MirrorPeptidizer Benchmark

Benchmark data accompanying the Mirror-Peptidizer paper. Mirror-Peptidizer designs D-peptide binders by inverting the chirality of the target receptor, generating backbone poses with Chroma, and optimizing sequences with ProteinMPNN (optionally fine-tuned for synthetic feasibility with Bayesian optimization).

This repository contains everything needed to inspect and reproduce the paper's benchmark analyses:

• Main D/L benchmark — for three pharmacologically relevant targets (PD-L1, MDM2, IL-23R), D-peptide designs (mirror-image pipeline) are compared against L-peptide controls (same pipeline without inversion): ~4,000 candidates per condition across multiple peptide lengths.
• New-target generalization — D-peptide designs for TNF-α, CXCR2 and CXCR4.
• Ablation study — chroma+proteinmpnn (Tier 1 only), chroma+bo (Chroma sequences as BO seeds, no ProteinMPNN), and the full chroma+proteinmpnn+bo pipeline.
• Figure source data (figure_data/), rendered figures (figures/) and the curated Supplementary-Information package (si_landscape_recovery_package/, design-landscape characterisation + ProteinMPNN sequence-recovery test).
• Analysis / plotting scripts and the two-stage pose-filtering implementation cited in the paper's Methods.

The design pipeline itself (Chroma, ProteinMPNN and BO wrappers) lives in the project source code; this repository hosts the benchmark data.

Repository layout

README.md                              # this dataset card
PDL1_D.pdb                             # input complexes (receptor + known active D-peptide);
MDM2_binder.pdb                        #   these carry a "fixed_" filename prefix
MDM2_2.pdb                             #   in the source repository
MDM2_3.pdb
IL23R_D1_2.pdb
PDL1/                                  # main benchmark, one dir per target
  D_peptide/all_results.csv            #   mirror pipeline: ~4000 candidates
  D_peptide/receptor_D.pdb             #   D (mirror-image) receptor used for design
  L_peptide/all_results.csv            #   control pipeline (no inversion)
MDM2/ ...
IL23R/ ...
new_targets/
  cleaned/{TNFalpha,CXCR2,CXCR4}_receptor.pdb     # prepared receptor inputs
  designs/<target>/D_peptide/all_results.csv      # ~340 candidates per target
  designs/new_target_design_summary.csv
ablation/
  ablation_summary.csv                 # raw summary (3 targets x 3 modes)
  ablation_summary_pose_filtered.csv   # summary after two-stage pose filtering (paper)
  ablation_tradeoff_pose_filtered.csv  # score / synthesis-penalty trade-off (paper)
  <target>/selected_poses.csv
  <target>/chroma_proteinmpnn/all_results.csv
  <target>/chroma_bo/all_bo_results.csv
  <target>/chroma_proteinmpnn_bo/all_bo_results.csv
ld_benchmark_summary.csv               # D/L benchmark headline summary
dl_*.csv                               # D/L sequence similarity / composition tables
buried_pose_inventory.csv              # poses rejected by the geometric surface filter
geometric_pose_classification.csv      # per-pose geometric metrics
filter_comparison_polypro_vs_geometry.csv
poly_alanine_diagnosis.csv
filtered_bad_poly_p_poses.csv
figure_data/                           # source CSVs for every paper figure + README
figures/                               # rendered figures (PDF/PNG/SVG)
si_landscape_recovery_package/         # self-contained SI package (figures, CSVs,
                                       #   representative PDBs, tables, MANIFEST)
recovery_test/mirror_axis_equivalence* # mirror-axis equivalence check (data + script)
scripts/                               # analysis / figure-generation scripts
utils/pose_filtering.py                # two-stage pose filter (cited in Methods)
structures/                            # raw generated structures, packed per experiment
  <target>_<condition>.tar.gz          #   one archive per experiment, see below
  ablation_<target>_<mode>.tar.gz
  ablation/<target>/chroma_proteinmpnn/tier1_source.txt

Raw structures (structures/)

The generated structure PDBs are packed into per-experiment tar.gz archives (~80,000 PDB files in total) to keep the repository browsable. All paths inside an archive are relative to the experiment directory, so the filename column of all_results.csv maps onto archive contents as <experiment-dir>/<relative-path>.

• Main benchmark and new-target archives (PDL1_D_peptide.tar.gz, ..., CXCR4_D_peptide.tar.gz) contain, per length (len_<n>/):
  • Poses/*.pdb — Chroma-generated receptor+binder backbone poses;
  • Binders/*.pdb — ProteinMPNN-designed binder structures built on those poses (all-D structures in D_peptide conditions);
  • results.csv — the per-length candidate table (a subset of the uploaded all_results.csv). In D_peptide conditions the receptor side of every structure is the mirror-image (D) receptor (receptor_D.pdb, uploaded at the dataset root of each target).
• Ablation archives (ablation_<target>_chroma_bo.tar.gz, ablation_<target>_chroma_proteinmpnn_bo.tar.gz) contain the BO run folders: len_<n>/pose_<i>/BO/Eval_PDBs/*.pdb (BO-evaluated structures per round) and BO bookkeeping files.
• ablation/<target>/chroma_proteinmpnn/tier1_source.txt — the chroma+proteinmpnn ablation mode generates no structures of its own; it re-scores the Tier-1 poses of the main benchmark (this file records which source directory was used).

Example:

mkdir -p PDL1/D_peptide && tar -xzf structures/PDL1_D_peptide.tar.gz -C PDL1/D_peptide

Data dictionary

all_results.csv (main benchmark and new targets; one row per designed candidate):

column	meaning
length	peptide length (number of residues)
pose	index of the Chroma-generated backbone pose
sequence	designed sequence in one-letter amino-acid code; in D_peptide conditions the actual binder is the all-D enantiomer of this sequence
score	ProteinMPNN score of the design (lower is better)
filename	path of the corresponding binder PDB on the generation machine; the same relative path is contained in the matching structures/*.tar.gz archive

Ablation all_bo_results.csv additionally contains the Bayesian-optimization record: Variants (sequence), Fitness (BO objective), MPNN_score, ~50 synthetic-feasibility descriptors (spps_*, hydrophobic/aromatic/charge fractions, retention index, ...), synthesis_penalty, synthesis_feasibility_score, synthesis_risk_class, plus score, length, pose.

Candidate filtering (as used in the paper)

Reported analyses apply a two-stage structural filter, implemented in utils/pose_filtering.py:

1. Geometric surface filter — a pose is rejected when the binder is buried inside the receptor (radial percentile ≥ 0.60, outward-nearby fraction ≤ 0.25, occupied octants ≤ 5, binder contact atoms ≥ 20). The rejected poses are listed in buried_pose_inventory.csv.
2. Sequence-complexity filter — designs with any homopolymer run ≥ 6 (e.g. poly-Pro / poly-Ala / poly-Gly) are dropped.

For D/L score-distribution comparisons the L-peptide candidates are additionally downsampled (numpy.random.default_rng(seed=42)) to match the D-peptide count; the filtered candidate sets are provided in figure_data/fig_dl_score_dist_*.csv and the downsampled summaries in figure_data/fig_dl_score_dist_summary.csv.

Reproducing

# main benchmark (requires the full Mirror-Peptidizer source checkout)
python run_benchmark.py --gpu 0
python analyze.py

# figure source data / figures
python scripts/export_figure_data.py
python scripts/plot_ablation_dl_figures.py
python scripts/plot_new_target_score_distribution.py

run_benchmark.py, run_ablation.py and run_new_targets.py import the design pipeline from the project source tree (utils.chroma_sample, utils.protein_mpnn, ...) and therefore need the full repository, not only this dataset. The filtering/figure scripts listed above run from this repository alone (pandas / numpy / scikit-learn / matplotlib).

License

Apache License 2.0.

Citation

If you use this benchmark, please cite the Mirror-Peptidizer paper (citation to be added upon publication).