# Supplementary Information

**De novo design of neutralizing protein binders against CCL26 and POSTN, two convergent targets in eosinophilic esophagitis**

This document contains supplementary figures S1–S12, supplementary tables S1–S5, and supplementary media (two 3D rotation videos). All data products are available as project artifacts.

---

## Supplementary Figures

### Figure S1 — Meta-analysis effect-size landscape

![Figure S1]({{artifact:563ed14c-138f-4fe8-9852-a6a0ddca6345}})

**Figure S1.** Volcano plot of the nine-cohort EoE random-effects meta-analysis (25,654 genes): pooled log₂ fold-change vs. −log₁₀ adjusted *p*. CCL26 and POSTN (bold, colored) are among the most strongly up-regulated genes; concordantly down-regulated barrier genes (DSG1, SPINK7, FLG) appear on the left. Dashed horizontal line = FDR 0.05; dotted vertical lines = ±1 log₂FC.

### Figure S2 — Literature attention vs. omics effect size

![Figure S2]({{artifact:86fde4cf-dbd2-40a0-a232-157db8e3363c}})

**Figure S2.** Cross-reference of the meta-analysis against the EoE literature corpus (591 abstracts, 2015–2026) and interventional-trial landscape (157 trials). x-axis: meta-analysis pooled log₂FC; y-axis: number of abstracts mentioning the target; marker size proportional to interventional-trial count; color indicates literature↔omics direction concordance. CCL26 and POSTN combine high omics effect size with zero direct interventional trials, unlike the heavily-studied cytokine nodes (IL-13, IL-5, TSLP).

### Figure S3 — Single-cell cell-of-origin

![Figure S3]({{artifact:25789867-4c22-4d5f-9a8f-368270a9933b}})

**Figure S3.** Single-cell differential-expression dot plot for the up-regulated candidate targets across EoE epithelial and effector-cell compartments. Dot color = log fold-change, dot size = −log₁₀ adjusted *p*. CCL26 localizes to suprabasal epithelium and POSTN to basal epithelium; CAPN14 spans both epithelial layers, CPA3 marks mast cells, and ALOX15 marks suprabasal epithelium and myeloid cells. (Only genes with up-regulated single-cell markers in the source table are shown.)

### Figure S4 — Omics druggability shortlist

![Figure S4]({{artifact:4abd669a-e097-4c3f-b081-8ae7fc3ad3cb}})

**Figure S4.** Omics-derived druggability priority score for the 20-gene shortlist (filtered for effect size, direction concordance, single-cell specificity, and antibody tractability). POSTN ranks 3rd of 20 and CCL26 11th of 20; both are secreted and suitable for a ligand-trap modality.

### Figure S5 — RFdiffusion backbone quality control

![Figure S5]({{artifact:4bc82cff-d1f5-4de1-a39a-005bbf58547c}})

**Figure S5.** Distributions of binder length, radius of gyration, and hotspot contacts (binder Cα within 10 Å of hotspot Cα) for 40 RFdiffusion backbones per target. All 40 CCL26 and 34/40 POSTN backbones engage ≥3 hotspot residues (dashed line).

### Figure S6 — Epitope definition and solvent accessibility

![Figure S6]({{artifact:743ca3f0-7a3d-421a-8409-23e7a9fee8b1}})

**Figure S6.** Per-residue solvent-accessible surface area (Shrake–Rupley) for the CCL26 and POSTN target structures, with design epitope hotspots highlighted. CCL26 hotspots (16, 17, 54, 55, 56) span the N-loop and C-terminal α-helix basic cluster (CCR3/GAG face); POSTN hotspots (83, 85, 108, 116) form an exposed patch on the FAS1-IV integrin-interaction surface. All epitope residues are surface-exposed.

### Figure S7 — SolubleMPNN score distributions

![Figure S7]({{artifact:403c2f20-30ea-4bbc-a71e-12cf979a039d}})

**Figure S7.** SolubleMPNN score distributions (lower = better) for all 1,920 designed sequences, split by target and sampling temperature (T = 0.1, 0.2, 0.3). Higher temperature broadens the score distribution as expected.

### Figure S8 — Sequence-complexity filter

![Figure S8]({{artifact:980ffd36-b4c7-40df-ae1a-f66e48dc3ca0}})

**Figure S8.** Sequence entropy vs. maximum single-amino-acid fraction for all designed sequences. The complexity filter (entropy ≥ 2.6 bits, max single-aa fraction ≤ 0.40, max single-residue run ≤ 5; dashed lines) removes low-diversity designs, retaining 1,074/1,920 sequences.

### Figure S9 — Full fold-back confidence landscape

![Figure S9]({{artifact:d6f7cc24-0d05-4e5c-bfcc-8df2b90c6261}})

**Figure S9.** Interface ipTM vs. complex pLDDT for all 60 Boltz-2–folded complexes; marker size proportional to epitope hotspots contacted. 59/60 pass the confidence thresholds (ipTM > 0.5, pLDDT > 0.7; dashed lines). Leads are circled.

### Figure S10 — Epitope-coverage distribution

![Figure S10]({{artifact:85e6db07-ca83-43b2-a8d5-363b76507082}})

**Figure S10.** Distribution of the number of epitope hotspots contacted per design (target atom within 5 Å of a binder atom). CCL26 designs engage the epitope more focally (13/30 with ≥3 hits) than POSTN designs (5/30 with ≥3 hits), consistent with the flatter FAS1 integrin face.

### Figure S11 — On-target score ranking

![Figure S11]({{artifact:f74cb274-a5c8-4eee-8e5c-7a7bebb3bc08}})

**Figure S11.** On-target composite score — ipTM × (0.5 + 0.5 × epitope-coverage) × min(pLDDT/0.7, 1) — for all 30 designs per target, ranked. The top-ranked design per target (marked) was selected as the lead.

### Figure S12 — Lead interface contact maps

![Figure S12]({{artifact:4c679536-95c8-4f62-821f-a0042da1ef59}})

**Figure S12.** Residue-level minimum inter-chain distance maps for the two lead complexes (target residue × binder residue). Colored horizontal lines mark epitope hotspot rows; the close-contact (yellow) bands align with the hotspots, confirming on-target engagement at the residue level.

### Figure S13 — Negative-control calibration (main Figure 5, full)

![Figure S13]({{artifact:36ccbb5f-dae3-48e2-88b1-c7959a5b47e4}})

**Figure S13.** Negative-control calibration of the Boltz-2 co-folding readout (identical to main Figure 5, reproduced here at full size). Designs vs. composition-matched scrambled-sequence decoys vs. random-composition decoys, all folded under the identical protocol. (A) interface ipTM; (B) complex pLDDT; (C) percentage crossing the ipTM > 0.5 & pLDDT > 0.7 gate. Scrambled decoys cross the confidence gate at design-level rates (92–100%), and design > scrambled ipTM is non-significant (Mann–Whitney CCL26 p = 0.16, POSTN p = 0.80).

### Figure S14 — Orthogonal-predictor concordance (main Figure 7, full)

![Figure S14]({{artifact:570deef0-1c87-4925-921b-af4d0b0b0940}})

**Figure S14.** Boltz-2 vs. Chai-1 interface ipTM for all 60 designs (identical to main Figure 7, reproduced here at full size). (A) per-design scatter; (B) median ipTM by model. Overall Spearman ρ = −0.13 (n.s.); 7% of designs pass Chai-1's ipTM > 0.5 threshold vs. 98% in Boltz-2.

---

## Supplementary Tables

- **Table S1 — Meta-analysis (top genes).** Nine-cohort random-effects meta-analysis results (pooled log₂FC, SE, z, p, adjusted p, I², k, direction) for the top 60 genes by |z|, including CCL26 and POSTN. `TableS1_meta_top60.csv`
- **Table S2 — Druggability shortlist.** 20-gene omics-derived shortlist with priority score, accessibility, antibody tractability, single-cell cell-types, and suggested modality. `TableS2_druggability_shortlist.csv`
- **Table S3 — Literature × omics cross-reference.** 45 targets with literature abstract/direction counts, trial counts and max phase, and meta-analysis effect sizes. `TableS3_lit_omics_crossref.csv`
- **Table S4 — Fold-back validation (all 60 designs).** Per-design Boltz-2 metrics: ipTM, complex pLDDT, pTM, confidence, interface residue count, epitope hits and coverage, on-target score. `TableS4_foldback_validation.csv`
- **Table S5 — SolubleMPNN summary.** Per-design/temperature sequence summary: sequence count, best/mean MPNN score, mean entropy (240 backbone×temperature groups). `TableS5_mpnn_summary.csv`
- **Table S6 — Negative-control calibration and orthogonal-predictor concordance.** Per-design Boltz-2 metrics (ipTM, complex pLDDT, pTM, epitope hits, on-target score) alongside Chai-1 interface ipTM and the Boltz−Chai ipTM difference for all 60 designs; the companion `decoy_control_results.csv` gives the same Boltz-2 metrics plus epitope hits for the 36 negative-control complexes (24 scrambled, 8 random, 4 reproduction controls). `TableS6_concordance_and_controls.csv`, `decoy_control_results.csv`, `decoy_control_spec.csv`

---

## Supplementary Media

- **Movie S1 — CCL26 lead complex.** 360° rotation of the CCL26 lead binder–target complex (target grey surface, epitope hotspots blue, binder orange cartoon). `CCL26_binder_complex_360.mp4`
- **Movie S2 — POSTN lead complex.** 360° rotation of the POSTN lead binder–target complex. `POSTN_binder_complex_360.mp4`

---

## Supplementary Methods — key parameters

| Stage | Tool | Key parameters |
|---|---|---|
| Target prep | Biopython | 1G2S chain A (CCL26, 71 aa); 5WT7 chain A (POSTN FAS1-IV, 140 aa); first NMR model, standard residues |
| Hotspots | Cα spatial compactness | CCL26 [16,17,54,55,56]; POSTN [83,85,108,116] |
| Backbones | RFdiffusion | Complex/PPI checkpoint; binder 65–90 aa; noise scale 0; 40/target |
| Sequences | SolubleMPNN (v_48_020) | design chain B, target fixed; 8 seqs × 3 T (0.1/0.2/0.3); 1,920 total |
| Complexity filter | — | entropy ≥ 2.6; max-run ≤ 5; top-aa ≤ 0.40 (1,074 pass); 30/target selected |
| Co-folding assessment | Boltz-2 (`boltz`) | `--use_msa_server`, 3 recycling steps, 3 diffusion samples; ipTM > 0.5 & pLDDT > 0.7 recorded as threshold-crossing rate (not a validation pass — see negative-control calibration) |
| On-target score | — | ipTM × (0.5 + 0.5 × epitope-coverage) × min(pLDDT/0.7, 1); epitope-coverage = hotspots with target atom < 5 Å of binder; weights/floor heuristic; ρ = 0.70 vs raw ipTM |
| Negative-control calibration | Boltz-2 | per target: 12 scrambled (residue-permuted, composition-matched) + 4 random-composition + 2 reproduction decoys (36 total), identical protocol; one-sided Mann–Whitney U on ipTM/pLDDT/epitope/on-target |
| Orthogonal predictor | Chai-1 (`chai_lab`) | all 60 designs; ESM2 embeddings, MSA server, 3 trunk recycles, 200 diffusion timesteps, seed 42; Spearman vs Boltz-2 ipTM |
| Druggability priority score | — | composite of pooled effect magnitude, direction concordance (I²-penalized), single-cell specificity, and antibody/ligand-trap tractability, normalized 0–100; heuristic, weights in released code |
| Lead structures | Boltz-2 / Biopython | buried SASA = (SASA_target + SASA_binder − SASA_complex)/2, Shrake–Rupley |
| Rendering | PyMOL + ffmpeg | 180-frame 360° turn, libx264 30 fps |