AIMMD: AI for Molecular Mechanism Discovery

• AIMMD is a machine-learning-assisted framework that couples transition path sampling with on-the-fly committor probability estimation to discover rare-event mechanisms.
• It adaptively targets the transition state ensemble, enabling efficient sampling and extraction of physically interpretable reaction coordinates.
• The method has been applied to diverse processes such as ion dissociation, polymer folding, and membrane-protein assembly, offering new mechanistic insights.

Artificial Intelligence for Molecular Mechanism Discovery (AIMMD) is a machine-learning-assisted rare-event sampling framework introduced to discover and characterize molecular transition mechanisms in systems where the relevant events occur on timescales far longer than straightforward molecular dynamics can reach. In its core formulation, AIMMD couples transition path sampling (TPS) to on-the-fly estimation of the committor probability $p_B(\mathbf{x})$, then uses that learned committor both to improve sampling efficiency and to derive a human-interpretable reaction coordinate. The method was introduced in the context of rare-event mechanism discovery rather than generic acceleration, with examples including ion association and dissociation, nucleation, polymer folding, and membrane-protein assembly (Minh et al., 5 Feb 2026).

1. Problem setting and conceptual basis

AIMMD addresses two coupled difficulties in molecular simulation. The first is the standard rare-event problem: trajectories that cross from reactant state $A$ to product state $B$ are extremely infrequent, so direct molecular dynamics spends most of its effort on waiting in metastable basins. The second, and scientifically harder, difficulty is that the governing reaction coordinate is generally unknown. AIMMD was introduced precisely for systems where direct simulation is inefficient and where hand-designed collective variables are not reliable surrogates for true transition progress (Minh et al., 5 Feb 2026).

The central theoretical object is the committor probability $p_B(\mathbf{x})$, defined as the probability that a trajectory started from configuration $\mathbf{x}$ reaches state $B$ before returning to $A$. In transition path theory, the committor is the optimal dynamical progress variable; the transition-state ensemble is the $p_B=1/2$ level set. AIMMD is therefore committor-centered: instead of computing the committor afterward by massive shooting analysis, it learns an approximation during path sampling itself. Earlier work in the AIMMD lineage had already established the essential idea that AI can sit inside the simulation loop, infer the reaction coordinate from trajectory outcomes, and continuously redirect sampling toward mechanistically informative regions of configuration space (Jung et al., 2019).

This committor-centered view is also what distinguishes AIMMD from many enhanced-sampling methods that primarily seek acceleration. The method is designed to yield mechanistic information: localization of the transition-state ensemble, identification of relevant order parameters, recognition of pathway heterogeneity, and extraction of low-dimensional coordinates that can be interpreted in terms of physically meaningful descriptors. In later autonomous formulations, the same logic was applied to self-organization phenomena such as ion association, gas-hydrate crystal formation, and membrane-protein assembly, explicitly coupling deep reinforcement learning, transition path theory, and symbolic regression into a closed learning cycle (Jung et al., 2021).

2. Core algorithmic framework

The AIMMD workflow is a self-consistent alternation between path sampling and model fitting. States $A$ and $B$ are first defined by physically motivated collective variables; these definitions specify reactant and product basins but need not provide the full reaction coordinate. An initial reactive trajectory is then generated to seed the transition-path ensemble. From there, AIMMD performs TPS shooting moves, records the shooting-point configurations and whether the resulting trajectories are reactive $A$0 or nonreactive $A$1 or $A$2, and uses those outcomes as committor-training data (Minh et al., 5 Feb 2026).

In the standard neural formulation, a configuration $A$3 is encoded by $A$4 physically chosen descriptors, and a feed-forward neural network outputs a logit-committor $A$5. The committor estimate is obtained by the logistic transform

$A$6

The network is trained by maximum likelihood on recent shooting outcomes. In the path-type-specific encoding discussed in the commentary, reactive $A$7 paths use $A$8, $A$9 paths use $B$0, and $B$1 paths use $B$2. The corresponding negative log-likelihood is

$B$3

Once updated, the learned committor guides future shooting moves. AIMMD selects frames along the current transition path with a Lorentzian weighting centered near the transition-state region,

$B$4

where $B$5 controls the breadth of exploration. This adaptive policy preferentially launches shoots from configurations with $B$6, i.e. $B$7, while retaining some exploration away from the barrier top. In equivalent transition-path-theory form, the probability that a configuration lies on a transition path is

$B$8

which is maximal at the transition state (Jung et al., 2021).

Sampling and training continue until self-consistency is reached. One stopping diagnostic compares the observed number of transition paths with the expected number implied by the current committor estimate,

$B$9

Convergence is declared when predicted transition-path statistics are consistent with the actual TPS outcomes. This makes AIMMD not just an estimator of $p_B(\mathbf{x})$0, but a closed-loop controller that uses its current mechanistic hypothesis to decide where to acquire the next trajectory data (Minh et al., 5 Feb 2026).

3. Mechanism extraction and interpretability

AIMMD is explicitly designed to discover mechanisms rather than merely to produce more reactive trajectories. The learned committor localizes the transition-state ensemble, orders configurations by dynamical progress rather than geometric proximity, and identifies which molecular degrees of freedom control barrier crossing. This is why the framework couples machine learning to symbolic regression: the neural network serves as a flexible nonlinear approximator, but the intended scientific endpoint is a symbolic reaction coordinate written in terms of known physical descriptors (Minh et al., 5 Feb 2026).

Earlier work in the AIMMD lineage developed additional interpretability machinery. One route is gradient-based atomic attribution near the transition state,

$p_B(\mathbf{x})$1

which quantifies the relative contribution of atom $p_B(\mathbf{x})$2 to the learned reaction coordinate near the bottleneck. A second route is feature-destruction relevance analysis, where a descriptor is replaced by noise and the increase in loss is measured as

$p_B(\mathbf{x})$3

Features can then be ranked, and symbolic regression is used to fit a simpler analytic surrogate $p_B(\mathbf{x})$4 to the neural committor while penalizing symbolic complexity (Jung et al., 2019).

This interpretability layer is mechanistically consequential. It allows AIMMD to answer questions that matter in chemistry and biophysics: which interactions govern barrier crossing, whether solvent organization or coordination number is decisive, which contact patterns define assembly progress, and how nonlinear coupling between descriptors shapes the transition. Even when the resulting symbolic expression is not minimal, it can still expose variable importance and suggest experimentally testable hypotheses. The framework therefore occupies a middle ground between black-box latent-variable learning and traditional hand-crafted reaction-coordinate design.

A recurrent misconception is that AIMMD is simply an acceleration strategy for TPS. The method does improve transition-path generation, but the literature repeatedly distinguishes that computational gain from its mechanistic function. Its scientific objective is to connect reactive path ensembles, committor structure, and interpretable coordinates into one self-consistent account of molecular change (Jung et al., 2021).

4. Demonstrated systems and mechanistic findings

AIMMD and closely related committor-centered frameworks have been applied to rare events in solution chemistry, conformational change, nucleation, polymer physics, and membrane-protein assembly. Across these examples, the recurring theme is that the learned mechanism often differs from the most obvious geometric coordinate.

System	Mechanistic result	AIMMD feature
Ion association/dissociation	Solvent and ionic environment can dominate over simple pair distance	Transfer learning across monovalent salts
Alanine dipeptide	$p_B(\mathbf{x})$5 torsion is dominant, with weaker solvent descriptors	Symbolic low-dimensional RC
LiCl dissociation	Many-body solvent and counterion rearrangement are crucial	Feature relevance and atomic saliency
Gas hydrate / methane clathrate nucleation	Temperature, surface water, and cage count govern nucleation	Recovery of CNT-like variables
Homopolymer folding	Progress is defined by combinations of structural descriptors	Nonlinear descriptor learning
Mga2 membrane-protein assembly	Two distinct pathways coexist	Parallel TPS and pathway discovery

In ion association and dissociation, AIMMD was shown to learn effective mechanisms from descriptors such as interionic distance and solvation-related coordination variables. A notable result was transfer learning: a model trained for LiCl was adapted to other monovalent salts by retraining only the final neural-network layer, suggesting that the learned representation captured partially transferable mechanistic motifs (Minh et al., 5 Feb 2026).

In alanine dipeptide, the dominant descriptor recovered was a transformed $p_B(\mathbf{x})$6 torsion,

$p_B(\mathbf{x})$7

but the model also identified weaker solvent descriptors involving nearby waters. A symbolic regression fit,

$p_B(\mathbf{x})$8

nearly matched the full neural network, indicating that the conformational mechanism is torsion-dominated but not purely intramolecular (Jung et al., 2019).

In LiCl dissociation, the mechanistic conclusion was more striking: the dissociating ion-pair distance itself did not emerge as the most important learned descriptor. Instead, collective solvent and counterion rearrangements dominated. One symbolic surrogate required both distance and environmental descriptors,

$p_B(\mathbf{x})$9

implying that dissociation couples pair separation to hydration and counterion organization. Autonomous AIMMD-style self-organization studies reported a related result for ion association: successful LiCl association requires inner-shell water molecules around Li$\mathbf{x}$0 to reorient and open a gap for incoming chloride (Jung et al., 2021).

In methane clathrate nucleation, the most important variables were temperature $\mathbf{x}$1, the number of surface water molecules $\mathbf{x}$2, and the number of $\mathbf{x}$3 cages $\mathbf{x}$4. The resulting low-dimensional mechanism recovered variables associated with classical nucleation theory and captured a temperature-dependent switch from amorphous growth at low temperature to increasingly crystalline growth at higher temperature. In Mga2 transmembrane dimer assembly, AIMMD combined data from parallel TPS simulations and identified two distinct reaction pathways; in a related autonomous formulation, a symbolic model in only two interhelical contact variables was sufficient to represent the committor and support pathway clustering (Minh et al., 5 Feb 2026).

5. Relation to neighboring methods and methodological extensions

AIMMD is best understood as an extension of TPS rather than a replacement for trajectory-based path sampling. Conventional TPS samples reactive trajectories without requiring a predefined reaction coordinate, but its practical efficiency still depends on selecting productive shooting regions. AIMMD replaces static intuition with an adaptive, learned shooting strategy based on an on-the-fly committor estimate. It therefore preserves the unbiased path-generation logic of TPS while reducing the dependence on hand-crafted reaction-coordinate intuition (Minh et al., 5 Feb 2026).

The framework also sits next to, rather than inside, traditional post hoc committor analysis. Classical committor analysis remains the gold standard for validating a reaction coordinate, but it is computationally expensive because it launches many trial trajectories from many configurations. AIMMD amortizes this cost by learning a committor surrogate directly from ordinary TPS shooting outcomes. The methodological advantage is economy and integration; the methodological caveat is that surrogate quality must still be validated independently.

Two substantial extensions have already been described. Waste-recycling AIMMD-TPS uses nonreactive $\mathbf{x}$5 and $\mathbf{x}$6 trajectories, together with additional equilibrium data, instead of discarding them. This yields more complete characterization of the reaction landscape, improved estimation of the rate constant, estimation of the free energy surface as a function of the learned committor, and better committor prediction away from the transition-state region. A key technical change is that shooting points are selected using a distribution that is uniform in the committor rather than Lorentzian around $\mathbf{x}$7, directly addressing the sampling imbalance of the original method (Minh et al., 5 Feb 2026).

A second extension, AIMMD-TIS, integrates the learned logit-committor from an AIMMD-TPS run with Transition Interface Sampling. The approximate committor is used to place interfaces resembling isocommittor surfaces, after which TIS refines rates, free energies, and the committor model while providing broader coverage away from the transition state. The commentary also highlights Replica Exchange Transition Interface Sampling (RETIS) as a stronger baseline for systems with multiple well-separated pathways, and explicitly suggests that AIMMD combined with RETIS would be a promising future direction for pathway-diverse systems (Minh et al., 5 Feb 2026).

6. Limitations, validation debates, and broader significance

The strongest conceptual criticism of AIMMD is that the framework does not itself prove that the learned reaction coordinate is the true committor. Agreement with numerically estimated committors is encouraging, but the literature argues that a histogram test should be incorporated to assess whether configurations predicted to share the same committor actually exhibit statistically identical shooting outcomes. This validation issue is central because AIMMD’s mechanistic authority depends on committor fidelity, not only on improved path yield (Minh et al., 5 Feb 2026).

A second limitation is sampling imbalance. Because the standard Lorentzian shooting distribution concentrates training data near the transition-state ensemble, committor prediction can degrade for configurations with $\mathbf{x}$8 near 0 or 1. Third, AIMMD-TPS inherits a familiar TPS weakness in discovering multiple well-separated pathways; parallel TPS simulations can help, but interface-based methods such as RETIS remain more powerful when pathway diversity is large. Fourth, the learned reaction coordinate still depends on the user-supplied descriptor pool: if important descriptors are absent, the learned mechanism may be systematically incomplete. Fifth, very diffusive processes with extremely long reactive trajectories remain computationally expensive because AIMMD improves how trajectory data are used, not the cost of the underlying dynamics itself (Minh et al., 5 Feb 2026).

Earlier formulations identify additional structural assumptions. Metastable states $\mathbf{x}$9 and $B$0 must still be defined by the user, and the quality of mechanistic inference depends on the molecular representation. In practice, AIMMD-style studies often rely on physically informed inputs such as internal coordinates, symmetry functions, contact variables, or other domain-specific descriptors rather than raw Cartesian coordinates. This preserves symmetry and can make mechanisms interpretable, but it also means the method is not fully unsupervised in representation learning (Jung et al., 2019).

Within the broader AI-for-molecules literature, some later systems are best regarded as AIMMD-adjacent rather than direct continuations of the committor-learning tradition. Auditable multi-agent molecular optimization platforms emphasize provenance, molecule-centered reasoning trajectories, docking-guided edits, and scientific traceability, thereby supporting mechanism-relevant ligand design without directly solving mechanism discovery (Ünlü et al., 5 Aug 2025). MCP-based binder-design agents similarly integrate MaSIF, Rosetta, ProteinMPNN, and AlphaFold3 into a protocol-driven workflow for end-to-end binder generation, but their main contribution is workflow automation and infrastructure integration rather than committor-based mechanistic inference (Ge et al., 16 Jan 2026). The distinction is substantive: core AIMMD literature is centered on reaction coordinates, transition paths, and mechanism extraction from dynamics, whereas these adjacent systems primarily automate molecular design workflows.

The broader significance of AIMMD lies in three linked contributions. It embeds machine learning into trajectory-space sampling rather than using ML only for post-processing; it makes the committor operational during simulation rather than treating it as an expensive diagnostic; and it turns rare-event acceleration into mechanism discovery by insisting on interpretable reaction coordinates and pathway analysis. In that sense, AIMMD occupies a distinctive position at the intersection of transition path theory, machine learning, and mechanistic molecular science (Jung et al., 2021).