Predicting drug-target residence times (koff) from structure within a factor of three
Rare-event sampling plus force-field error and no prospective benchmark make this harder than co-folding generalisation (47) and close to scaffold-agnostic affinity (63) in difficulty. Application value 42: concrete benefit to kinetics-driven optimisation, much smaller than affinity (76) because koff is usually measured. Leverage 28 as a drug-discovery-specific capability that reuses rather than unlocks other methods.
The problem
Predicting how long a drug stays stuck to its target before it lets go.
- 1Drug sits in pocket
- 2Rare event: drug escapes
- 3Extra push speeds up simulation
- 4Convert to real-world rate
- 5Compare with measured residence time
Kinetics versus thermodynamics. Binding strength is . Two drugs with the same can have different . Residence time has been linked to duration of action in vivo (Copeland and colleagues, 2006).
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- A transferable exchange–correlation functional free of delocalization and self-interaction errorComputational Chemistry · D72
- Predicting fold-switching proteins and both of their folds from sequenceComputational Medicine · D58
- Step 2
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsComputational Chemistry · D52
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
- Step 3
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Gravitational N-body simulationAlgorithm
- Step 2
- Molecular dynamics for biomoleculesAlgorithm · host
- Then
- Predicting drug-target residence times (koff) from structure within a factor of threeThe problem
State of the art
- Enhanced-sampling MD. Residence times are typically orders of magnitude longer than directly simulable timescales, so methods use biased or accelerated dynamics. tauRAMD (random acceleration MD; Kokh et al.) estimates relative residence times from the time to dissociation under a small random force, and is reported to give useful rank correlations within congeneric series for targets such as HSP90 and GPCRs (no single correlation value is quoted here, as reported performance varies by target and protocol); metadynamics, weighted-ensemble and milestoning variants give pathways and absolute rates in favourable systems (reviewed in Frontiers in Molecular Biosciences 2022 and the Annual Review of Physical Chemistry 2019).
- Absolute rates. Absolute predictions within a factor of three remain rare and depend strongly on force-field and water-model choice; most successful reports are retrospective, on one or a few targets, and with protocols tuned on those targets.
- Machine learning. ML on tauRAMD trajectory features and on structure-based descriptors improves correlation in-series (Kokh et al., 2019), but, as reviewed for GPCRs, scarce kinetics data are the main bottleneck.
- Generative ensemble emulators (BioEmu, AlphaFlow) sample equilibrium ensembles, not transition kinetics; no emulator is known to predict unbinding rates.
- Not shown. No prospective, multi-target blinded evaluation has demonstrated the target accuracy; no community benchmark analogous to D3R for kinetics is in wide use as of October 2026.
Realistic applications
Residence time often tracks in vivo efficacy and duration of action better than equilibrium affinity, particularly for slowly dissociating covalent-like or induced-fit inhibitors; predicting it would enable kinetics-guided optimisation (selectivity via off-rates, durable target occupancy at lower dose, safety) before synthesis.
What counts as solved
A refereed prospective evaluation, with experimental measured after predictions were frozen (SPR, TR-FRET, jump dilution) for at least 100 ligands on at least 6 target classes, in which a single protocol (not tuned per target) satisfies the factor-of-three and rank-correlation criteria above at a stated compute cost.
How it gets verified
ObservationSettled by new measurements or observations.
Needs new k_off measurements (SPR, TR-FRET or jump dilution) taken after predictions are frozen, for at least 100 ligands across at least 6 target classes: a single untuned protocol must meet the factor-of-three and rank-correlation criteria at a stated compute budget.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Rare-event simulation cost dominates, with force-field accuracy and scarce kinetics data as the other walls.
Bottlenecks
ranked by votes- #1ComputeSampling rare unbinding events
Residence times of minutes to hours exceed direct simulation by 6+ orders of magnitude; enhanced sampling needs good collective variables and many replicas.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Jun 1, 2022Review of koff simulation methodssource
Frontiers review surveys simulation methods, force-field effects and HPC/ML prospects for computing dissociation rates.
- Jan 1, 2019ML on tauRAMD trajectoriessource
Trajectory features from random-acceleration MD improve residence-time models for GPCR and HSP90 ligands.
- Jan 1, 2006Residence time as a drug propertysource
Copeland, Pompliano and Meek argue that drug-target residence time rather than affinity can determine in vivo duration of action.