Datasets:

Tc-43
/

FKBP12_F36V_Binder_Designs_1BL4

license: cc-by-4.0
pretty_name: FKBP12 F36V Binder Designs (PDB 1BL4)
task_categories:
  - tabular-regression
  - graph-ml
  - other
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: designs.csv
language:
  - en
tags:
  - chemistry
  - biology
  - drug-discovery
  - FKBP12
  - FKBP1A
  - FKBP
  - F36V
  - FKBP12-F36V
  - P62942
  - immunophilin
  - PPIase
  - peptidyl-prolyl-isomerase
  - bump-hole
  - bumped-ligand
  - chemical-biology
  - chemically-induced-proximity
  - induced-proximity
  - dTAG
  - degron
  - degradation-tag
  - AP1497
  - SLF
  - synthetic-ligand-of-FKBP
  - rapamycin
  - FK506
  - tacrolimus
  - chemical-inducer-of-dimerization
  - CID
  - molecular-glue
  - bifunctional
  - heterobifunctional
  - PROTAC
  - RIPTAC
  - targeted-protein-degradation
  - protein-protein-interaction
  - ternary-complex
  - 1BL4
  - molecular-docking
  - autodock-vina
  - protein-ligand-complex
  - structure-based-drug-design
  - de-novo-design
  - generative-design
  - generative-ai
  - ai-generated
  - genetic-algorithm
  - GA-II
  - synthetic-data
  - small-molecule
  - SMILES
  - SDF
  - cheminformatics
  - computational-chemistry
  - virtual-screening
  - TC-43
  - technetium

FKBP12 F36V Binder Designs (PDB 1BL4)

165 de novo, generative-AI–designed small molecules docked into the engineered FKBP12 F36V cavity of 1BL4 (1.9 Å), each supplied as a complete protein–ligand complex.

Molecules were generated by the Technetium TC-43.ai engine (GA-II generation). Target setup, receptor validation and dataset curation by Claude Code.

Target

FKBP12 (P62942) is a 12 kDa peptidyl-prolyl isomerase and the receptor for FK506 and rapamycin. The F36V mutation enlarges the ligand cavity, creating a "hole" that accepts correspondingly "bumped" synthetic ligands while wild-type FKBP12 does not — the basis of the bump-hole systems used throughout chemically induced proximity, including dTAG degron tagging and heterobifunctional proximity-inducing molecules.

PDB 1BL4 is FKBP mutant F36V complexed with a remodeled synthetic ligand — the reference bumped ligand AP1497 (693.8 Da). The receptor here is the deposited dimer: chains A and B, 107 residues each (mature FKBP12, initiator Met removed, so the mutation sits literally at position 36 and receptor numbering is UniProt − 1). AP1497 is retained in the chain B cavity; the designs occupy the chain A cavity.

Binding site

All 165 designs bind the chain A FKBP cavity. Contact frequencies across the set:

residue	% of poses	residue	% of poses
Trp59	100%	Asp37	77%
Tyr82	100%	Val36	70%
Ile56	100%	Glu54	60%
Val55	100%	Phe99	47%
Tyr26	98%	Ile91	27%
Ile90	95%	Gln53	15%
Phe46	87%	Gly28	15%

Trp59, Tyr82, Ile56 and Val55 form the invariant floor of the cavity — every pose touches all four.

Val36 engagement — read this before selecting compounds

Val36 is the mutation. It is the only residue that distinguishes this receptor from wild-type FKBP12, so a design that does not contact it has no structural basis for preferring F36V over wild type. Wild-type FKBP12 is one of the most abundant cytosolic proteins in mammalian cells, which makes that distinction the difference between a usable chemical-biology tool and one swamped by the endogenous pool.

val36_contact = True — 115 of 165 designs (70%). Use this subset where F36V discrimination matters.
The remaining 50 bind the conserved cavity floor only and should be treated as generic FKBP12 binders.

This dataset does not include wild-type FKBP12 counter-docking; Val36 contact is a structural proxy, not a measured selectivity.

file	description
`designs.csv`	one row per design — 29 columns
`ligands.sdf`	all 165 poses, bond orders and formal charges assigned from the SMILES template
`receptor_1BL4_FKBP12_F36V.pdb`	chains A + B plus the AP1497 reference ligand
`structures/`	165 complete protein–ligand complexes

Columns

design_id, rank, smiles, pose_smiles, stereo_check, vina_score, ligand_efficiency, mw, clogp, tpsa, qed, hbd, hba, rotatable_bonds, heavy_atoms, rings, aromatic_rings, fsp3, formal_charge, murcko_scaffold, chemotype_cluster, n_contact_residues, contact_residues, val36_contact, min_dist_protein_ang, min_dist_ap1497_ang, steric_clash, ecfp4_similarity_to_ap1497, pose_file

Set characteristics

	designs (mean)	AP1497 reference
Vina score	−10.56 (best −12.30)	—
ligand efficiency	0.34	—
molecular weight	433	694
heavy atoms	31	50
cLogP	3.17	5.99
TPSA	87.6	139.3
rotatable bonds	6.3	17
QED	0.44	0.17

All 165 SMILES are unique, spanning 142 Murcko scaffolds and 66 chemotypes (Butina on ECFP4 at Tanimoto 0.4), with 46 singletons and a largest cluster of 33.

The designs are chemically independent of the reference. Maximum ECFP4 similarity to AP1497 across the set is 0.19 (mean 0.13), and none of the 165 contains a pipecolate ester — the canonical FKBP-binding motif shared by FK506, rapamycin, SLF and AP1497. The set explores the cavity with different chemistry rather than re-deriving the known pharmacophore. Whether that is an advantage or a liability depends on how much the pipecolate anchor contributes to the binding you need.

Pose quality

Docking was rigid-receptor with a fixed receptor across the whole set (verified: a single set of receptor and AP1497 coordinates in all 165 files). Poses are sterically clean:

closest heavy-atom contact to protein: median 2.97 Å (5th–95th percentile 2.67–3.11)
only 3 of 165 poses have any atom within 2.5 Å of the protein (steric_clash)
zero poses clash with AP1497 in the adjacent chain B cavity
all 165 ligands round-tripped 3D pose → bond orders → SMILES with matching connectivity

Scores were pre-filtered upstream at −10.0 kcal/mol, so the set contains nothing above that value and the distribution describes the surviving tail only. Treat scores as a ranking signal, not affinity.

Citation

Molecules generated by the Technetium TC-43.ai engine. Receptor validation and dataset curation by Claude Code. Released under CC-BY-4.0.