + mFOLFOX6 in stage III dMMR colon cancer, HR 0.5035. 3-year DFS 86.3% vs 76.2%. Extends ICI benefit into adjuvant dMMR. OS not yet mature. A new large population will now experience prolonged adjuvant ICI exposure, creating a fresh substrate for late-resistance biology.
  • emerging mouse KLRG1 nominated as a novel inhibitory checkpoint in anti-PD-1-resistant melanoma36; novel anti-human KLRG1 mAb reduces tumor progression in humanized KI mice via combined CD8, NK, and γδ-T effects. Distinct from PD-1/CTLA-4/LAG-3/TIM-3.
  • emerging human Anti-TIM-3 (TQB2618) + anti-PD-1 penpulimab achieves 52% ORR in PD-1-pretreated classical Hodgkin lymphoma (n=21, phase Ib)37. Salvage signal in a setting where re-engaging checkpoint biology was not expected to help.
  • emerging mouse TROP2-claudin-7 tight junctions functionally exclude T cells from TNBC38 — reframes TROP2 from ADC target to barrier-mechanism target; mechanistic rationale for TROP2 ADC + anti-PD-1 combinations beyond cytotoxic payload delivery.
  • emerging mouse PKMYT1 inhibition (clinical-grade RP-6306) activates cGAS-STING in castration-resistant prostate cancer39. Converts cold CRPC to anti-PD-L1-responsive in preclinical models; prostate cancer has been stubbornly ICI-refractory.
  • emerging mouse IFN-γ → IRF1 → AGPAT3 axis sensitizes tumors to ferroptosis40. Extends IFN-γ consequences beyond antigen presentation and apoptosis to a lipidomic vulnerability.
  • emerging human 8-gene k-TSP + mucinous composite biomarker identifies a 15% dMMR/MSI-H mCRC subgroup with extreme benefit (HR 0.10) from anti-CTLA-4 addition to anti-PD-141. Retrospective (n=25 in positive subgroup); prospective validation called for. If it holds, would resolve a decade-long open question.
  • emerging human LOAd703 (CD40L/4-1BBL oncolytic adenovirus) + atezolizumab in anti-PD-1-refractory melanoma restores ICI-responsive immune signatures in 24 patients42. Biomarker-level evidence of myeloid-compartment rescue in a refractory population.
  • emerging human Single-cell spatial profiling defines six niches stratifying neoadjuvant cSCC response better than PD-L1 IHC43. Niche-composition biomarkers may eventually outperform single-marker IHC.
  • emerging human hMENA TGF-β-driven CAF signature validated against OAK phase III44 — a clinically anchored exclusion biomarker despite the TGF-β clinical translation gap on the drug side.

Human-study evidence

Study N Feature Effect 95% CI / p Method
Sinicrope 2026 n=712 (phase 3 ATOMIC (NCT02912559), stage III dMMR colon cancer) atezolizumab + mFOLFOX6 vs mFOLFOX6 alone 3-yr DFS / HR 86.3% vs 76.2%; HR 0.50 95% CI 0.35–0.73, p<0.001 phase 3 RCT
Hong 2026 n=21 (phase Ib (NCT05400876), PD-1-pretreated relapsed/refractory cHL) anti-TIM-3 TQB2618 + anti-PD-1 penpulimab ORR 52% (1 CR, 10 PR); grade ≥3 TRAE 24% — phase Ib clinical trial
Ambrosini 2026 n=163 (pooled across 2 independent dMMR/MSI-H mCRC cohorts) Cluster A + mucinous histology (15% biomarker-positive subgroup) 24-mo PFS / HR (Combo vs Mono) 72.2% vs 13.8%; HR 0.10 95% CI 0.02–0.39, p<0.001 8-gene k-TSP classifier + histology
Grauers 2026 n=24 (single-arm phase I/II, anti-PD-1-refractory stage IV melanoma) intratumoral LOAd703 (CD40L/4-1BBL oncolytic adenovirus) + atezolizumab immune-signature biomarker increased DC markers, T-cell infiltration, EM CD8⁺; decreased circulating Tregs — multi-parameter flow + TME transcriptomics
Lee 2026 n=27 (3 cSCC cohorts incl. 2 phase II trials) six spatial tissue niches (high-APC / B-plasma / inflammatory-keratinocyte vs. proliferative / low-APC-myeloid / fibroblast-EMT) pathologic response prediction niche profiling outperformed PD-L1 IHC — single-cell spatial transcriptomics
Melchionna 2026 TCGA NSCLC + SU2C + OAK (NCT02008227) hMENA⁺ TGF-β-driven CAF 9-gene signature prognosis / ICT resistance direction signature-high → worse prognosis, ICT resistance — 9-gene RNA signature

Practical questions & quick answers

  • Is PD-L1 IHC still clinically useful? Yes, for 1L NSCLC selection (TPS ≥50% standard for pembro monotherapy). Less useful elsewhere; increasingly replaced by tumor-type-specific algorithms.
  • Does dMMR status guarantee benefit? No. KEYNOTE-1776 PFS was biphasic with ~29% primary progression. dMMR is necessary-but-not-sufficient in mCRC.
  • Does peri-ICI antibiotic exposure matter? Yes — the antibiotic-harm signal is the most robustly replicated microbiome observation. Specific probiotic or taxonomy-guided interventions are not clinically ready.
  • Can anti-PD-1-refractory patients be rescued? Emerging positive signals: FMT from a responder donor (~20–30% response in small trials), TIM-3 combination in PD-1-pretreated cHL (52% ORR n=21), LOAd703 oncolytic in melanoma (biomarker-level). None are standard of care yet.
  • Is STK11/KEAP1 testing informative in KRAS-mutant NSCLC? Yes — it guides prognosis on ICI and supports clinical trial stratification, though no FDA-level companion biomarker yet.
  • Is there a prospective biomarker beyond PD-L1 / TMB / dMMR / STK11 ready for clinical use? Not yet. HLA LOH, HED/deltaHED, microbiome, IPRES, PBRM1, bTMB — all research-only or have failed prospective validation.
  • What about hyperprogression? Definitions vary; the literature is contested. Natural history and pseudoprogression confound the phenomenon. Not actionable today.

See per-mechanism deep dives in the navigation. The papers appendix lists every paper ingested (N>1,000) with links.

  1. Tumeh 2014 Nature — pre-existing CD8 infiltrate predicts PD-1 response. Link. ↩

  2. Chen & Mellman 2017 Nature — cancer-immunity set-point framework. Link. ↩

  3. Rizvi 2015 Science — TMB in NSCLC under pembrolizumab. Link. ↩

  4. Snyder 2014 NEJM — TMB under anti-CTLA-4 in melanoma. Link. ↩

  5. Le 2017 Science — pembrolizumab in MMR-deficient tumors across 12 histologies. Link. ↩

  6. André 2020 NEJM — KEYNOTE-177 pembrolizumab 1L dMMR mCRC. Link. ↩↩↩

  7. Zaretsky 2016 NEJM — JAK1/2 and B2M LOF in acquired anti-PD-1 resistance. Link. ↩↩

  8. Shin 2017 Cancer Discov — JAK1/2 LOF in primary resistance. Link. ↩

  9. Manguso 2017 Nature — Ptpn2 as ICI sensitizer. Link. ↩

  10. Skoulidis 2018 Cancer Discov — STK11/LKB1 in KRAS-mutant NSCLC ICI. Link. ↩

  11. Larkin 2015 NEJM — CheckMate 067 nivo+ipi vs mono in melanoma. Link. ↩

  12. Reck 2016 NEJM — KEYNOTE-024 pembrolizumab 1L NSCLC PD-L1 ≥50%. Link. ↩

  13. Siddiqui 2019 Immunity — stem-like Tcf1+PD-1+ CD8 subset. Link. ↩

  14. Sade-Feldman 2018 Cell — CD8 T cell states and melanoma ICI response. Link. ↩

  15. Alfei 2019 Nature — TOX reinforces exhausted CD8 phenotype. Link. ↩

  16. Mariathasan 2018 Nature — TGF-β excludes T cells in urothelial ICI. Link. ↩

  17. Tauriello 2018 Nature — TGF-β in MSS CRC metastasis GEMM. Link. ↩

  18. Sharma 2017 Cell — Primary/Adaptive/Acquired resistance review. Link. ↩

  19. McGranahan 2017 Cell — allele-specific HLA-I LOH in NSCLC. Link. ↩

  20. Gopalakrishnan 2018 Science — gut microbiome and anti-PD-1 in melanoma. Link. ↩

  21. Routy 2018 Science — antibiotics, Akkermansia, and PD-1 therapy. Link. ↩

  22. Sivan 2015 Science — Bifidobacterium enhances anti-PD-L1 in mice. Link. ↩

  23. Vétizou 2015 Science — microbiota and anti-CTLA-4 in mice. Link. ↩

  24. Baruch 2021 Science — FMT rescue in anti-PD-1-refractory melanoma (phase I). Link. ↩

  25. Davar 2021 Science — FMT + pembrolizumab in anti-PD-1-refractory melanoma (phase II). Link. ↩

  26. Gandara 2018 Nat Med — blood TMB in NSCLC atezolizumab (retrospective). Link. ↩

  27. Miao 2018 Science — PBRM1 and anti-PD-1 response in ccRCC. Link. ↩

  28. Hugo 2016 Cell — IPRES signature in anti-PD-1 resistant melanoma. Link. ↩

  29. Spranger 2015 Nature — WNT/β-catenin and T-cell exclusion. Link. ↩

  30. RICD protection by PD-1 on expanding T cells. Link. ↩

  31. AARS1 lactylation of PD-L1 K280. Link. ↩

  32. HILPDA-KLF5-palmitoylation of PD-L1. Link. ↩

  33. deltaHED metric across three ICI cohorts. Link. ↩

  34. Hypoalbuminemia and macrophage arginine in ICI resistance. Link. ↩

  35. ATOMIC phase 3 — adjuvant atezolizumab + FOLFOX in stage III dMMR. Link. ↩

  36. KLRG1 as novel checkpoint. Link. ↩

  37. Anti-TIM-3 + penpulimab in PD-1-pretreated cHL. Link. ↩

  38. TROP2-claudin-7 tight junction barrier in TNBC. Link. ↩

  39. PKMYT1 inhibition activating cGAS-STING in CRPC. Link. ↩

  40. IFN-γ–IRF1–AGPAT3–ferroptosis axis. Link. ↩

  41. 8-gene k-TSP + mucinous biomarker for anti-CTLA-4 addition in dMMR/MSI-H mCRC. Link. ↩

  42. LOAd703 + atezolizumab in anti-PD-1-refractory melanoma. Link. ↩

  43. Spatial niches stratifying neoadjuvant cSCC response. Link. ↩

  44. hMENA CAF signature validated in OAK phase III. Link. ↩