How to dock a compound against a kinase mutation

The short version

To compare how a compound binds the wild-type and mutant forms of a kinase, you need three things: a clean apo or ligand-bound crystal structure of the wild-type protein, a way to introduce the mutation in silico, and a docking engine that returns a comparable score for both. The Δ score (mutant − wild-type) tells you whether the mutation is predicted to weaken (positive Δ, resistance), strengthen (negative Δ, selectivity), or leave binding unchanged.

On Liganx the entire workflow takes about two minutes from choosing a target to having a side-by-side wild-type/mutant score with a 3D pose, ProLIF interaction fingerprint, and PoseBusters geometric validation. The catalog covers 13 of the highest-impact oncology kinase targets and 40 actionable mutations.

Step 1 — Pick the right PDB structure

Most failed mutation-docking studies fail because of the structure choice, not the docking algorithm. The questions worth asking before you pick a PDB ID:

Is the mutated residue resolved in the structure?

If the residue you want to mutate isn't present in the crystal (a common problem with flexible loops), you can't apply the mutation to it. Liganx pre-flight validation flags this and suggests alternate PDB IDs that resolve the residue.

Is the residue in the binding pocket?

A mutation 30 Å away from the docking box can't possibly affect a Vina-scored pose. Liganx labels this as “outside docking box” and treats Δ ≈ 0 as the expected, honest outcome rather than a failure.

Was the structure solved with a ligand bound?

Apo structures often have a collapsed or rearranged pocket. For drug-discovery work prefer a holo (ligand-bound) structure, then strip the co-crystal ligand before docking.

Is the resolution adequate?

Below 2.5 Å is generally fine for docking. Above 3.0 Å, side-chain placements get noisy and small Δ differences become unreliable.

Step 2 — Apply the mutation

There are two common ways to introduce a point mutation into a docking-ready receptor. PDBFixer-style residue replacement rebuilds the side chain in a default rotamer; FoldX BuildModel additionally minimises the local environment to relieve steric clashes. For gatekeeper mutations like EGFR T790M and BCR-ABL T315I — where a small Thr → Met / Ile substitution introduces a bulky side chain into a tight pocket — the local relaxation matters because an unrelaxed Met side chain can clash with the docking box and produce spurious resistance scores.

Liganx applies PDBFixer for the basic mutation step, then OpenMM-minimises the receptor before running the docking engine. In our published validation suite this reproduces the literature direction on the cases rigid-receptor docking can resolve, with noise-floor and method-limit cases flagged.

Step 3 — Choose a docking engine (or run all three)

The three engines available on Liganx have different strengths:

AutoDock Vina (QuickVina2-GPU)

The standard physics-based scoring function. Fast on GPU, well-validated, and returns scores in kcal/mol that are directly comparable across runs. The right default for most mutation-comparison work.

GNINA

A convolutional neural network rescorer trained on the PDBbind dataset. Often improves pose ranking when Vina returns several near-degenerate poses. Slower than Vina; worth the cost for the final shortlist.

Boltz-2

A 2025 ML co-folding model that predicts the protein-ligand complex geometry directly from sequence and SMILES, then reports a model-predicted (uncalibrated) affinity in log₁₀(IC₅₀ µM). Slowest of the three (~5 minutes per pose) but the only engine that natively handles induced fit. The Imatinib × BCR-ABL T315I gatekeeper test reproduces the published resistance signal (Δ = +0.60).

Step 4 — Read the Δ score honestly

The wild-type vs. mutant Δ is the headline number. A few things worth keeping in mind when interpreting it:

• Δ > +1 kcal/mol is a meaningful predicted resistance signal. Anything smaller is in the noise floor for Vina-style scoring.
• Δ < -1 kcal/mol is a meaningful predicted selectivity signal. Drugs that gain ≥1 kcal/mol of binding to the mutant over the wild-type are the strongest mutant-selective candidates.
• |Δ| < 0.5 kcal/mol means the mutation doesn’t differentiate this compound from the wild-type at the resolution Vina can see. That’s often the right answer — many mutations are silent for many compounds.
• Compare the 3D poses, not just the numbers.Two compounds with the same Δ can have completely different mechanisms. The 3D viewer and ProLIF interaction fingerprint show you what binding mode each compound adopts in each variant.

Worked examples

Five of these are documented in our scientific validation suite with the live job links. Each starts from a published cellular IC₅₀ ratio and reproduces the expected direction of the Liganx Δ score.

BCR-ABL T315I (Imatinib resistance)

The classic gatekeeper mutation. Imatinib binds the inactive DFG-out conformation; Thr315 → Ile blocks the entrance and the published cellular IC₅₀ shifts ~30-fold. Liganx Δ on Boltz-2 reproduces +0.60 in log₁₀(IC₅₀ µM).

EGFR T790M (Gefitinib/Erlotinib resistance)

Acquired resistance to first-generation EGFR inhibitors. The bulky Met side chain partially fills the pocket and increases ATP affinity, indirectly lowering inhibitor binding. Published cellular IC₅₀ shift ~10x; Liganx Δ reproduces the resistance direction.

BRAF V600E (Vemurafenib selectivity)

Vemurafenib was specifically designed to preferentially bind the V600E mutant. The Δ should be negative (mutant-selective). Liganx Δ on the V600E case reproduces the published selectivity once OpenMM minimisation is disabled for this target — see the validation page case study for the discussion.

BTK C481S (Ibrutinib resistance)

Ibrutinib is a covalent inhibitor that bonds to Cys481; the C481S mutation removes the covalent anchor. Vina scores will not capture the loss of covalent bonding — this is a documented method limit, not a Liganx bug.

Common pitfalls

• Using a different PDB for wild-type vs. mutant.The Δ then conflates structural noise with the mutation effect. Always start from the same crystal and apply the mutation in silico.
• Treating |Δ| < 0.5 as a failure.Most mutations are silent for most compounds. A near-zero Δ is information, not a bug.
• Ignoring covalent inhibitors.Vina/GNINA/Boltz-2 are non-covalent scoring functions. For covalent ligands (Ibrutinib, Osimertinib, Sotorasib) the Δ ranks reversible binding only.
• Docking against a low-resolution apo structure.A 3.5 Å apo crystal often has a collapsed pocket; Vina will dock the ligand into a sterically wrong site and return a meaningless Δ. Prefer holo, < 2.5 Å.