ccRCC Target Atlas

Summary of the findings

Written by the orchestrating agent from the five agent reports. Every sentence cites evidence records and passed the automatic checks.

CA9 is the candidate on which omics, literature, external knowledge, single-cell and spatial data agree most. It is strongly up in tumour RNA CA9 · rna · log2fc = 6.312, replicates in TCGA-KIRC, survives purity extrapolation, and its tumour transcripts come from the cancer cells themselves rather than from stroma CA9 · share_of_tumor_transcripts = 0.976. Girentuximab-based antibody and imaging agents are already in clinical development and CA9 PET has proof-of-concept literature. It is best read as a surface antigen and imaging biomarker rather than a dependency: DepMap shows no ccRCC dependency, expression is heterogeneous across tumour cells, and stomach staining is a safety flag for any conjugate. ^{[1,2,3,4,5,6,7,8,9]}

A second tier of tumour-up, druggable surface or metabolic candidates follows CA9. In the target profile SCARB1, CD70, SLC16A3, NNMT and ENPP3 hold their places under random re-weighting. CD70 and ENPP3 already have clinical antibody or conjugate programmes, and CD70 has the best conjugate suitability of the controls. SLC16A3 (MCT4) and NNMT replicate at the protein level in the independent confirmatory CPTAC cohort, and NNMT is tumour-cell derived in two single-cell datasets and linked to worse survival in TCGA. SCARB1 tops the target profile on expression, but its cell attribution does not replicate and few tumour cells express it in the weakest donors, so its rank overstates its case. None of these five shows a selective DepMap dependency, so their case rests on expression restriction and tractability, not on functional need. ^{[10,11,12,13,14,15,16,17,18,19,20,21,22,23]}

The strongest biomarker calls are different genes. PSAT1, PCK1 and SLC6A19 are the most robust members of the biomarker profile. PSAT1 and PCK1 are lower in tumour than in normal kidney, and both protein losses replicate in the confirmatory cohort, so they are loss-of-identity or metabolic markers, not drug targets. PSAT1 additionally tracks the aggressive ccB state. ALDOB is a literature-supported prognostic loss marker, and COXFA4L2 (NDUFA4L2) is a HIF-driven tumour marker with dense preclinical literature but a mitochondrial location that makes it a tissue biomarker rather than a target. ^{[24,25,26,27,28,29,30]}

The convergence ranking rewards literature and existing drugs, so it can never put a little-studied gene on top; the discovery screen is where those genes appear, and two tumour-cell genes from it are identified and prioritised for experimental validation. COL23A1, a transmembrane collagen, is made mainly by RCC cells COL23A1 · source_share = 0.881, keeps a large increase when extrapolated to a pure tumour COL23A1 · rna_lfc_at_purity1 = 4.88 and replicates in TCGA-KIRC, with no clinical drug and no vital-organ protein enrichment. What is missing: SURFY and CSPA do not confirm surface display, the AlphaFold model is low confidence, there is no prognostic signal, no selective dependency and no confirmatory-cohort data; immunohistochemistry on tumour and normal tissue arrays plus surface proteomics would test it. RAB42, a small GTPase, is also made mostly by RCC cells RAB42 · source_share = 0.627 and persists in a pure tumour RAB42 · rna_lfc_at_purity1 = 3.189, but it is unlikely to be antibody-accessible, so knockdown or CRISPR experiments in ccRCC models are the natural next test. The best stroma candidates (OLFML2A, JAG2) are purity-robust but come from fibroblasts and endothelium, and the immune-track genes fade when extrapolated to a pure tumour. Novel here means few ccRCC papers and no clinical drug, not that nothing was published: this literature corpus did not cover these genes, so earlier preclinical work cannot be excluded, and nothing here shows that targeting them would help a patient. ^{[31,32,33,34,35,36,37,38,39,40,41,42]}

Across all layers, the one candidate with a selective cell-line dependency is the established target EPAS1 (HIF-2 alpha, belzutifan), and the established VEGFR target KDR is endothelial rather than tumour-intrinsic. The top-ranked new candidates are therefore antigen, imaging or biomarker leads rather than vulnerabilities, and no gene separates the main treatment arms in the trial data after correction. ^{[43,44,45,46]}

Prioritised candidates

The orchestrator picked these from the numeric ranking and the verified agent reports. It may depart from the ranking and then says why. The tags show the role (target, biomarker or both), the rank, the cell type that makes the gene, the best pocket score and any drug that already reached the clinic.

A guided tour of the tabs

Each tab answers one question. Click a card to open it. Every figure, table and counter has a round ? button that explains the axes, how to read it and what we see.

Evidence coverage and proposed use

Coverage counts measured layers, not independent votes. Missing data, measured results that do not meet a criterion, and contradictory observations are kept separate. These descriptive rules do not alter ranking weights. “Favorable” applies only to the named test; it is not proof of therapeutic benefit.

The convergence ranking

Each gene gets one score built from seven kinds of evidence: how much it changes in tumour (RNA, protein and phospho), whether a copy-number change really moves its product, what the literature says, genetics, how druggable the protein looks, whether cancer cell lines need it, and whether a second cohort (TCGA-KIRC) sees the same change. Every kind is scaled from 0 to 1 and multiplied by a weight (omics 2.0, cnv 0.5, literature 1.0, genetics 2.0, tractability 0.5, dependency 1.0, replication 1.0). The weights live in config.yaml and were set before the results were known. 19,161 genes are scored.

Top 25 ?

Integration

Integrating the layers gives three groups. Targets with convergent support are CA9, CD70, NNMT, SLC16A3 and ENPP3: tumour-up, mostly tumour-intrinsic, replicated, and tractable or already in trials, with no dependency. Biomarkers are PSAT1, PCK1, ALDOB, COXFA4L2 and UMOD, mostly loss-of-identity markers. Mechanisms are VHL, EPAS1, PBRM1 and HILPDA. The main disagreements are SCARB1, which ranks first on data but weakens on cell attribution and conjugate suitability; COXFA4L2, which has strong literature but no dependency; and CD70, whose literature survival link does not replicate in TCGA. ^{[2,49,14,13,11,24,25,27,29,105,53,106,17,58]}

Tumour versus healthy kidney (CPTAC discovery cohort)

CPTAC tumour-versus-normal data separate two groups. HIF-driven genes rise in tumour: CD70 is up at protein CD70 · protein · log2fc = 2.421 and RNA CD70 · rna · log2fc = 4.464, replicates in TCGA and shows no vital-organ protein enrichment CD70 · vital_protein_ts_max = 0.34, and the same pattern holds for CA9, NNMT, SLC16A3, ENPP3 and SCARB1. ^{[47,48,49,50]}

Proximal-tubule genes fall in tumour. ALDOB is lost at protein ALDOB · protein · log2fc = -2.273 and RNA ALDOB · rna · log2fc = -7.028, and low expression is associated with worse survival in TCGA; such down-regulated genes (with PSAT1, PCK1 and UMOD) are biomarkers of lost tubule identity, not targets. ^[51,52,28]

VHL explains the tumour-up programme rather than joining it: it is the recurrent driver VHL · freq = 0.483 with frequent 3p copy-number loss, and its own expression change is small, so it is a mechanism, not a DE-based biomarker. ^[53,54]

Association with tumour grade ?

This shows how closely each gene follows tumour grade (G1 to G4) across the discovery tumours, for RNA and for protein. The CPTAC clinical file has grade, stage and tumour size but no survival, so grade is the only outcome-like measure available here. The survival analysis uses TCGA-KIRC instead.

Omics agent findings (verified)

CA9 target and biomarker CA9 is the top convergence gene: tumour-up in RNA CA9 · rna · log2fc = 6.312 and protein CA9 · protein · log2fc = 2.269, replicated in TCGA-KIRC and robust to purity extrapolation. Safety caveat: the vital-organ proteome shows stomach enrichment CA9 · vital_protein_ts_max = 5.19. ^{[1,108,2,3,109]}

CD70 target CD70 is a membrane target absent in normal kidney and up in tumour at protein CD70 · protein · log2fc = 2.421 and RNA CD70 · rna · log2fc = 4.464, with TCGA replication. It is not enriched in vital organs CD70 · vital_protein_ts_max = 0.34, and antibody and ADC programmes exist. ^{[47,48,49,50]}

ENPP3 target ENPP3 is a surface enzyme up in tumour protein ENPP3 · protein · log2fc = 1.544 and RNA ENPP3 · rna · log2fc = 3.895, with no vital-organ enrichment ENPP3 · vital_protein_ts_max = 0.0. An ADC (AGS-16C3F) reached phase 2. ^[99,110,111]

NNMT target and biomarker NNMT is tumour-intrinsic and up in protein NNMT · protein · log2fc = 2.709 and RNA NNMT · rna · log2fc = 4.67. Higher expression is associated with worse survival in TCGA-KIRC, which makes it a prognostic biomarker with a metabolic-target rationale. ^[98,112,16]

SCARB1 target SCARB1 (SR-B1) ranks first in the target profile: up in protein SCARB1 · protein · log2fc = 1.837 and RNA SCARB1 · rna · log2fc = 2.974, replicated in TCGA. It is a membrane cholesterol-uptake receptor consistent with the lipid-laden ccRCC phenotype, but it is also expressed in normal tissues, so selectivity is not established. ^[96,113,92]

SLC16A3 target SLC16A3 (MCT4, lactate exporter) is up in protein SLC16A3 · protein · log2fc = 2.279 and RNA SLC16A3 · rna · log2fc = 3.435, replicated in TCGA, consistent with HIF-driven glycolysis. Cis copy-number effect is positive. ^[97,114,90]

COXFA4L2 biomarker COXFA4L2 (NDUFA4L2) is a strong HIF-axis tumour marker, up in protein COXFA4L2 · protein · log2fc = 3.768 and RNA COXFA4L2 · rna · log2fc = 6.752, replicated in TCGA. It is a mitochondrial protein, so it is better suited as a tissue biomarker than as a target. ^[30,115,107]

ALDOB biomarker ALDOB is lost with proximal-tubule identity: down in protein ALDOB · protein · log2fc = -2.273 and RNA ALDOB · rna · log2fc = -7.028. Low expression is associated with worse survival in TCGA. It is a diagnostic and prognostic marker, not a target. ^[51,52,28]

PSAT1 biomarker PSAT1 tops the biomarker profile: down in tumour at protein PSAT1 · protein · log2fc = -1.963 and RNA PSAT1 · rna · log2fc = -3.867. Within TCGA-KIRC, higher expression predicts worse survival, a paradoxical pattern. Use it as a prognostic marker with caution. ^{[100,116,117]}

VHL mechanism VHL is the disease driver, with recurrent somatic mutation across the combined CPTAC and TCGA cohort VHL · freq = 0.483, plus frequent 3p copy-number loss with a positive cis dosage effect. Loss of VHL explains HIF activation and the up-regulated CA9, NDUFA4L2, SLC16A3 and NNMT programme. Its own RNA and protein changes are small, so it is not a DE-based biomarker. ^[53,54]

Which cells make each candidate?

The cell-type calls come from single-cell data of Li et al. 2022: 178,817 cells from 12 donors, 146,558 of them from tumour tissue. "Share of tumour transcripts" is the part of a gene's tumour RNA that each cell type makes. It separates a gene the tumour cells make themselves from one whose signal comes from infiltrating cells. "Tumour cell vs proximal tubule" compares the cancer cells with the normal cell they most likely came from, donor by donor (6 donors had both). The last block on this tab repeats the attribution in a second, independent dataset.

Single-cell attribution confirms that CA9 and MET are tumour-intrinsic. MET transcripts sit mainly in RCC cells MET · share_of_tumor_transcripts = 0.935, though only a minority of RCC cells express it MET · pct_expressing = 35.2, so expression is heterogeneous. ^[70,71]

Several bulk-omics or clinical genes are not tumour-intrinsic. KDR is endothelial, CD74 sits in macrophages and dendritic cells, CXCR4 in infiltrating lymphocytes, FCGR3A is almost entirely immune-derived, and most EPAS1 mRNA in this dataset is endothelial rather than tumour-cell, so its bulk level is not a tumour-cell readout. CD70 is partly an immune-compartment signal as well. Across a second single-cell dataset NNMT and PCK1 keep their tumour-cell attribution, whereas SCARB1 switches between myeloid and tumour cells, which weakens it as a tumour-specific target. ^{[45,72,73,74,75,76,15,77,17]}

Cell type of each candidate ?

Does a second dataset agree?

Zhang et al. 2021 profiled 7 ccRCC donors (20,509 cells) and 5 normal kidneys, with 4 donors in both. The two studies name cell types differently, so both were folded into seven shared groups. For genes in the discovery screen, the dominant source in Zhang matches the track from Li in 183 of 313 immune genes, 33 of 34 stroma genes, 59 of 63 tumor genes. Across 17,919 genes the tumour-cell share correlates between studies at rho 0.618, and the tumour-versus-tubule fold change at rho 0.421. Immune genes agree least: the two studies capture very different amounts of immune cells.

Where in the tissue, and what is next to it?

The tissue data are 38,836 Visium spots from 16 sections and 8 donors. Each spot mixes a few cells, so its cell makeup is estimated by fitting it to the study's own single-cell profiles. The candidate genes were left out of that fit, which makes them a fair test of it: CA9 follows the tumour-cell share (rho 0.277), KDR follows endothelium (0.205), CD74 follows macrophages (0.304) and PTPRC follows T cells (0.123). The correlations are positive but modest. The Tissue viewer tab shows the same data on the real H&E sections.

In Visium sections, high-CA9 spots are enriched for RCC-cell neighbours CA9 · log2_enrichment = 0.381, with only marginal macrophage enrichment, and tumour-cell pseudobulk shows CA9 far above proximal tubule CA9 · log2fc = 7.739. This supports CA9 as a tumour-cell niche marker. Spatial analysis was run for CA9 only. ^[78,79,80]

Use the menu above each map to pick a cell share or a candidate gene. These are the largest core and interface sections.

For each gene, we take the top tenth of spots and look at the ring of six spots around each one (about 100 micrometres). We compare the cell makeup of those rings with rings around random spots from the same sections, as a ratio, so common cell types get no head start. A permutation test with 1,000 shuffles gives each p-value, adjusted over all genes and cell types.

Tissue context of each candidate ?

Cell-context agent findings (verified)

CA9 target and biomarker CA9 transcripts in tumour are almost entirely tumour-cell derived, with RCC cells holding nearly all of them CA9 · share_of_tumor_transcripts = 0.976, while only about half of RCC cells express it CA9 · pct_expressing = 54.7; the bulk signal is tumour-intrinsic and CA9 is the top convergence-ranked gene, but it is already clinically drugged (girentuximab) and expression is heterogeneous across tumour cells. ^[4,5]

CA9 biomarker Paired tumour-cell versus proximal-tubule pseudobulk shows CA9 strongly up in tumour cells CA9 · log2fc = 7.739, and in Visium the high-CA9 spots are enriched for RCC neighbours CA9 · log2_enrichment = 0.381 (macrophages show only a marginal enrichment), consistent with a tumour-cell niche marker. ^[80,78,79]

CA9 target The AlphaFold CA9 model has mean pLDDT CA9 · mean_plddt = 76.56 and its top detected cavity has volume CA9 · volume_a3 = 126.8 A3 with one nearby accessible lysine; detected cavities recover half of experimental ligand-contact residues CA9 · recall = 0.5, which only validates cavity detection and says nothing about degron or ligase compatibility. ^[81,82,83]

MET target MET is mainly tumour-cell derived in tumour tissue, with RCC cells holding most transcripts MET · share_of_tumor_transcripts = 0.935, but only a minority of RCC cells express it MET · pct_expressing = 35.2, and endothelium carries a smaller per-cell share, so MET is a tumour-cell target with heterogeneous expression. ^[70,71]

KDR microenvironment KDR is an endothelial gene, not a tumour-cell target: endothelial cells hold almost all transcripts KDR · share_of_tumor_transcripts = 0.951, and a substantial fraction of endothelial cells express it KDR · pct_expressing = 40.5; anti-VEGFR therapy acts on the vasculature. ^[45,176]

EPAS1 mechanism In this scRNA-seq dataset EPAS1 (HIF2A) transcripts sit mostly outside RCC cells: endothelial cells hold a large share EPAS1 · share_of_tumor_transcripts = 0.365 against a small share for RCC cells EPAS1 · share_of_tumor_transcripts = 0.051, so mRNA level in bulk is not a tumour-cell readout (HIF2A activity is better read through targets such as CA9). ^[75,177,178]

CD74 microenvironment CD74 is an antigen-presenting-cell and lymphocyte transcript, not tumour-cell: macrophages hold 0.436 of tumour transcripts CD74 · share_of_tumor_transcripts = 0.436, with dendritic cells showing the highest per-cell level and nearly all of them expressing it, so it is a microenvironment marker. ^[72,179]

CXCR4 microenvironment CXCR4 is carried by infiltrating lymphocytes: CD8 T cells hold 0.563 of tumour transcripts CXCR4 · share_of_tumor_transcripts = 0.563 and CD4 T, B and pDC cells also express it; it is an immune-axis target, not tumour-intrinsic. ^[73,180]

COL23A1 discovery Tumour track: COL23A1 is made mainly by RCC cells COL23A1 · source_share = 0.881, stays up when extrapolated to a pure tumour COL23A1 · rna_lfc_at_purity1 = 4.88, replicates in TCGA-KIRC, has no clinical drug, a surface call of surface (UniProt cell membrane; SURFY and CSPA do not support it), a few ccRCC papers, and no protein enrichment in vital-organ proteome; the AlphaFold model is low confidence (see model record) with one tiny cavity, no dependency (DepMap not selective) and no prognostic signal, so it remains a hypothesis. ^{[31,32,33,36,34,35]}

RAB42 discovery Tumour track: RAB42 is expressed mostly by RCC cells RAB42 · source_share = 0.627, persists in a pure tumour RAB42 · rna_lfc_at_purity1 = 3.189 and replicates in TCGA, with very few ccRCC papers and no clinical drug; its location is unknown for antibody access (UniProt membrane, not in SURFY or CSPA), survival is not associated and no dependency is shown, so it is an unvalidated hypothesis. ^{[37,38,39,84]}

OLFML2A discovery Stroma track: OLFML2A is a fibroblast-derived secreted protein OLFML2A · source_share = 0.647, purity-robust OLFML2A · rna_lfc_at_purity1 = 1.527 and TCGA-replicated with no clinical drug, but it is a microenvironment target with notable colon expression, no survival association and no functional evidence. ^{[40,85,86,87]}

JAG2 discovery Stroma track: JAG2 is a surface Notch ligand made mainly by endothelium JAG2 · source_share = 0.487, replicated in TCGA and purity-robust, with no clinical drug; it is less novel than other candidates (several ccRCC papers), has modest fold change, heart-muscle RNA expression, and no dependency evidence. ^{[41,181,182,183]}

MILR1 discovery Immune track: MILR1 is a surface macrophage (mast-cell receptor family) gene with no ccRCC papers, with most tumour transcripts in macrophages MILR1 · source_share = 0.841; its bulk increase largely vanishes in a pure tumour MILR1 · rna_lfc_at_purity1 = 0.414, flagged infiltration-driven, although TCGA shows adverse prognosis; it is a macrophage-state marker hypothesis, not a tumour-cell target. ^{[42,184,88,89]}

GAL3ST4 discovery Immune track: GAL3ST4 is almost exclusively macrophage GAL3ST4 · source_share = 0.972 and is TCGA-replicated with adverse prognosis, but it is an internal Golgi enzyme (not antibody-accessible) and its fold change extrapolated to a pure tumour is small GAL3ST4 · rna_lfc_at_purity1 = 0.401, so at most a macrophage biomarker. ^{[185,186,187,188]}

Structure and pockets

MGD: experimental degrons and ternary references

No new MGD or ternary prediction is made. Reference matching is annotation, not a classifier. Recruitment, ubiquitination and degradation are separate observations. The structural-context heuristic below does not establish any of them.

Ligase context in the ccRCC data

These are bulk and cell-type expression/alteration observations, not proof of an active ligase or intracellular colocalization. VHL recruitment requires functional VHL in the selected model; its availability is not assumed. No candidate-specific ternary model is generated without a specified target, ligase and compound.

Sources: gene_universe.parquet, layer_mutations.parquet and sc_celltype_expression.parquet. Frequencies refer to their respective cohorts, not the same individuals. Reference controls above come from the published experimental systems, not from a new experiment in this project.

Cavities are detected on AlphaFold monomers after removing residues below pLDDT 50. The structural-context heuristic combines coordinate confidence, cavity geometry, nearby accessible lysines and an exposed non-helical segment. The latter is not a degron annotation.

The AlphaFold CA9 model is of moderate confidence CA9 · mean_plddt = 76.56 and its largest detected cavity is small CA9 · volume_a3 = 126.8, with one nearby accessible lysine; cavity detection recovers half of experimental ligand-contact residues CA9 · recall = 0.5. These are descriptive heuristics on a predicted monomer: they validate cavity detection only and are not evidence of a degron, ligase compatibility or a degrader. No curated experimental degron or ternary-complex reference was cited for any candidate in this run. ^[81,82,83]

Descriptive references: known glue-degrader substrates ?

These known substrates provide structural reference points only. Their monomer scores are not a test of degradation prediction: this comparison has no matched experimental negatives and does not model the compound or ligase interface. The heuristic summarizes pocket features; neither a high nor a low value establishes susceptibility to molecular-glue degradation.

Best pocket per protein ?

Target discovery: promising, little studied, no drug yet

The screen starts from 410 genes that are up in tumour in both CPTAC RNA (log2 fold change at least 1.0) and protein (at least 0.5) and that single-cell data tie mostly to one cell type (at least 40 % of the gene's tumour transcripts). Each gene is then checked for how much ccRCC literature it has (Europe PMC, counting only aliases that belong to that gene), whether a clinical drug or potent compounds already exist (Open Targets, ChEMBL), where the protein sits (UniProt), how much it is expressed in vital organs (Human Protein Atlas), and whether it follows grade, replicates in TCGA-KIRC, survives correction for tumour purity, tracks survival or is a selective cell-line dependency. A gene is novel when it has no clinical-stage drug and at most 15 ccRCC papers (196 genes). It is a first-in-class opportunity when it is also tractable by a small molecule or an antibody, stays at or below 50 nTPM in every vital organ, replicates in TCGA-KIRC and, for tumour and stroma genes, keeps its up-regulation at purity one (55 genes: 49 immune, 3 tumor, 3 stroma). These are hypotheses ranked by data, not validated targets.

In the tumour track, COL23A1 is the lead. It is a transmembrane collagen made mainly by RCC cells rather than stroma COL23A1 · source_share = 0.881. Its bulk increase survives extrapolation to a pure tumour COL23A1 · rna_lfc_at_purity1 = 4.88, it replicates in TCGA-KIRC, it has no clinical drug and few ccRCC papers, and no vital-organ protein enrichment. Missing are a confirmed surface display (SURFY and CSPA disagree with UniProt), a confident structural model, a prognostic signal and any selective dependency. ^{[31,32,33,34,35,36]}

RAB42, a small Rab GTPase, is the second tumour-track candidate. It is made mostly by RCC cells RAB42 · source_share = 0.627 and its increase persists in a pure tumour RAB42 · rna_lfc_at_purity1 = 3.189, with TCGA replication, very few ccRCC papers and no drug. It is probably intracellular, survival is not associated and no dependency is shown, so it is a functional hypothesis for knockdown studies rather than an antibody target. ^{[37,38,39,84]}

In the stroma track, OLFML2A is a secreted protein made mainly by fibroblasts OLFML2A · source_share = 0.647. Its change is purity-robust and replicates in TCGA, but it is a microenvironment candidate with colon expression, no survival link and no functional evidence. JAG2, an endothelial Notch ligand, is similar but less novel. ^{[40,85,86,87]}

The immune track is weak. MILR1 is a macrophage surface receptor with no ccRCC papers, but its bulk increase largely vanishes when extrapolated to a pure tumour MILR1 · rna_lfc_at_purity1 = 0.414, so it reflects infiltration; with GAL3ST4, an internal Golgi enzyme, it is at most a macrophage-state marker. All discovery genes are data-ranked hypotheses, and none is a selective dependency. ^[42,88,89]

Spotlight: novel, tractable candidates

These are the best candidates of each track (ranked inside the track) with no clinical-stage drug, few ccRCC papers and a small-molecule or antibody route, whatever their other flags. "Off-tumor risk" means expression above 50 nTPM in a vital organ. "Not replicated" and "not purity-robust" mean the gene failed one of those two tests. Most novel tumour-cell genes are expressed in some vital organ, which is why only a few reach first-in-class.

Screen results (top 30 per track by novelty score) ?