Code and model outputs for:
STRIVE: An Integrated In Silico Framework for Structure-Guided De Novo Design and Off-Target Risk Prioritization of CLDN4-Targeting scFv Candidates
STRIVE (Structural Therapeutic Risk and Interaction Viability Engine) is a model-derived
prioritization score, not an affinity or safety predictor. Its measured discrimination on
an independent panel of solved antibody–antigen complexes is weak (AUROC 0.606, 95% CI
0.348–0.846) and it does not outperform interface ipTM used alone. Please read
revision_analyses/output/benchmark_discrimination.csv before reusing it.
pipeline/
design/ RFantibody (RFdiffusion + ProteinMPNN + RF2) driver scripts, PRODIGY, AF3 input prep
alphafold3/ Nextflow workflow for AF3 MSA + inference; AF3 output parser
scoring/ STRIVE implementation and the refined/clean OTR variants
md/ GROMACS .mdp inputs, topology, run scripts, and analysis .xvg outputs
expression/ CLDN4 target selection: TCGA/LuCA pseudobulk, CPS scoring, HPA and IHC
notebooks/ Figure-generating notebooks
results/
scoring_tables/ on-target and off-target STRIVE scores, RF2 scores for all 1,000 designs,
PRODIGY ΔG tables, clinical monoclonal panel
benchmark/ independent decoy-panel AF3 metrics and scored outputs
revision_analyses/
scripts/ R1–R8: every number added during peer review, reproducible end to end
output/ their outputs (specificity margins, benchmark, seed variability,
paired IHC statistics, LuCA FDR statistics, HPA pan-tissue, MD interface)
figures/ Supplementary Figure 3 (vector)
environment/ pinned Python environment for revision_analyses
SOFTWARE_VERSIONS.md full provenance: versions, model checkpoints, database snapshots
MANIFEST.md what each file is
pip install -r environment/analysis_requirements.txt
cd revision_analyses
python3 scripts/R1_specificity_reanalysis.py # Table 1, worst-case margins
python3 scripts/R2_strive_benchmark.py # decoy benchmark + ablation
python3 scripts/R3_af3_seed_variability.py # AF3 seed-to-seed variability
python3 scripts/R4_ihc_paired_stats.py # paired IHC statistics
python3 scripts/R5_luca_pseudobulk_stats.py # LuCA extraction (needs the atlas)
python3 scripts/R5b_luca_stats_from_profiles.py # FDR + donor-level statistics
python3 scripts/R6_hpa_pan_normal_tissue.py # HPA pan-normal-tissue / pan-cancer
python3 scripts/R7_md_interface_analysis.py # MD interface metrics (needs trajectory)
python3 scripts/R8_supp_fig3_mab_heatmap.py # Supplementary Figure 3R5 needs the LuCA atlas (202603_luca.h5ad) and R7 needs the 500 ns trajectory. Both are too
large for git, so each script reads its location from the environment and says so if it is
unset: point DESCARTES_LUCA and DESCARTES_MD at them. Every other input the scripts read
is in results/, and they resolve it from their own location, so the tree runs wherever it is
unpacked and from any working directory. See revision_analyses/scripts/deposit_paths.py.
Deposited on Zenodo because of size, under the concept DOI above: the 500 ns production
trajectory (step7_production_500ns.xtc), the periodic-boundary-corrected trajectory used for
analysis, the GROMACS run input (.tpr, which embeds the topology and the force-field
parameters), the topology, the run-parameter files, the equilibration coordinates and the run
logs.
A public download, not ours to redistribute: the LuCA single-cell atlas, from cellxgene.cziscience.com.
Not available anywhere. Stated here rather than implied, because an earlier version of this
file claimed otherwise. The raw AlphaFold3 structure outputs for the 975 antibody–antigen pairs
were not retained; the per-pair confidence metrics they produced, from which every reported
value derives, are in results/. The CHARMM-GUI toppar/ bundle is superseded by the run
input and is not deposited separately. The designed antibody sequences are not deposited: the
modelled heavy- and light-chain variable sequences of the lead candidate, including its CDRs,
are available from the authors on reasonable request.
Stated plainly, because they affect what this code can be used for:
- The AlphaFold3 container was referenced as
alphafold3:latest, not by digest, so the exact build used for the April 2026 screens cannot be recovered. SeeSOFTWARE_VERSIONS.md. - Each antibody–antigen pair was predicted with a single seed and a single diffusion sample. Measured run-to-run variability is ~0.03 STRIVE units (max 0.068); differences smaller than ~0.07 units should not be interpreted.
- The 1,000 designed scFv sequences are not yet deposited here as a single FASTA — see
MANIFEST.md.
Please cite the paper and the archive (concept DOI 10.5281/zenodo.22680233, which resolves to the most recent version).