# Steering into competitive whitespace: a public-data, computation-driven campaign nominates GUCY2C and DKK1 and delivers de novo binder leads across the esophageal adenocarcinoma trajectory

**Authors:** [Author list to be finalized]¹  
**Affiliations:** ¹[Institution]  
*Correspondence:* [email]

**Preprint — not peer reviewed. All results are computational (in-silico); no experimental validation has been performed.**

---

## Abstract

Esophageal adenocarcinoma (EAC) is a somatic, chromosomally-unstable cancer whose incidence continues to rise in Western populations, yet whose targeted-therapy landscape is concentrated on a small number of validated axes — HER2, PD-1/PD-L1, VEGF, and, most recently, Claudin-18.2. We asked whether a reproducible, competition-aware computational campaign built entirely on public data could identify differentiated therapeutic entry points and deliver credible molecular leads against them. Mining TCGA genomics (cBioPortal *esca_tcga_pan_can_atlas_2018*, n=182) confirmed a copy-number-driven disease — *TP53* altered in 87% of tumors, *CDKN2A* deleted in 39%, *CCND1* amplified in 35% — in which the drug-tractable drivers are dominated by amplified surface receptors and angiogenic factors. A structured competitive-landscape analysis (351 interventional EAC trials, 106 active; Open Targets known-drug sets) mapped where clinical effort is concentrated and, by contrast, where it is not. We deliberately steered target nomination away from the crowded axes and applied an explicit druggability filter that down-weighted lost tumor suppressors and secreted biomarkers not addressable by an engineered biologic. Two leads emerged spanning the Barrett's→dysplasia→carcinoma trajectory: **GUCY2C**, a gastrointestinal-luminal-restricted surface antigen selected for a T-cell-engager treatment arm in advanced EAC, and **DKK1**, a secreted Wnt-pathway modulator with existing clinical precedent selected for a neutralizing-trap interception arm in high-risk Barrett's. For both targets we designed de novo protein binders using an RFdiffusion→ProteinMPNN→Boltz-2 pipeline. In a pilot-scale campaign (eight backbones per target, single generation run), fifteen of sixteen designs cleared an interface-confidence threshold (ipTM > 0.5); the top GUCY2C binder reached ipTM 0.915 (complex pLDDT 0.896) and the top DKK1 binder ipTM 0.858 (complex pLDDT 0.908), each folding at a contiguous target interface. A sequence-composition liability screen tempers this: several high-ipTM designs — including the top GUCY2C binder — carry high alanine content and long homopolymer runs characteristic of a generic-scaffold failure mode, whereas the DKK1 lead is well-composed on both interface confidence and developability. ipTM is a single-model docking-confidence proxy, not a measured affinity; these are pilot-scale, in-silico leads, not experimentally tested therapeutics. The program's two decisive questions — whether GUCY2C's luminal restriction provides a genuine therapeutic window for a T-cell engager, and whether a dysplasia-regression surrogate endpoint is acceptable for DKK1 interception — remain open and lie downstream of the work reported here.

---

## Introduction

Esophageal adenocarcinoma is the predominant subtype of esophageal cancer in Western populations and the only one whose incidence has continued to rise over recent decades ([Coleman et al. 2018](https://doi.org/10.1053/j.gastro.2017.07.046)). It develops through a well-characterized histologic sequence — Barrett's metaplasia, low- and high-grade dysplasia, intramucosal carcinoma, and finally invasive, locally advanced or metastatic disease ([Spechler & Souza 2014](https://doi.org/10.1056/nejmra1314704)). Despite this uniquely legible natural history, most patients present with advanced disease, and outcomes remain poor.

The molecular basis of EAC is now well described. It is a somatic, chromosomally-unstable adenocarcinoma initiated by early loss of *TP53* and *CDKN2A* and driven forward by whole-genome doubling and focal amplification of cell-cycle and receptor-tyrosine-kinase genes, rather than by a small set of recurrently mutated oncogenic drivers ([TCGA Research Network 2017](https://doi.org/10.1038/nature20805)). This genomic architecture has shaped its therapeutic development: the drug-tractable drivers are amplified surface receptors (HER2, EGFR) and secreted angiogenic factors (VEGFA), and the approved and late-stage agents cluster tightly around them and around immune checkpoints. Trastuzumab and trastuzumab-deruxtecan (HER2) ([Bang et al. 2010](https://doi.org/10.1016/s0140-6736%2810%2961121-x)), pembrolizumab and nivolumab (PD-1, including the adjuvant setting) ([Kelly et al. 2021](https://doi.org/10.1056/nejmoa2032125)), ramucirumab (VEGFR2) ([Fuchs et al. 2014](https://doi.org/10.1016/s0140-6736%2813%2961719-5)), and zolbetuximab (Claudin-18.2, approved 2024) ([Shitara et al. 2023](https://doi.org/10.1016/s0140-6736%2823%2900620-7)) define a competitive but narrow target space.

Two consequences follow. First, a differentiated program requires deliberate movement into competitive whitespace — biology that is credible but under-exploited. Second, the disease's legible trajectory creates an opportunity that a treatment-only strategy ignores: the chance to intercept malignant progression in the large, identifiable, high-risk Barrett's population before invasive cancer develops, a setting in which no molecular therapeutic is approved and endoscopic surveillance plus ablation remains the standard of care.

Here we report a computational discovery-to-design campaign designed around both of these considerations. Working entirely from public genomics, target-association, and clinical-landscape data, we nominated candidate targets under an explicit novelty steer and an honest druggability filter, selected two leads that together span the disease trajectory, and designed de novo protein binders for each using contemporary generative structure tools. We present the campaign as a reproducible, honesty-gated template: every quantitative claim traces to a named public source or a generated computational result, and every unvalidated inference is flagged as such.

---

## Results

### EAC is a copy-number-driven disease whose tractable drivers are clinically crowded

We first characterized the somatic landscape of EAC using the TCGA PanCancer Atlas cohort (cBioPortal *esca_tcga_pan_can_atlas_2018*, n=182 after restriction to the adenocarcinoma subtype). The alteration profile is dominated by tumor-suppressor loss and copy-number change rather than recurrent activating point mutation: *TP53* is altered in 87% of tumors, *CDKN2A* is deep-deleted in 39%, and *CCND1* is amplified in 35%, followed by *WWOX* deletion (34%) and amplification of *MYC* (22%), *ERBB2* (15%), *VEGFA* (12%), *EGFR* (12%), and *GATA6* (11%) (**Figure 1a**). This confirms a chromosomally-unstable, amplification-driven biology in which the readily drug-tractable events are amplified cell-surface receptors and secreted factors, while the earliest and most penetrant events — *TP53* and *CDKN2A* loss — are intracellular and hard to drug directly.

We then mapped the clinical landscape. Across ClinicalTrials.gov we identified 351 interventional EAC trials (106 active) and, separately, 203 interventional trials in the Barrett's/dysplasia precursor setting; Open Targets known-drug sets returned 106 agents associated with the indication. Ranking targets by active-trial intervention mentions against their development maturity (**Figure 1b**) shows clinical effort concentrated on PD-1/PD-L1 (the most-mentioned axis by a wide margin), HER2, and VEGF, with Claudin-18.2 and FGFR2b emerging. By contrast, several biologically credible axes are nearly untouched in EAC: the surface antigens B7-H3, TROP2, and GUCY2C; the secreted Wnt modulator DKK1; the immunometabolic target CD73; and, as a category, molecular interception of Barrett's progression.

### A trajectory-spanning, two-arm strategy

Rather than compete on a crowded axis or pursue a single point on the disease course, we defined a two-arm strategy that attacks EAC at two under-exploited positions on its trajectory (**Figure 1c**): an **interception arm** acting across Barrett's metaplasia and dysplasia to slow or prevent progression, and a **treatment arm** acting on established early-through-advanced disease. This framing was reinforced by a patient-priorities synthesis emphasizing preservation of swallowing function, quality-of-life-weighted survival, minimization of treatment toxicity, and — in the precursor setting — a well-tolerated, less-invasive way to reduce progression risk and defer or avoid esophagectomy.

### Honest druggability triage nominates GUCY2C and DKK1

We nominated sixteen candidate targets spanning both lanes and scored them on a transparent additive composite, `composite = tumor_selectivity + novelty + tractability_score + 5 × disease_association` (component definitions, ordinal anchors, and weights in Table S0). A weight-perturbation analysis (Monte Carlo over ±50% multiplicative variation in all four weights, 20,000 draws; Table S0) tested rank robustness: among biologic-addressable treatment-lane targets, GUCY2C is the top-ranked target in 80.7% of draws and top-three in 94.2%, with CDH17 the nearest competitor. This rank-1 dominance is, however, contingent on GUCY2C's optimistic tumor-selectivity score: recalibrating it from 5 to 4 to reflect that the safety window is unresolved (O-2) drops GUCY2C's composite to 12.00, lets CDH17 (12.38) overtake it, and reduces its Monte-Carlo rank-1 frequency to ~0% (top-three ~48%; Table S0) — so O-2 is decisive for lead selection, not only safety, and CDH17 is carried as a credible backup; in the interception lane, DKK1 is the only biologic-addressable candidate after the druggability filter, so its selection is driven by the filter, not by out-competing a field — an honest reflection of how thin the addressable interception space is. Critically, we then applied an explicit biologic-druggability filter, because a high composite score does not by itself make a target reachable by an engineered protein (**Figure 2**). The single highest raw-composite candidate, *GPX7*, is a tumor suppressor lost in Barrett's — a restoration or synthetic-lethal problem, not an antibody-addressable one — and was down-weighted accordingly. Three further high-scoring interception candidates (*REG4*, *OLFM4*, *TFF3*) are secreted progression biomarkers whose therapeutic value is unproven, and the crowded validated receptors (*ERBB2*, *EGFR*, *CLDN18*) were deprioritized under the novelty steer.

After this filter, the highest-scoring biologic-addressable target for the treatment lane was **GUCY2C** (guanylyl cyclase C), a gastrointestinal-epithelial surface antigen whose apical, luminal-facing restriction in normal gut is the basis for a candidate therapeutic window, and which has antibody, ADC, bispecific, and CAR precedent in colorectal cancer. Normal-tissue RNA profiling (GTEx v8; **Figure S1**, Table S4) supports the selectivity premise: 85% of GUCY2C's normal expression is in GI tissues and its highest non-GI tissue reaches only ~1.7 TPM, a profile matched among comparators only by CDH17 and clearly distinct from the pan-epithelial TROP2 (TACSTD2). Two caveats are essential and unresolved. First, GTEx measures *normal* tissue only; it establishes tissue selectivity but does **not** establish that GUCY2C is expressed on EAC tumor cells at therapeutically actionable levels — the clinical precedent is colorectal, and EAC tumor-cell expression remains to be confirmed from tumor RNA/protein data (a gap we flag rather than paper over — GUCY2C tumor-mRNA quantification in the esca_tcga_pan_can_atlas_2018 cohort is the single highest-priority follow-up query, and the treatment-arm rationale should be considered provisional until it is reported). Second, GUCY2C is genuinely expressed in normal small intestine and colon (~28 TPM median), so "luminal restriction" is a polarity argument, not an absence-of-antigen argument, and the normal-GI safety window for a T-cell engager is precisely open question O-2. For the interception lane we selected **DKK1** (Dickkopf-1), a secreted Wnt-pathway modulator addressable by a neutralizing antibody or ligand trap. The DKN-01 program ([Klempner et al. 2024](https://doi.org/10.1200/jco.24.00410)) establishes that pharmacologic DKK1 neutralization is clinically tractable — but that precedent is in *advanced* gastric/gastro-esophageal adenocarcinoma, a late-stage treatment population with different safety tolerances and endpoints. It is **not** evidence for DKK1 interception in a largely pre-malignant Barrett's surveillance population; direct precursor-setting therapeutic evidence for DKK1 is essentially absent beyond progression-biomarker associations. We therefore carry DKK1 into the interception arm on mechanistic and druggability grounds, with the population mismatch stated explicitly as a risk (open question O-4). GUCY2C was routed to an anti-GUCY2C × anti-CD3 T-cell-engager format for advanced/peri-operative EAC; DKK1 to a neutralizing-trap format for high-risk Barrett's interception.

### De novo binders were designed for both arms

For each target we prepared a design-ready structure from the AlphaFold database and isolated the relevant domain: the GUCY2C extracellular domain (UniProt P25092, residues 24–430; the intracellular kinase/cyclase domain was excluded as irrelevant to a surface binder) and the DKK1 CRD2 domain (UniProt O94907, residues 178–256; the LRP6-binding functional region and the high-confidence structured segment). Surface-exposed, high-confidence residue patches were selected as hotspots by a neighbor-count and pLDDT heuristic. We generated eight binder backbones per target with RFdiffusion, designed sequences with ProteinMPNN, and folded each best-per-backbone binder in complex with its target using Boltz-2 (**Figure 3**).

### In-silico validation and lead complexes

In this pilot-scale run (eight backbones per target from a single generation seed), fifteen of sixteen designs cleared the interface-confidence threshold of ipTM > 0.5 (**Figure 3**). We report this as a pilot observation, not a generalizable success rate: n=8/target is too small to estimate a rate, ipTM > 0.5 is a permissive bar, and the values are single-best-of-five diffusion samples. The top treatment-arm binder by ipTM, gucy2c_bb2 (71 aa), reached ipTM 0.915 with a complex pLDDT of 0.896 (seven of eight GUCY2C designs passed); the top interception-arm binder, dkk1_bb1 (63 aa), reached ipTM 0.858 with a complex pLDDT of 0.908 (all eight DKK1 designs passed). A sequence-composition liability screen (Table S5) materially qualifies the treatment-arm lead: gucy2c_bb2 is 44% alanine with an eight-residue homopolymer run, and several other high-ipTM designs (dkk1_bb2/bb3/bb5) exceed 50% alanine — the signature of a generic low-complexity scaffold that can fold with spurious confidence. By this combined criterion, **dkk1_bb1 is the most robust lead** (ipTM 0.858, 17.5% alanine, maximum run 2), whereas gucy2c_bb2's headline ipTM must be read against its composition liability and treated as provisional pending redesign toward a lower-alanine backbone. In the predicted complexes (**Figure 4**), both binders dock against a contiguous, well-localized surface on their target: the GUCY2C binder contacts a patch of 13 target residues (Cα < 8 Å) on the extracellular domain (UniProt P25092 positions 153–402), and the DKK1 binder contacts an eight-residue patch mapping to UniProt O94907 positions 187–208 — confirming engagement squarely within CRD2 (178–256). A pairwise sequence-identity check across the design set (length-penalized) found maximum identities of 21.4% (GUCY2C set) and 32.9% (DKK1 set; the most-similar pair, dkk1_bb2/bb3, are both alanine-rich), well below any high-similarity threshold — the designs are diverse solutions, not a single repeated motif (Table S5). These are pilot-scale in-silico leads generated in a single design cycle. All structural confidence rests on single-model Boltz-2 ipTM/pLDDT without orthogonal consensus (AlphaFold-Multimer or Chai-1 fold-back, and monomer/decoy specificity controls, are the immediate next computational step, O-1); ipTM is a docking-confidence proxy, not a measurement of affinity; and no experimental characterization has been performed.

---

## Discussion

This campaign shows that a public-data, competition-aware computational workflow can move from open genomics to differentiated therapeutic leads while deliberately routing around crowded target space. Its principal output is not a single molecule but a coherent, trajectory-spanning hypothesis backed by concrete designed leads: intercept EAC progression in Barrett's by neutralizing secreted DKK1, and treat established disease by redirecting T cells against the GI-restricted surface antigen GUCY2C — two axes chosen precisely because the HER2/PD-1/VEGF field is saturated.

Several features distinguish the work from a purely aspirational target list. The druggability triage was applied honestly: the highest raw-scoring candidate was rejected because it is not addressable by the chosen modality, and secreted biomarkers with no therapeutic evidence were down-weighted rather than promoted for narrative convenience. The design step produced real, structurally-validated binders for both arms rather than a specification, with high interface and fold confidence.

The program's strengths are matched by two decisive, unresolved questions, both of which lie downstream of anything demonstrated here. The first is the **GUCY2C therapeutic window**: the luminal-restriction argument is a polarity- and species-dependent hypothesis, and whether it constitutes a genuine safety margin for a T-cell engager against an antigen also expressed on normal enterocytes can only be answered with human normal-tissue immunohistochemistry and redirected-cytotoxicity selectivity data. The second is the **DKK1 interception surrogate**: intervening in a largely pre-malignant Barrett's population sets a very high safety bar and requires regulatory acceptance of histologic dysplasia regression (or a comparable molecular surrogate) as a registrational endpoint — acceptance that is not established. Either question could halt its arm.

The limitations of the computational work should be stated plainly. (i) All structural validation rests on single-model Boltz-2 confidence without orthogonal consensus; AlphaFold-Multimer / Chai-1 fold-back and monomer/decoy specificity controls (O-1) are the immediate next step and have not yet been run, so ipTM values are provisional and cannot be distinguished from generic-scaffold artifacts without them. (ii) The sequence-composition screen (Table S5) already shows that the top GUCY2C design by ipTM carries a high-alanine, low-complexity liability, so interface-confidence rank and developability rank disagree and the treatment-arm lead should be regarded as a starting point for redesign rather than a finished binder. (iii) GTEx establishes GUCY2C's *normal-tissue* selectivity but not its EAC *tumor-cell* expression, which must be confirmed from tumor data before the treatment-arm rationale is secure. For DKK1, the interception arm targets an otherwise-well Barrett's surveillance population, so the safety bar is fundamentally higher than for the advanced-disease DKN-01 precedent: DKK1 has canonical roles in Wnt-dependent bone and skeletal homeostasis, its normal-tissue RNA (Table S4) is dominated by a cultured-fibroblast signal (a cell-culture artifact) with low solid-tissue levels consistent with a stromal/paracrine source, and systemic neutralization in well patients carries skeletal and developmental-signaling risks that late-stage oncology dosing does not have to clear. We therefore list skeletal/systemic toxicity explicitly, alongside GI on-target effects, as a threshold risk for the interception arm (gate G5, O-4). (iv) The epitope hotspots were chosen by a structural heuristic rather than experimental functional-site mapping, so it remains to be confirmed that the DKK1 binder blocks the LRP6 interface and that the GUCY2C binder engages a tumor-exposed epitope. (v) The GUCY2C binder is only one arm of a bispecific whose CD3 arm and format engineering are not done. (vi) The triage operated over a manually assembled sixteen-target shortlist rather than a fully enumerated candidate universe, and the trial-count extraction lacks recorded query strings and snapshot dates. No biophysical, cellular, or in-vivo data exist. The immediate next steps are correspondingly computational and experimental: an ESM-based developability and liability screen with orthogonal fold-back for confidence consensus, followed by expression and biophysical characterization, functional and normal-tissue-selectivity assays, and — for the interception arm — an early regulatory interaction to settle the surrogate-endpoint question before any long prevention trial is contemplated.

Framed appropriately, the campaign is a reproducible template for early, differentiated, honesty-gated target and lead generation, and a concrete dual-arm starting point for EAC that a wet-lab program could take forward.

---

## Methods

### Genomic characterization
Somatic alteration frequencies were obtained from cBioPortal for the TCGA PanCancer Atlas esophageal cohort (*esca_tcga_pan_can_atlas_2018*), restricted to the adenocarcinoma subtype (n=182). Alteration classes (amplification, deep deletion, mutation) were taken as reported by cBioPortal; the drivers shown in Figure 1a are the most frequently altered genes in the cohort.

### Target-association and competitive-landscape analysis
Target–disease associations and tractability were drawn from Open Targets (EAC MONDO_0005028; Barrett's esophagus MONDO_0013662) and known-drug sets. The competitive landscape was assembled from ClinicalTrials.gov interventional-trial counts (351 EAC, 106 active; 203 Barrett's/dysplasia) reported in the upstream program artifacts, and from approved-agent records confirmed against Drugs@FDA. The per-axis "active EAC trial intervention mentions" (Figure 1b, x-axis; Table S3) are axis-level annotations of active interventional trials referencing each target/axis; they were not de-duplicated across combination trials and therefore do not sum to the 106 active-trial total, and should be read as relative crowding indicators rather than exact disjoint counts. The exact extraction query strings and snapshot dates were not recorded in the upstream artifacts and should be captured with a raw trial-ID list before external submission — we flag this as a reproducibility gap in the Limitations.

### Target triage
Candidates were scored on the additive composite `composite = tumor_selectivity + novelty + tractability_score + 5 × disease_association`, where tumor_selectivity and novelty are 1–5 ordinal analyst scores (anchors in Table S0), tractability_score maps antibody-tractability tier to an integer (surface-confirmed = 3, clinical-stage = 4, clinical-validated = 5), and disease_association is the Open Targets association score (EAC for treatment-lane targets, Barrett's for interception-lane targets; 0 when unavailable). This rubric exactly reproduces all sixteen published composite values (maximum residual 0.005). A separate biologic-druggability category (STRONG/GOOD/MODERATE = addressable by an engineered biologic; CROWDED = deprioritized under the novelty steer; WEAK-BIOMARKER and NOT-A-BINDER-TARGET = down-weighted) was assigned from target biology and applied as an explicit filter over the raw composite (Figure 2). Rank robustness was assessed by Monte Carlo (20,000 draws, each weight scaled by an independent U(0.5,1.5) factor and the association weight by U(2.5,7.5)), restricted to biologic-addressable targets within each lane (Table S0). The candidate universe was the manually assembled sixteen-target shortlist; a fully enumerated association-ranked universe is noted as a limitation.

### Target-structure preparation
Design-ready structures were taken from the AlphaFold database: GUCY2C (UniProt P25092) and DKK1 (UniProt O94907). The GUCY2C extracellular domain (residues 24–430) and the DKK1 CRD2 domain (residues 178–256) were isolated; disordered and functionally-irrelevant regions were excluded. Hotspot residue patches were selected by a combined neighbor-count and pLDDT heuristic (GUCY2C ECD patch [239,240,267,422,424,425]; DKK1 CRD2 patch [181,182,183,194,195], in the isolated-domain numbering).

### De novo binder design
Binder backbones were generated with RFdiffusion (Complex_base checkpoint, noise 0), eight backbones per target, binder length 70–100 aa (GUCY2C) and 60–90 aa (DKK1). Sequences were designed with ProteinMPNN (model v_48_020, binder chain only, eight sequences per backbone at sampling temperature 0.1). Each best-per-backbone binder was folded in complex with its target using Boltz-2 (target chain with `--use_msa_server`, binder as single sequence, five diffusion samples). Designs were assessed by interface ipTM (pass threshold > 0.5) and complex pLDDT (fold threshold > 0.7). Interface residues reported here were computed directly from the predicted complex coordinates as target residues with a Cα within 8 Å of any binder Cα, then mapped to canonical UniProt positions (GUCY2C → P25092 153–402; DKK1 → O94907 187–208, within CRD2; full offset mapping in Table S6). All GPU steps were run on cloud A100 hardware.

### Reproducibility and honesty gating
Every quantitative claim in this manuscript traces to one of: cBioPortal, Open Targets, ClinicalTrials.gov, Drugs@FDA, or a generated computational result reported in the design tables. In-silico results are labeled as such throughout and are not represented as experimental validation; clinical, regulatory, and commercial inferences appear only in the referenced strategy documents (Supplementary Information) and are flagged there as assumptions requiring expert confirmation.

---

## Data and code availability
The upstream program artifacts — machine-readable dossier, target-nomination and competitive-landscape tables, the full binder-design table (all sixteen designs with sequences and confidence metrics), the two lead complex structures (PDB), and the scientific/regulatory/commercial strategy briefs — accompany this manuscript as Supplementary Information. Public data sources (cBioPortal *esca_tcga_pan_can_atlas_2018*; Open Targets MONDO_0005028 and MONDO_0013662; ClinicalTrials.gov; AlphaFold DB entries P25092 and O94907) are openly accessible. Design tools (RFdiffusion, ProteinMPNN, Boltz-2) are publicly available.

## Author contributions and AI-assistance disclosure
This campaign was executed as a human-supervised, AI-assisted computational workflow, and we distinguish three categories of contribution. (1) **Deterministic bioinformatics** — retrieval of TCGA/cBioPortal alteration frequencies, Open Targets associations, ClinicalTrials.gov and Drugs@FDA records, GTEx expression, and AlphaFold structures; the additive triage composite and its Monte-Carlo robustness analysis; interface-residue and sequence-composition computations — was reproducible code, not model judgment. (2) **Published ML methods as objects of study** — RFdiffusion (backbone generation), ProteinMPNN (sequence design), and Boltz-2 (complex fold-back) — generated the binder designs and their confidence metrics. (3) **AI-assisted synthesis and drafting** — an LLM-based agent proposed the novelty steer, the two-arm framing, candidate shortlisting, the ordinal selectivity/novelty annotations in Table S2, and the manuscript prose; every such judgment was reviewed by a human supervisor, and every quantitative claim was regenerated from, or traced to, a deterministic source (see `claims_to_sources.csv`). The ordinal annotations and the "whitespace verdict" column reflect analyst judgment informed by the cited data, not an independent measurement. Final author list, individual contributions (CRediT), and human-supervisor identities to be completed.

## Competing interests
[To be declared.]

## Acknowledgements
This work was performed as part of a life-sciences computational drug-program exercise. No experimental or clinical work was conducted.

## References

1. Coleman et al. (2018). *The Epidemiology of Esophageal Adenocarcinoma*. [10.1053/j.gastro.2017.07.046](https://doi.org/10.1053/j.gastro.2017.07.046)

2. Spechler & Souza (2014). *Barrett's Esophagus*. [10.1056/nejmra1314704](https://doi.org/10.1056/nejmra1314704)

3. TCGA Research Network (2017). *Integrated genomic characterization of oesophageal carcinoma*. [10.1038/nature20805](https://doi.org/10.1038/nature20805)

4. Bang et al. (2010). *Trastuzumab in combination with chemotherapy versus chemotherapy alone for treatment of HER2-positive advanced gastric or gastro-oesophageal junction cancer (ToGA)*. [10.1016/s0140-6736(10)61121-x](https://doi.org/10.1016/s0140-6736%2810%2961121-x)

5. Kelly et al. (2021). *Adjuvant Nivolumab in Resected Esophageal or Gastroesophageal Junction Cancer (CheckMate 577)*. [10.1056/nejmoa2032125](https://doi.org/10.1056/nejmoa2032125)

6. Fuchs et al. (2014). *Ramucirumab monotherapy for previously treated advanced gastric or gastro-oesophageal junction adenocarcinoma (REGARD)*. [10.1016/s0140-6736(13)61719-5](https://doi.org/10.1016/s0140-6736%2813%2961719-5)

7. Shitara et al. (2023). *Zolbetuximab plus mFOLFOX6 in patients with CLDN18.2-positive, HER2-negative, untreated, locally advanced or metastatic gastric or gastro-oesophageal junction adenocarcinoma (SPOTLIGHT)*. [10.1016/s0140-6736(23)00620-7](https://doi.org/10.1016/s0140-6736%2823%2900620-7)

8. Klempner et al. (2024). *DKN-01 in Combination With Tislelizumab and Chemotherapy as First-Line Therapy in Advanced Gastric or Gastroesophageal Junction Adenocarcinoma (DisTinGuish)*. [10.1200/jco.24.00410](https://doi.org/10.1200/jco.24.00410)