# Figure Legends

**Figure 1. EAC is a copy-number-driven cancer whose tractable drivers are clinically crowded, motivating a two-arm, trajectory-spanning strategy.**
(a) Somatic alteration frequencies in the TCGA PanCancer Atlas esophageal adenocarcinoma cohort (cBioPortal *esca_tcga_pan_can_atlas_2018*, n=182), colored by alteration class. The earliest and most penetrant events (*TP53* mutation 87%, *CDKN2A* deletion 39%) are intracellular and hard to drug; the tractable drivers are amplified surface receptors and secreted factors. (b) Competitive landscape of EAC target axes, positioning each axis by active-EAC-trial intervention mentions (x, log scale; higher = more crowded) against development maturity (y), colored by target biology. Clinical effort concentrates on PD-1/PD-L1, HER2, and VEGF; the selected leads GUCY2C and DKK1 (black outline) sit in competitive whitespace. (c) The two-arm strategy on the disease trajectory: a DKK1 neutralizing-trap interception arm acting across Barrett's metaplasia and dysplasia, and a GUCY2C T-cell-engager treatment arm acting on early-through-advanced disease.

**Figure 2. Honest druggability triage nominates GUCY2C and DKK1.**
Sixteen candidate targets scored on a composite of disease association, tumor selectivity, antibody tractability, and novelty (x-axis), then filtered by whether an engineered biologic can actually reach them (color: green = biologic-addressable; grey = crowded/poor-window and deprioritized under the novelty steer; red = not a binder target or an unproven secreted biomarker). Marker shape encodes trajectory lane (circle = treatment, diamond = interception). The single highest raw-composite candidate, *GPX7*, is a lost tumor suppressor and is not antibody-addressable; the selected leads are the highest-scoring targets an engineered biologic can reach in each lane. Source: `target_nomination.csv`.

**Figure 3. De novo binder design and in-silico validation.**
Binder campaign for both arms: RFdiffusion backbone generation -> ProteinMPNN sequence design -> Boltz-2 fold-back of each binder in complex with its target. Interface confidence (ipTM) and fold confidence (complex pLDDT) for all designs; 15 of 16 clear the ipTM>0.5 interface threshold. (Reproduced from the design campaign; source `design_results.png` / `design_leads.csv`.)

**Figure 4. Predicted lead complexes fold at contiguous target interfaces.**
Boltz-2 predicted complexes for the two lead binders, shown as Calpha backbone traces (target grey, binder colored). Target residues within 8 A (Calpha-Calpha) of the binder are marked (navy). (a) gucy2c_bb2 (71 aa) on the GUCY2C extracellular domain (ipTM 0.915, complex pLDDT 0.896), contacting a 13-residue patch. (b) dkk1_bb1 (63 aa) on the DKK1 CRD2 / LRP6-binding region (ipTM 0.858, complex pLDDT 0.908), contacting an 8-residue patch. Backbone traces depict topology and interface localization, not experimentally-determined secondary structure; ipTM is a docking-confidence proxy, not a measured affinity. Source: `gucy2c_bb2_complex.pdb`, `dkk1_bb1_complex.pdb`.