Docking Game in Molecular Docking

• Docking game is a multifaceted concept that reformulates molecular docking as interactive puzzles and multi-agent games.
• It spans educational tools like Bioblox2.5D, two-player Loop Self-Play for flexible binding, and cooperative games for multimeric rigid docking.
• These frameworks simplify complex docking physics by using structured actions, iterative feedback, and scalar scoring to assess binding quality.

Docking game denotes a set of related constructs in molecular docking research in which docking is recast as an explicitly interactive system. In one usage, it is an educational puzzle that translates protein docking into drag, rotation, charge matching, and score feedback. In another, it is a formal game-theoretic framework in which ligand and protein modules, or multiple protein chains, act as players whose updates seek accurate binding modes or stable assembly equilibria. Across these usages, the shared idea is to represent docking through structured actions, intermediate feedback, and a scalar criterion—such as a percent match, a player-specific loss, or a game potential—that stands in for molecular complementarity or stability (Leymarie et al., 2022, Zhang et al., 7 Aug 2025, Somnath et al., 2023).

1. Scope and nomenclature

The term appears in three distinct but related contexts in the cited literature.

Usage	Docking task	Core representation
Bioblox2.5D	Educational protein docking	Interactive “slice-and-fit” puzzle
The Docking Game	Blind flexible protein–ligand docking	Two-player game with Loop Self-Play
DockGame	Multimeric rigid protein docking	Cooperative game over $(SO(3)\times T(3))^N$

In "Bioblox 2.5D -- Developing an Educational Game Based on Protein Docking" (Leymarie et al., 2022), docking is rendered as a touch-based educational puzzle aimed at teenagers. The game focuses on three learning outcomes: the three-dimensional shape complexity of proteins, the role of electrostatic charges in guiding molecular association, and the nature of a docking score as a proxy for binding affinity.

In "The Docking Game: Loop Self-Play for Fast, Dynamic, and Accurate Prediction of Flexible Protein--Ligand Binding" (Zhang et al., 7 Aug 2025), docking is formalized as a two-player game between a ligand player and a protein player. The associated Loop Self-Play algorithm alternates cross-player exchange with per-player self-refinement and is accompanied by a convergence argument to a Nash equilibrium.

In "DockGame: Cooperative Games for Multimeric Rigid Protein Docking" (Somnath et al., 2023), docking is treated as a cooperative game among $N$ proteins or sub-assemblies. Stable assemblies are modeled as equilibria of an underlying potential, approached either through surrogate-potential optimization or through diffusion-based sampling on the manifold of rotations and translations.

A plausible implication is that "docking game" in current literature is not restricted to literal gameplay. It also names a family of optimization viewpoints in which docking becomes a multi-agent decision problem.

2. Bioblox2.5D as an educational docking game

Bioblox2.5D was developed through a collaboration between the Computing Department at Goldsmiths, University of London, and the Structural Bioinformatics group at Imperial College London. The Imperial team provided the content requirements and validated the technical solution, while the Goldsmiths team designed and implemented the requirements into an educational puzzle game intended to support teaching and motivate students to engage with biology (Leymarie et al., 2022).

The game embodies molecular docking as an interactive “slice-and-fit” puzzle. Source data for each docking pair comes from the Protein Data Bank, where the two chains are already in their experimentally determined complex. Within a 3D-modelling environment, each chain is cut exactly at its primary interface, yielding complementary halves. These are rendered as 3D surface meshes and then flattened to form 2D slices. In play, the slices retain the silhouette and visible protrusions or recesses of the original surface. Players drag them into a docking window and rotate them so that convex and concave regions interlock, echoing the lock-and-key metaphor.

This representation is a deliberate simplification. The real docking problem requires continuous fitting in three rotational and three translational degrees of freedom, whereas Bioblox2.5D reduces this to two-dimensional translation plus discrete $0^\circ$, $90^\circ$, $180^\circ$, or $270^\circ$ rotation. Complementarity is then scored by comparing the placed slice to a hidden target outline through a pixel-overlap or contour-distance measure converted into an on-screen percentage score.

Electrostatic charges are introduced through an additional layer of annotation. A short Perl script identifies all inter-chain salt bridges, defined as charged side-chain atom pairs within a threshold distance, and records their 3D coordinates on the surface of each protein. In the game, each charged locus is represented by a small colored sphere attached to the 2D slice: blue for positively charged residues and red for negatively charged ones. Tutorial levels initially hide these spheres; later levels reveal them so that charge complementarity becomes an additional docking cue. When opposite charges overlap they attract and boost the docking score, while same-sign or non-aligned charges trigger a repulsion animation and a sharp error sound cue.

The paper explicitly contrasts the educational score with a classical electrostatic term,

$$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$

which in docking algorithms is typically combined with van der Waals, desolvation, and entropy components. Bioblox2.5D replaces this with a continuously updated “accuracy” metric:

$$S(\%)=100\times\left[\alpha_{\rm shape}\cdot\left(1-\frac{d_{\rm contour}}{d_{\max}}\right)+\alpha_{\rm charge}\cdot\left(1-\frac{\sum_c \min(d_c,d_0)}{d_0\,N_c}\right)\right].$$

Here $d_{\rm contour}$ is the average distance between candidate and hidden outline, the sum runs over annotated charge spheres, and $\alpha_{\rm shape}+\alpha_{\rm charge}=1$. In practice, the engine discretizes both $N$0 and $N$1 into bins, recomputes $N$2 each frame, and displays it as a horizontal bar plus a numeric percent. A score below approximately $N$3 percent triggers an unpleasant buzzing sound, while reaching at least $N$4 percent completes the dock and plays a gratifying chime.

The level structure is scaffolded. Levels 1 and 2 introduce drag and rotate mechanics with no visible charges and only three candidate slices per round. Levels 3 through 5 make charges visible, increase the number of unique protein pairs to 10–17, and require rotation in some rounds. Levels 6 and 7 add shaking pieces, gravity, and four-way choice. Each round has a fixed time and a life system; quick, accurate docks earn bonus points, and cumulative score, average time, and selection precision are shown at level end. Optional multiple-choice quizzes tailored to UK GCSE and A-Level curricula appear between levels, provide extra points and textual feedback when answered correctly, and can be skipped to preserve flow.

Since mid-2019, the game has been deployed in-class and at public engagement events. Teachers report that once introduced, roughly 80 percent of students continue to play during breaks. The paper states that a formal, controlled study on knowledge retention is pending, but qualitative feedback highlights improved intuition for why shapes and complementary charges must both be considered, greater confidence with vocabulary such as interface, salt bridge, and binding affinity, and heightened motivation to explore real PDB structures beyond the game (Leymarie et al., 2022).

3. The Docking Game for flexible protein–ligand binding

The 2025 Docking Game paper models blind flexible protein–ligand docking as a two-player game

$N$5

where the ligand player $N$6 predicts binding-pocket residues on the protein and the 3D coordinates of the holo ligand, and the protein player $N$7 predicts the 3D coordinates of the holo protein-pocket (Zhang et al., 7 Aug 2025). The action spaces are the model parameters: $N$8 for the ligand player’s pocket-prediction and ligand-docking network, and $N$9 for the protein player’s pocket-docking network.

The players minimize losses built from three constituent terms:

$0^\circ$0

$0^\circ$1

The paper specifies $0^\circ$2 as weighted binary-cross-entropy plus Huber center loss for selecting residues, $0^\circ$3 and $0^\circ$4 as Huber loss on 3D coordinates, and $0^\circ$5 as MSE on the inter-atomic distance map between predicted ligand atoms and pocket residues. A pair $0^\circ$6 is a Nash equilibrium when each parameter set minimizes its own loss given the other.

Loop Self-Play, or LoopPlay, is the optimization procedure proposed to solve this game. It uses a two-level loop. In the outer loop, the ligand and protein players exchange predicted poses, allowing each to incorporate the other’s structural predictions. In the inner loop, each player refines its own predicted ligand or pocket poses by feeding them back into its own model. After each outer-loop phase, only the active player’s parameters are updated by gradient descent on its own loss. The paper characterizes the outer loop as “cross-player” exchange and the inner loop as “self-refinement.”

The theoretical analysis states that if $0^\circ$7 and $0^\circ$8 are compact, the losses are bounded, and every self-play update of the active player yields a decrease of at least $0^\circ$9 in its own loss, then the sequence of alternating updates generated by LoopPlay converges to a Nash equilibrium of the Docking Game. The proof sketch proceeds by showing that the game is an exact potential game and then applying an $90^\circ$0-finite improvement argument in a compact potential game (Zhang et al., 7 Aug 2025).

The paper’s discussion attributes the method’s speed and accuracy to three factors stated explicitly: it is regression-based, with no slow diffusion or sampling loops; the outer loop enforces mutual adaptation and avoids “chicken-and-egg” mismatches; and the inner loop repeatedly refines heterogeneous-graph predictions, akin to implicit energy minimization but in a single-shot feedforward pass.

4. DockGame for multimeric rigid protein docking

DockGame addresses the less-studied multimeric docking problem, where two or more proteins must be assembled jointly rather than through a binary approximation (Somnath et al., 2023). In this framework, a rigid multimeric protein complex is composed of $90^\circ$1 constituent proteins or sub-assemblies indexed by $90^\circ$2, and each protein is treated as a player in a cooperative game.

The action space for player $90^\circ$3 is

$90^\circ$4

where a rotation $90^\circ$5 and a translation $90^\circ$6 determine the protein’s placement. If $90^\circ$7 is the original point cloud of $90^\circ$8 atoms for chain $90^\circ$9 with center of mass $180^\circ$0, then under action $180^\circ$1 it is mapped to

$180^\circ$2

The joint action space is

$180^\circ$3

The framework posits a scalar game potential

$180^\circ$4

intended in principle to encode the true biophysical energy of the resulting assembly. Equilibria are the joint actions at which no single player can unilaterally lower the potential; equivalently, they are local minima of $180^\circ$5. The equilibrium condition is written as

$180^\circ$6

Because the true energy $180^\circ$7 is not available in closed form, the paper develops two instantiations. The first learns a differentiable surrogate $180^\circ$8 parameterized by a graph neural network and supervised by a physics-based energy function from PyRosetta. Decoys are generated by perturbing the ground-truth complex, scored with PyRosetta, and used in a Bradley–Terry pairwise ranking loss:

$180^\circ$9

To avoid trivial equilibria in which chains fly apart, the surrogate is augmented by a soft distance penalty with threshold $270^\circ$0 Å. Equilibria are then computed by simultaneous Riemannian gradient descent on the product manifold.

The second instantiation treats true assemblies as samples from the Gibbs distribution

$270^\circ$1

and approximates this distribution by a score-based diffusion model over the manifold $270^\circ$2. The forward process is a variance-exploding SDE; on $270^\circ$3 it uses the isotropic Gauss–Weingarten kernel, and on $270^\circ$4 it uses a standard Gaussian. An equivariant score network is trained with a score-matching objective, and inference integrates the reverse SDE as a geodesic random walk on each component manifold.

The cooperative-game framing is used to motivate several claimed advantages: natural scaling to $270^\circ$5 proteins by decoupling the combinatorial roto-translation search into individual action spaces, a sound equilibrium concept that can yield multiple plausible assemblies, and flexibility in the choice of differentiable scoring function or learned surrogate.

5. Reported evaluation results

Framework	Evaluation setting	Reported outcome
Bioblox2.5D	In-class and public engagement since mid-2019	Roughly 80 percent of students continue to play during breaks; qualitative gains in intuition, vocabulary, and motivation
LoopPlay	PDBBind v2020, blind apo-structure setting	Approximately a 10 percent improvement in predicting accurate binding modes compared to previous state-of-the-art methods
DockGame	Multimeric subsets of DB5.5	DockGame-SM runs two orders of magnitude faster than ClusPro and three orders faster than Multi-LZerD

For LoopPlay, experiments were conducted on PDBBind v2020, ablated to 12,807 training, 734 validation, and 303 test complexes, with a further unseen-receptor subset of 114 complexes whose UniProt IDs do not appear in training. Metrics included ligand RMSD percentiles, mean RMSD, success rates below 2 Å and 5 Å, pocket-prediction accuracy, pocket RMSD, and inference time per complex. On all test complexes, FABFlex reported mean RMSD 5.44, success rates 40.6 percent at $270^\circ$6 Å and 68.3 percent at $270^\circ$7 Å, with time 0.17 s. LoopPlay(2,4) reported mean RMSD 4.89, success rates 39.6 percent and 69.9 percent, with time 0.23 s. LoopPlay(2,6) reported mean RMSD 4.90, success rates 41.9 percent and 71.9 percent, with time 0.32 s. The paper summarizes this as an approximately 10 percent improvement in mean RMSD, with gains of +1.3 percentage points at $270^\circ$8 Å and +3.6 percentage points at $270^\circ$9 Å. On the unseen-receptor subset, mean RMSD is reduced from 7.17 Å for FABFlex to approximately 6.65 Å for LoopPlay, described as an approximately 8 percent gain (Zhang et al., 7 Aug 2025).

The same study reports that self-play with $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$0 harms performance relative to single-phase training, whereas gains plateau around $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$1, $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$2, trading off depth against over-refinement. Case studies including PDB 6G24 and 6K1S are said to show better pocket localization with centroid error reduced by 4–10 Å and more accurate ligand conformations with RMSD reduced by 5–15 Å relative to FABFlex.

For DockGame, benchmarking was conducted on the multimeric subsets of the 253 complexes in Docking Benchmark 5.5. Compared methods included Attract, PatchDock, ClusPro, EquiDock, DiffDock-PP, Multi-LZerD, DockGame-E, and DockGame-SM. Reported median results on the test set were: ClusPro runtime $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$3 s, median C-RMSD 3.38 Å, median TM-score 0.8938; DiffDock-PP runtime 80 s, median C-RMSD 16.29 Å, median TM-score 0.7312; Multi-LZerD runtime $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$4 s, median C-RMSD 20.89 Å, median TM-score 0.6266; DockGame-E (20) runtime 182 s, median C-RMSD 17.74 Å, median TM-score 0.6820; DockGame-SM (40) runtime 157 s, median C-RMSD 9.44 Å, median TM-score 0.7773. The paper concludes that, despite tackling the harder $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$5-chain docking problem, DockGame-SM runs two orders of magnitude faster than ClusPro, three orders faster than Multi-LZerD, and achieves comparable median C-RMSD and TM-score to state-of-the-art binary methods (Somnath et al., 2023).

6. Abstractions, limitations, and interpretive significance

A defining feature of docking-game formulations is controlled abstraction. In Bioblox2.5D, full 3D manipulation is replaced by 2D slices, continuous Coulomb forces by discrete annotated spheres, and full physical energy calculation by an overlap-and-distance percent match. The paper presents these as compromises that preserve silhouettes, protrusions, cavities, charge complementarity, and an interpretable score while avoiding unwieldy six-degree-of-freedom controls and opaque real-time physics (Leymarie et al., 2022).

In the flexible protein–ligand Docking Game, the main stated limitation is that performance depends on the choice of outer and inner loop counts $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$6, with the convergence proof relying on assumptions that the self-play update decreases the active player’s loss by at least $$E_{\rm elec}=k_e\sum_{i,j}\frac{q_i q_j}{\epsilon r_{ij}},$$7 and that losses are bounded on compact parameter spaces. The paper adds that real optimizers only guarantee descent in expectation. It also proposes extensions including side-chain flexibility, richer physics-based priors in the distance-map loss, multi-player games such as solvent as a third agent, and meta-learning of loop counts (Zhang et al., 7 Aug 2025).

In DockGame for multimeric docking, the surrogate-learning route depends on the fidelity of physics-based supervision from PyRosetta, which the paper notes may not perfectly correlate with biological truth, while the diffusion score-based route demands abundant assembly data. The paper also states that new evaluation metrics are needed to judge the quality and diversity of multiple equilibria beyond a single ground-truth crystal structure. Extensions mentioned include general-sum games for per-protein objectives and incorporation of conformational flexibility via diffusion-bridge methods (Somnath et al., 2023).

Taken together, these works show that a docking game is not a single method but a family of representations for docking problems. In the educational setting, the representation is a scaffolded puzzle designed to teach molecular shape, salt bridges, and binding-affinity proxies. In the algorithmic setting, the representation is a game over model parameters or rigid-body actions, with docking interpreted through equilibrium, self-play, surrogate potentials, or diffusion on manifolds. This suggests that the term identifies a methodological pattern: docking is reformulated as an interactive system whose rules are simpler than full molecular physics but still structured enough to preserve core docking concepts.