論文紹介: Proximity-guided graph learning reveals tumour-associated proximity antigens

Nature | Vol 658 | 09 September 2026 Proximity-guided graph learning reveals tumour-associated proximity antigens Cody Scandore, Clare F. Malone, Christopher K. May et al. Corresponding authors: Rob C. Oslund and Olugbeminiyi O. Fadeyi InduPro, Cambridge, Massachusetts · Seattle, Washington, USA 2026.11.02 Full | Shunichi Ito

Proximity reveals therapeutic co-targets Fig. 1 2

A nearby antigen can refine tumour targeting Local membrane neighbourhoods TAA: tumour-associated antigen Fig. 1, enlarged details TAA + nearby TAPA TAPA: tumour-associated proximity antigen 3

The photocatalyst labels only nearby proteins Ir/Dz Short-lived carbene ≤ 5 ns RFT/BT Longer-lived radical ~0.1 ms Fig. 2a 4

Spatial specificity is fixed before lysis 248 micromaps: 12 RTKs across 28 tumour cell systems Streptavidin enrichment, LysC/trypsin digestion, TMTpro, SPS-MS3 Fig. 2a 5

Proximity mapping reveals unexplored RTK pairs Novel proximal pairs: EGFR–PTK7, MET–EPHB2, HER3–MST1R Only EGFR×MET, EGFR×HER3 and HER2×HER3 are clinically explored Fig. 2b,c 6

MetaMap infers neighbours from shared enrichment Spearman correlation across 248 micromaps Fig. 3a,b 7

MetaMap extends a known integrin community 2.7–6.5× over random protein pairs FAP, CD81, RECK and PAPPA join the integrin core without prior annotation Fig. 3d,e 8

Three models predict which proteins are associated Proximity enrichment Protein expression Protein association DIA measurement Reported Network (STRING) Graph Neural Network database of protein associations Three kinds of GNN Fig. 4b 9

Proximity adds predictive value beyond abundance Fig. 4c, GAT feature ablation 10

CDCP1 emerges as an EGFR-associated co-target MetaMap support Model-prioritized EGFR partners: HER2, HER3, MET and the non-RTK CDCP1 Fig. 4d,e 11

EGFR–CDCP1 proximity depends on cell context Surface expression Fig. 4f,g EGFR-anchored enrichment 12

EGFR–CDCP1 proximity persists across tumour contexts Fig. 4h 13

Dual engagement increases antibody internalization Unconjugated antibody Uptake is much lower in primary epithelial cells Fig. 5a,b. Representative SW48 result 14

The bispecific ADC increases tumour-cell killing vcMMAE payload, Drug / Ab (DAR) = 2 SW48 IC50: 11.4 nM Primary epithelial cells: 324–947 nM Fig. 5c. Primary-cell IC50: Supplementary Fig. 28c 15

The bispecific ADC improves control at a lower dose 1 mg/kg: complete TGI versus 43% with EGFR-only ADC Fig. 5d,e. SW48 xenografts, n = 8 mice per group 16

Coordinated targeting outperforms separate ADCs Fig. 5f,g. SW900 xenografts, n = 9 mice per group 17

The same target pair improves T-cell-directed killing Trispecific TCE: T-cell engagement Fig. 5h,i. In vitro T-cell-dependent cytotoxicity 18

Limitation of proximity mapping Local labelling does not resolve direct binding Avidity and trafficking also influence efficacy Proximity maps rest on technical triplicates only Tumour and normal tissues from the same patient remain essential 19

Surface proteomics can prioritize spatial co-targets Measure local neighbourhoods Prioritize TAA–TAPA pairs Validate therapeutic co-engagement 20

Appendix

Primary epithelial cells show lower uptake and killing Supplementary Fig. 28b,c B1

CDCP1 depletion reduces bispecific ADC activity TGI: 97% in parental tumours, 46% after CDCP1 knockdown Extended Data Fig. 2c,d. SW48 dual-flank study B2

MAD normalization makes micromaps comparable Supplementary Fig. 9. Methods B3

Graph construction and model architectures Fig. 4a. Supplementary Notes 1 and 2 B4

Reciprocal anchoring reproduces the EGFR–CDCP1 pair Reciprocal anchoring Monovalent direct conjugates Both directions, both chemistries, a different epitope, no secondary antibody Extended Data Fig. 1b,d B5

Both chemistries recover every intended target Target enriched in 248 of 248 maps Supplementary Fig. 8 Across chemistries: r = 0.679, bias −0.55 log2FC B6

Enrichment is independent of proteomic abundance Supplementary Fig. 13 Supplementary Fig. 14 B7

GAT results are consistent with DepMap Supplementary Fig. 18 B8

NanoBiT experiments also validated that bispecific antibody capture both EGFR and CDCP1 Supplementary Fig. 27 B9

Reciprocal EGFR-CDCP1 enrichment is reproduced across micromapping formats, binders and cell systems. Extended Data Fig. 1 B10

TMTpro-SPS-MS3 resolves chimera spectra Step 1: TMTpro labeling Step 3: Select fragment ion MS2 Synchronous precursor selection (SPS) Step 2: Select precursor ion Step 4: Quantitate reporter ion MS1 MS3 B11