Covalent docking: how it differs from the non-covalent kind

Most molecular docking rests on one quiet assumption: the ligand binds reversibly, and the score you compute is a non-covalent binding free energy. That assumption is wrong for a growing slice of the oncology drug list. Sotorasib, osimertinib, ibrutinib, afatinib — these all form an actual chemical bond with their target. Run them through standard docking and the score means very little. Covalent docking treats the bond as the whole point.

Why standard docking gets covalent binders wrong

A reversible docking score estimates the free energy of a non-covalent equilibrium: van der Waals contact, hydrogen bonds, electrostatics, desolvation. A covalent inhibitor’s binding is dominated by something that scoring function does not model at all — the formation of a covalent bond, typically between an electrophilic “warhead” on the ligand and a nucleophilic residue on the protein, most often a cysteine thiol.

Worse, nothing in a standard docking run forces the warhead to point at that cysteine. The docker will happily return a high-scoring pose with the warhead buried on the wrong side of the pocket: geometrically plausible, chemically impossible. The number looks fine and the pose is nonsense.

The warhead

A warhead is the reactive group that makes the bond. The most common one in approved drugs is the acrylamide, a Michael acceptor that reacts with cysteine thiols. Osimertinib, ibrutinib, and afatinib all carry acrylamide warheads; sotorasib and adagrasib carry acrylamide-type warheads aimed at the mutant cysteine of KRAS G12C. Other warhead chemistries exist — nitriles, β-lactams, epoxides, activated esters — but cysteine-targeting Michael acceptors dominate oncology.

The binding event is really two steps: a reversible recognition step, where the molecule docks into the pocket like any other ligand, then an irreversible chemical step, where the warhead reacts. Covalent docking has to respect both.

How covalent docking tools actually work

Covalent docking tools handle the two steps by tethering the ligand. You tell the software which protein residue is the nucleophile and which ligand atom is the warhead; it then constrains that atom pair to bonding distance and samples poses around the tether. A few established approaches:

• AutoDock covalent — Bianco et al. (2016) describe two variants, a two-point attractor method and a flexible side chain method. The flexible side chain version recovered the experimental pose in 75% of a 20-complex training set.
• CovDock — Zhu et al. (2014) combine Glide docking with Prime structure modeling in a parameter-free workflow; 76% of test inhibitors landed within 2.0 Å RMSD of the crystallographic pose.
• GNINA covalent mode — extends the CNN-scored docking engine to handle covalent constraints, so the same learned pose-quality model carries over.

Scarpino et al. (2018) benchmarked six covalent docking tools across a large, diverse complex set and found 40–60% of top-scoring poses within 2.0 Å RMSD, rising to 50–90% when the best of the top ten was counted. Success was meaningfully higher for Michael additions than for ring-opening reactions — the warhead chemistry matters to how well the method behaves.

What covalent docking does and does not tell you

Covalent docking gives you a credible bound pose and answers geometric questions well: can the warhead actually reach the cysteine, does the recognition scaffold fit the pocket, which mutant cysteine does a candidate prefer. What it does not give you is reactivity. Real covalent potency is governed by k_inact/K_i — the intrinsic chemical reaction rate divided by the reversible binding affinity. Docking can speak to the K_i side and to geometry; it does not predict k_inact.

A useful sanity check, noted by Scarpino et al.: non-covalent docking into a protein with the target cysteine mutated to alanine reproduced binding modes almost as well as full covalent docking, at a fraction of the computational cost. The lesson is that much of the value lives in getting the recognition pose right — the covalent constraint refines it rather than rescuing it.

Try the docking yourself

Open Studio and dock a warhead-bearing candidate against KRAS with the G12C mutation, or against EGFR — both are canonical covalent targets with a reactive cysteine in the pocket. Look at where the warhead lands relative to that cysteine: a pose that buries the acrylamide away from the thiol is telling you the recognition scaffold is fighting the chemistry. Then compare the ΔΔ between the mutant receptor (reactive cysteine present) and the wild-type receptor (no reactive cysteine) to see the selectivity the warhead is supposed to buy you.

Liganx puts molecular docking online and free in the browser, so you can test a warhead-bearing candidate against a reactive- cysteine target without a local install.

Primary sources

• Zhu K, et al. Docking Covalent Inhibitors: A Parameter Free Approach To Pose Prediction and Scoring. J Chem Inf Model 54, 1932–1940 (2014). doi:10.1021/ci500118s
• Bianco G, Forli S, Goodsell DS, Olson AJ. Covalent docking using AutoDock: Two-point attractor and flexible side chain methods. Protein Sci 25, 295–301 (2016). doi:10.1002/pro.2733
• Scarpino A, Ferenczy GG, Keserú GM. Comparative Evaluation of Covalent Docking Tools. J Chem Inf Model 58, 1441–1458 (2018). doi:10.1021/acs.jcim.8b00228