Assessing State-Specific Accuracy of Cofolding Models for Kinases and GPCRs

Supplementary Scripts/Structures Scripts and predicted structures from the co-folding benchmarking study "Assessing State-Specific Accuracy of Cofolding Models for Kinases and GPCRs". Where scripts were prepared per receptor/family, one representative example is provided. Repository Structure supplementary/ ├── 1_Structures/ # All predicted structures │ ├── GPCR_5HT2a/ │ │ ├── Alphafold/ │ │ │ └── <condition>/ # e.g. 5ht2a_inactive_custom_templates_custommsa │ │ │ ├── seed-0_sample-{0..4}/ # individual predictions (.cif, _aligned.pdb, confidences.json) │ │ │ └── *_confidences.json, *_ranking_scores.csv # aggregated outputs │ │ ├── boltz/ │ │ ├── Chai1/ │ │ └── RF3/ │ ├── GPCR_AA2A/ # same layout as above │ ├── GPCR_DOR/ │ ├── Kinase_ALK2_ATP/ │ ├── Kinase_ALK2_RK783/ │ ├── Kinase_LRRK2/ │ └── Kinase_PI3K/ ├── 2_input_preparation/ # JSON/A3M/M8 prep scripts ├── 3_predictions/ # Prediction run scripts ├── 4_alignment/ # PyMOL-based structural alignment ├── 5_ligand_rmsd/ # Ligand RMSD calculation ├── 6_extract_metrics/ # CSV metric extraction ├── 7_core_analysis_plots/ # RMSD bar/dot plots ├── 8_rmsd_scatter_plots/ # Template/MSA/pLDDT scatter plots └── 9_plddt_plots/ # Per-residue and atom-wise pLDDT plots Analysis Pipeline 1) State-biased MSA (MMseqs2) Creates custom kinase-state A3M files. mmseqs createdb <INPUT_FASTA> <QUERY_DB> mmseqs createdb <STATE_FASTA> <STATE_DB> mmseqs search <QUERY_DB> <STATE_DB> <RESULT_DB> <TMP_DIR> --max-seqs 1000 --threads 8 mmseqs result2msa <QUERY_DB> <STATE_DB> <RESULT_DB> <OUTPUT_A3M> --msa-format-mode 5 reformat.pl -M first -r a3m a3m <RAW_A3M> <FIXED_A3M> 2) Input preparation AF3 2_input_preparation/create_aa2a_active_json.py — GPCR JSON prep--fasta_path --smiles_path --a3m_path --templates_dir --output_dir → AF3 .json 2_input_preparation/create_alphafold3_jobs_new_a3m.py — kinase-state JSON prep--smiles --ligand_name --fasta_path --output_dir → AF3 .json RF3 2_input_preparation/normalize_a3m_for_rf3.py<input_a3m> <output_a3m> → normalized .a3m 2_input_preparation/create_rf3_jsons_from_af3.pyInternal BASE path → RF3 .json Chai-1 2_input_preparation/create_m8_gpcr_chai1.py--query-fasta --templates-dir --output-m8 → template-hits .m8 Helper 2_input_preparation/fix_templates_bio.py — retain single chain in template mmCIFs--templates_dir 3) Prediction execution AF3 3_predictions/run_aa2a_alphafold.sh → .cif RF3 rf3 fold inference_engine=rf3 inputs=<RF3_JSON> ckpt_path=<RF3_CHECKPOINT> diffusion_batch_size=5 seed=0 Chai-1 3_predictions/run_all_gpcrs_chai1.sh → .cif 4) Alignment and ligand RMSD 4_alignment/alignment_pymol.py<ground_truth.cif> <predictions_dir> [--save]→ alignment_log.txt, *_aligned.pdb, *_groundtruth.pdb 5_ligand_rmsd/lig_rmsd.py<ground_truth.pdb> <predictions_dir>→ ligand_rmsd_log.txt ALK2 split (ATP/RK783): 4_alignment/alignment_pymol_alk2_separate.py and 5_ligand_rmsd/lig_rmsd_alk2_separate.py. 5) Metric extraction 6_extract_metrics/extract_pdb_data.py — GPCRs (5HT2A, AA2A, DOR)Inputs: hardcoded ligand_rmsd_log.txt, alignment_log.txt, aligned PDB paths→ 5HT2a_GPCR_analysis_core.csv, AA2A_GPCR_analysis_core.csv, DOR_GPCR_analysis_core.csvColumns: PDB_Path, ligand RMSD, protein RMSD, mean protein pLDDT (core residues), mean ligand pLDDT 6_extract_metrics/extract_kinase_data.py — LRRK2, PI3KInputs: hardcoded <PROJECT_ROOT>, target folders, log files, aligned PDBs→ LRRK2_kinase_type2_analysis.csv, PI3K_kinase_allosteric_analysis.csvColumns: same as above; protein pLDDT is whole-chain mean 6_extract_metrics/extract_alk2_data.py — ALK2 (ATP/RK783 split)Inputs: hardcoded <PROJECT_ROOT>/target_data/ALK2_kinase_active, ligand-specific logs→ ALK2_ATP_analysis.csv, ALK2_Inhibitor_analysis.csv GPCR core residue ranges for pLDDT in extract_pdb_data.py: 5HT2A 69–397, AA2A 1–314, DOR 39–335. 6) Plotting 7_core_analysis_plots/plot_core_analysis_subpart_final.pyInputs: <PROJECT_ROOT>/Core-prot-analyses/*.csv→ plots/*_core_analysis_final.png, Core-prot-analyses/after-plotting/*_plotted.csvShows per-condition/method protein RMSD bars + ligand RMSD dots (best replicate by mean protein pLDDT); exports enriched per-model table consumed by steps 7–8. 8_rmsd_scatter_plots/plot_template_vs_ligand_rmsd.py--after-dir --output-dir --rmsd-cutoff (+ optional pLDDT thresholds)Input: *_plotted.csv from Core-prot-analyses/after-plotting/→ scatter .png + summary/method-stats .csvPanels: (i) template category vs ligand RMSD, (ii) MSA category vs ligand RMSD, (iii) protein pLDDT vs ligand RMSD (Huber), (iv) ligand pLDDT vs ligand RMSD (Huber; RF3 excluded) 9_plddt_plots/plot_plddt_values.py--target_data --outputInput: prediction CIF trees under <PROJECT_ROOT>/target_data→ per-target pLDDT figures + summary_report.txtShows method/condition-level pLDDT (best replicate), residue-wise and atom-wise confidence profiles 9_plddt_plots/plot_atomwise_plddt_from_after_csv.py--after-dir [--csv --output-dir --filter-selected]Input: *_plotted.csv with PDB_Path and Method→ <name>_atomwise_plddt_best_single.pngShows per-method residue pLDDT from the single best model, with aggregate background range and GPCR core-region shading