Drug Interactions
Whether a drug's binding pocket survives alternative splicing.
This layer asks a therapeutic question directly: when a gene is alternatively spliced, does the pocket a drug binds still exist in the resulting isoform? We start from 20,091 AlphaFold 3 co-folded drug–protein complexes spanning 3,725 compounds and 2,889 target genes, define each drug's pocket from its actual bound pose, then carry that pocket onto every splice isoform of the target.
Co-folded Complexes
Each complex is a single AlphaFold 3 prediction of the target protein together with the ligand, so the binding pose is modelled in structural context rather than docked after the fact. Confidence is summarised by a holo grade derived from the interface ipTM:
ipTM ≥ 0.75 — high-confidence pose.
0.50 ≤ ipTM < 0.75 — moderate.
ipTM < 0.50 — low; interpret with care.
Each drug page shows the compound's 2D skeleton (from its SMILES) and its real bound conformer lifted from the co-fold, alongside chemistry pulled from PubChem — molecular formula, molar mass, InChIKey, and cross-references. Two per-pocket pLDDT values are reported: the ligand pLDDT (confidence in the drug's placement) and the pocket pLDDT (confidence in the contacting residues).
Contact Residues
A pocket is the set of protein residues within 4.5 Å of any ligand atom in the co-folded structure, stored as 1-based positions on the drugged transcript's sequence. These residues define what splicing must preserve for the drug to still bind.
Pocket Fate Across Isoforms
The pocket is anchored on the drugged transcript and aligned onto each sibling isoform with a splice-aware (BLOSUM62) alignment — never onto an arbitrary canonical, so the comparison is always drug-relative. The fraction of contact residues retained in aligned register gives the fate:
≥ 75% of contact residues retained.
25–75% retained — partial pocket.
< 25% retained — pocket effectively gone.
No-call gate: a pocket is only scored when its residues sit in a confident region of both structures (per-pocket pLDDT ≥ 70). Below that, the fate is reported as no_call rather than guessed — a low-confidence region cannot be said to preserve or lose a pocket. Gene-level summaries are computed over decidable calls only.
The result reproduces known biology independently — for example AR-V7, the treatment-resistant androgen-receptor variant, is scored Lost for its ligand-binding pocket because the C-terminal ligand-binding domain is spliced out.
Where It Appears
- The Drug Pockets lens ranks drugged genes by how subtractive their splicing is for drug binding.
- Each /drugs/[gene] page pairs the drugged-complex structure with the selected isoform, highlighting which contact residues survive.
- Gene and isoform pages carry a drug panel whenever co-fold data exists for that target.
Scope: a pocket call speaks to structural presence, not binding affinity — a preserved pocket means the contact residues are there, not that the drug binds with unchanged potency. Co-fold selections are keyed on protein_md5, so predictions attach to the exact sequence they were computed for.