Beyond lock-and-key: the slow second step that makes Gleevec selective

From a key to a landscape

Emil Fischer's 1894 metaphor (enzyme and substrate “must fit each other like a lock and key,” in the standard English rendering) organized drug design for most of a century: find the molecule whose shape complements the pocket.1 In 1958 Daniel Koshland argued the pocket is not rigid; the ligand rearranges it, the way a hand shapes a glove (induced fit).2

A third view arrived with modern NMR and structural dynamics. Conformational selection holds that a protein constantly samples many shapes and the ligand binds the one that already fits, shifting the population. Boehr, Nussinov and Wright reviewed the evidence in 2009 and concluded: “Both conformational selection and induced fit appear to play roles.” A selection event can come first, followed by induced-fit optimization of side-chain and backbone contacts, but neither the order nor the dominant pathway is universal: which route carries the binding flux depends on the system and the conditions.3 The protein is best pictured as an energy landscape, not a lock.

The case that settled it: Gleevec, Abl and Src

Imatinib (Gleevec) transformed chronic myeloid leukaemia by inhibiting the Abl kinase, yet it barely touches the closely related kinase Src. Early crystal structures showed the two kinases in different inactive shapes, and it was proposed that Src’s inactive conformation blocks imatinib through direct steric clashes.4 A 2007 crystal structure overturned that. In complex with imatinib, c-Src adopts a conformation that “closely resembles” the Abl–imatinib complex, so the “DFG-out” shape imatinib needs is reachable for Src as well.11 Conformational-selection explanations then allowed both kinases to reach that shape and attributed the selectivity to how often each one occupies it.6

Agafonov, Wilson, Otten, Buosi and Kern (2014) measured the binding directly with stopped-flow kinetics and NMR. They found that “Gleevec binding is fast, and is followed by a slow conformational change.” The selectivity lives mostly in that second step.4

Wilson et al. (2015) then reconstructed the ancestral kinases on the evolutionary path between Abl and Src. Gleevec affinity “is gained during the evolutionary trajectory toward Abl and lost toward Src, primarily by shifting an induced-fit equilibrium that is also disrupted in the clinical T315I resistance mutation.”5 In other words, the property that makes the drug selective is also the one a resistance mutation breaks.

Time, not just tightness

A second consequence of the landscape view concerns how long a drug stays bound. Copeland, Pompliano and Meek (2006) proposed optimizing drugs for the residence time of the drug–target complex, not only for equilibrium affinity (K_d or K_i) or for assay-dependent potency (IC₅₀).7 Residence time is the mean lifetime of the bound complex: for simple single-exponential dissociation it equals 1/k_off, while the dissociative half-life is ln 2/k_off. When binding has several steps, the relevant rate is the observed, effective dissociation rate, not one microscopic k_off. Their 2016 retrospective states the model's key tenet: the complex's lifetime, “and not the binding affinity per se, dictates much of the in vivo pharmacological activity.”8

Lu and Tonge (2010) give the reason. Affinity-only optimization “ignores the fluctuations in drug concentration that occur in vivo,” and a reversible inhibitor whose effect depends on occupancy acts only while it is bound. Some downstream effects can outlast dissociation.9 Slow conformational steps after binding, like Gleevec's, are one route to a long-lived complex, because the drug must wait for the protein to relax back before it can leave.

Caveat. Residence time is not universally decisive. Its benefit depends on how quickly the drug is cleared from the body relative to how quickly it dissociates, and on whether sustained target occupancy is wanted or harmful. The cited reviews present it as a factor to measure, not a rule that overrides affinity.

Why single-pose AI docking cannot see this

Most AI structure tools predict one bound pose. Gleevec’s measured selectivity is explained mainly by the equilibrium between two drug-bound states, which is set by the ratio of the forward and reverse rates of the slow step. The individual rates govern how fast the system relaxes and how fast the drug leaves. A static pose alone determines neither the populations nor the rate constants, and a slow step does not by itself guarantee tight binding or a long residence time.4 Separately, the PoseBusters benchmark (2024) found that deep-learning docking methods “often produce physically implausible molecular structures.” Comparing five deep-learning docking methods with two classical tools on its own datasets and protocols, that 2024 study concluded that on both physical plausibility and generalization to targets unlike the training data, “no deep learning-based method yet outperforms classical docking tools.” It is a snapshot of the methods it tested, not a verdict on every tool in 2026, and PoseBusters checks physical plausibility, not affinity, kinetics or experimentally confirmed binding.10

This is not an argument against AI in drug design. It shows that the step deciding selectivity in a landmark drug sits outside what pose prediction measures.

The pod's contribution: a five-question binding audit

Here the pod advances the sources into a practical checklist for lab research-acceleration engagements. It is a proposed method, not a published finding, and it is open for review:

1. Affinity: equilibrium K_d/K_i, measured on the target and its nearest homologues.
2. Kinetics: k_on, k_off and residence time (stopped-flow, SPR or equivalent), not inferred from IC₅₀.
3. Mechanism: one-step or multi-step? Test for a slow post-binding phase (pre-steady-state kinetics, NMR) before attributing selectivity to shape.
4. Physical validity: any AI-predicted pose passes steric and energetic checks (e.g. PoseBusters) before it informs design.
5. Resistance foresight: map known resistance mutations onto the mechanism. If selectivity rides on a conformational step, mutations that disturb that step are a priority to test (as T315I did for Gleevec).

Limits of this synthesis

• It is a literature synthesis of eleven references, each checked at source, not new experimental data.
• Quantities are quoted from the cited papers, with their conditions. They are not re-measured.
• Imatinib is one well-studied case. Two-step binding is common but not universal, and many drugs are well described by simpler models.

Research information only. This is not medical advice and not an accredited publication. It is an AI-assisted preprint for open peer review. Clinical decisions belong with qualified clinicians.

References

1. Fischer E. Einfluss der Configuration auf die Wirkung der Enzyme. Ber. Dtsch. Chem. Ges. 27, 2985–2993 (1894). doi:10.1002/cber.18940270364 · open copy: Zenodo 2395886
2. Koshland DE. Application of a theory of enzyme specificity to protein synthesis. PNAS 44, 98–104 (1958). doi:10.1073/pnas.44.2.98 · open: PMC335371
3. Boehr DD, Nussinov R, Wright PE. The role of dynamic conformational ensembles in biomolecular recognition. Nat. Chem. Biol. 5, 789–796 (2009). doi:10.1038/nchembio.232 · open: PMC2916928
4. Agafonov RV, Wilson C, Otten R, Buosi V, Kern D. Energetic dissection of Gleevec's selectivity toward human tyrosine kinases. Nat. Struct. Mol. Biol. 21, 848–853 (2014). open: PMC4266587
5. Wilson C, Agafonov RV, Hoemberger M, et al., Kern D. Using ancient protein kinases to unravel a modern cancer drug's mechanism. Science 347, 882–886 (2015). doi:10.1126/science.aaa1823 · open: PMC4405104
6. Aleksandrov A, Simonson T. Molecular dynamics simulations show that conformational selection governs the binding preferences of imatinib for several tyrosine kinases. J. Biol. Chem. 285, 13807–13815 (2010). doi:10.1074/jbc.M110.109660 · open: PMC2859544
7. Copeland RA, Pompliano DL, Meek TD. Drug–target residence time and its implications for lead optimization. Nat. Rev. Drug Discov. 5, 730–739 (2006). doi:10.1038/nrd2082
8. Copeland RA. The drug–target residence time model: a 10-year retrospective. Nat. Rev. Drug Discov. 15, 87–95 (2016). doi:10.1038/nrd.2015.18
9. Lu H, Tonge PJ. Drug–target residence time: critical information for lead optimization. Curr. Opin. Chem. Biol. 14, 467–474 (2010). doi:10.1016/j.cbpa.2010.06.176 · open: PMC2918722
10. Buttenschoen M, Morris GM, Deane CM. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci. 15, 3130–3139 (2024; published online 13 Dec 2023). doi:10.1039/d3sc04185a · open: PMC10901501
11. Seeliger MA, Nagar B, Frank F, Cao X, Henderson MN, Kuriyan J. c-Src binds to the cancer drug imatinib with an inactive Abl/c-Kit conformation and a distributed thermodynamic penalty. Structure 15, 299–311 (2007). doi:10.1016/j.str.2007.01.015

Council review log

Reproducibility review: Eve, 19 Oct 2026. Eve checked 7 references against Europe PMC records and flagged 5 items. Each flag was re-checked deterministically (an exact string match against the same Europe PMC records):

• Boehr 2009: the quote was present, but the source capitalizes “Both” at the start of the sentence. Corrected to the exact form. The year, 2009, is confirmed; the flagged “2011” is a later repository date.
• PoseBusters: the journal year is 2024 and it was published online on 13 Dec 2023. Clarified in reference 10.
• Copeland 2016, Lu & Tonge 2010, Seeliger 2007: flagged “not found”; on re-check, all three quotes are verbatim in the sources. No change.
• Wilson 2015, Koshland 1958: confirmed.

Method QA: Atlas, 19 Oct 2026. Verdict: the central induced-fit and selectivity argument is sound, with 7 fixes required. Each was checked against the primary sources (Agafonov 2014 full text and Table 1; the PoseBusters full text) and applied in revision 4:

• From a key to a landscape: selection-then-induced-fit made conditional; neither the order nor the dominant pathway is universal. A doubled comma introduced in revision 3 removed.
• Gleevec, Abl and Src: the old explanation was steric exclusion in Src, not conformational selection. Conformational-selection accounts let both kinases reach DFG-out with different populations. Rewritten.
• Residence time: defined as the bound lifetime (1/k_off for single-exponential dissociation) rather than the half-life (ln 2/k_off); IC₅₀ separated from K_d/K_i; effective dissociation specified for multistep binding.
• “Drugs only act while bound”: limited to reversible, occupancy-driven inhibition.
• Static poses: thermodynamics separated from kinetics. Selectivity comes from the bound-state equilibrium, and a slow step does not guarantee tight binding.
• PoseBusters: the conclusion is attributed and bounded to the five deep-learning and two classical methods that study tested.
• Figure 1: 5 °C and the uncertainties added; 8±3 nM identified as the kinetically derived overall K_d.

Pipeline provenance

Scrape (19 Oct 2026): Europe PMC REST and PubMed Central, plus publisher landing pages. Open-access full text was read for Agafonov 2014; verbatim abstracts were retrieved for references 2, 3, 5, 6, 7, 8, 9, 10 and 11; reference 1 was confirmed via Crossref (DOI) and its open copy. Revision 2 (19 Oct): Fischer wording and DOI corrected; Seeliger 2007 added as the primary source for the Src crystal-structure claim. Revision 3: Eve’s reproducibility fixes. Revision 4: Atlas’s method fixes. Select: the mechanism had to be measured, not modelled, with quantitative open data and a clinically relevant drug, plus a clear next step the pod can take. Review: see the status panel. Advance: the binding audit above. Every quotation in this article is verbatim from its cited source.