ReactMech: Multiscale Mechanistic Modeling

• ReactMech is a dual-framework system integrating spatial hybrid simulation of active biological matter with graph-based prediction of chemical reaction mechanisms.
• It unifies discrete particle or atom-level dynamics with continuous reaction-diffusion processes through clear mechanistic rules and learned patterns.
• Its applications span active matter simulations, impurity screening, and complete mechanistic elucidation in both biological and chemical domains.

ReactMech (also referred to as DeepMech) denotes two distinct, independently developed frameworks for mechanistic modeling: (1) a spatial hybrid-systems simulation language and engine for active biological matter, and (2) a machine learning/data framework for predicting full chemical reaction mechanisms. Despite their disciplinary divergence—computational biology versus computational chemistry—both frameworks embody the common principle of explicitly coupling dynamically evolving discrete and continuous entities (particles/objects or atoms/molecules) via mechanistic, interpretable rules or learned patterns. The term "ReactMech" has thus emerged as an umbrella for interpretable, multiscale, and mechanistically faithful computational paradigms in both soft-matter physics and chemical reactivity prediction (Somogyi et al., 2017, Joung et al., 2024, Das et al., 19 Sep 2025).

1. Core Conceptual Foundations

Biological Hybrid Systems

The original ReactMech framework addresses the modeling of biological cells as exemplars of "active matter": systems whose constituents both exert and respond to complex mechanical and chemical stimuli. This necessitates a formalism bridging particle-level (Lagrangian) mechanical models and reaction-diffusion systems traditionally described by continuum equations (Somogyi et al., 2017). Here, ReactMech introduces a modeling language where:

• Mechanical structure is discretized into particles and links, each with positions, velocities, and user-defined force laws.
• Chemical concentrations "ride" on these particles or spatial regions, undergoing advection, reaction, and diffusion.
• Discrete state transitions (e.g., cell division, link formation/breaking) are encoded declaratively using process rules.

The approach is unified: chemical and mechanical rules coexist in a formally coupled description, and the compiler auto-generates simulation code to resolve both domains concurrently.

Chemical Mechanism Prediction

In the molecular domain, ReactMech/DeepMech refers to a supervised learning/data framework for deriving stepwise reaction mechanisms (CRMs—complete reaction mechanisms) from curated datasets of atom-mapped, mass-balanced chemical steps. The approach couples:

• Large-scale, curated datasets of elementary steps (ReactMech dataset, ∼30K CRMs, 100K+ steps) with explicit atom conservation and mass/charge balancing (Das et al., 19 Sep 2025).
• A graph-based deep neural architecture (DeepMech) which leverages message-passing, atom- and bond-level attentions, and subgraph-based mechanistic templates (TMOp).
• Training and evaluation protocols that stress both in-distribution and out-of-distribution (OOD) generalization, interpretability, and the explicit tracking of byproducts and intermediates.

This framework elevates mechanistic prediction above simple product identification, incorporating mechanistic imputation, intermediate and byproduct enumeration, and pathway-level inference (Joung et al., 2024).

2. Formal Syntax, Semantics, and Data Construction

ReactMech Simulation Language

Every simulation model consists of object type definitions and process rules (Somogyi et al., 2017):

• Object Types: MaterialRegion, Particle, Composite.
• Attributes: Continuous (via conc/amount), discrete states, or binding sites.
• Processes:
  • Continuous rules (proc) for reactions, transport, and diffusion.
  • Link declarations for mechanical coupling; e.g., springs with chemistry-modulated rest lengths.
  • Conditional predicates (when, while) for event-driven rewriting.
• Spatial Scoping: Unbound symbols in process bodies are resolved hierarchically (from local object to global context).

An example script for a chemomechanically-coupled secreting cell includes chemically familiar constructs (e.g., proc (A) -> (B)) and mechanically motivated links (e.g., spring constants as functions of local concentrations).

ReactMech Dataset for Chemistry

The chemical ReactMech dataset is constructed via:

• Manual annotation of multistep CRMs for key classes from the USPTO corpus, transition-metal coupling collections, and diverse organic reactions (Das et al., 19 Sep 2025).
• Atom-mapping for every reactant, intermediate, and product; checks guarantee mass and charge conservation.
• Extraction of per-step SMARTS-based templates (generalized mechanistic operations).
• Final dataset: >29,000 full mechanisms; ~105,000 elementary steps; coverage across 67 mechanistic classes and explicit prebiotic transformations.
• Datasets are stratified into in-distribution and OOD test splits, supporting rigorous assessment of mechanistic generalization.

A similar approach—using large-scale patent-derived reaction records, manually encoded templates (175 elementary-step types for 86 core classes), and template-guided sequential imputation—yields over 5.8 million steps in (Joung et al., 2024), with intermediates and byproducts annotated throughout.

3. Mathematical and Algorithmic Frameworks

Hybrid Dynamical Simulation (Biological ReactMech)

• Chemical kinetics and transport: Discretized reaction-diffusion equations over mobile particles,

$$\frac{dC_i^p}{dt} = \sum_{q \in N(p)} D_i \frac{C_i^q - C_i^p}{\|\mathbf{r}_q-\mathbf{r}_p\|} w(\mathbf{r}_p,\mathbf{r}_q) + \sum_{\alpha} \nu_{i\alpha}\,k_\alpha(C(\mathbf{r}_p))$$

• Mechanical dynamics: Newtonian force-balance for each particle,

$$m_p\frac{d^2\mathbf{r}_p}{dt^2} = \sum_{q \neq p} [\mathbf{F}^L_{pq} + \mathbf{F}^C_{pq} + \mathbf{F}^D_{pq} + \mathbf{F}^R_{pq}] + \mathbf{F}_p^{\rm ext}$$

where $F^L_{pq}$ are user-declared, possibly concentration-dependent link forces.

• Coupling: Mechanochemical feedback arises by making mechanical parameters (e.g., spring stiffness) explicit functions of interpolated chemical concentrations at link midpoints.

Mechanism Prediction Models (DeepMech)

• Graph Representation: Each molecular entity encoded as disconnected or connected graph $G=(V,E)$ with atom-level and bond-level attributes (Das et al., 19 Sep 2025).
• Message-Passing and Attention:
  • Atom representations updated via $T$ layers of

  $$h_u^{(t+1)} = U\left(h_u^{(t)}, \sum_{v \in N(u)} M(h_u^{(t)}, h_v^{(t)}, e_{uv})\right)$$ - Global reactivity attention (GRA) incorporates topological distance through a learned bias.

• Template Mechanistic Operations (TMOp):

  • 545 classes indexed as (template, operation, hydrogen/charge delta).
  • Pooling and classification over attended bond embeddings to choose both the reactive bonds and the mechanistic operation.
• Inference: Apply top-k bond selections and the best TMOp class to generate the next intermediate structure; beam search to traverse full multistep CRMs.
• Losses: Binary cross-entropy for bond reactivity; multiclass for TMOp; total loss is their sum.

The combined approach integrates data-driven identification of reactive sites (via attention), operation class, and a physically and chemically valid update of molecular graphs.

4. Compiler, Simulation, and Training Protocols

Biological Simulation Pipeline (Somogyi et al., 2017)

• Parsing/AST generation: Recursive-descent parser interprets model scripts.
• Semantic analysis: Build global symbol tables, generate ODE and force systems, and event-predicate logic.
• Code generation: Emit C code to match the structure and requirements of the simulation backends (CVODE for ODEs, mdcore for DPD/dynamics); JIT compilation to native shared libraries.
• Runtime: Three-phase loop per $\Delta t$:
  • Execute discrete events (when/while conditions).
  • Integrate chemical dynamics.
  • Advance mechanical positions/velocities.

Optimizations include contiguous memory, neighbor-search acceleration (Verlet/cell lists), and pending parallelization or GPU offload.

DeepMech/ReactMech Training and Inference (Das et al., 19 Sep 2025, Joung et al., 2024)

• Optimization: Adam with weight decay, scheduled learning rate; batch sizes scaling with step and model complexity.
• Early stopping and regularization: Patience-based halting and dropout.
• Beam search for CRM inference: Sequentially apply the model at every step, terminating on special templates or cycle detection.
• Evaluation metrics: Top-$k$ accuracy at the elementary and CRM levels; precision/recall for bond and byproduct prediction.

Model variants include baseline graph-edit predictors (WLDN), Graph2SMILES (D‐MPNN encoder + Transformer decoder), and vanilla sequence Transformers.

5. Performance and Generalization Characteristics

Simulation Engine (Biological) (Somogyi et al., 2017)

• The mesh-free, Lagrangian engine supports dynamic object creation/destruction, seamless handling of division, death, or migration, and physically consistent evolution without grid artifacts.
• JIT compilation makes simulations efficient: interpreter overhead is eliminated.
• Currently, runtime is single-threaded, but parallelization and GPU acceleration are straightforward due to data contiguity and neighbor-list locality.

Mechanistic Model Predictive Accuracy (Das et al., 19 Sep 2025, Joung et al., 2024)

Task	DeepMech Top-1 (%)	G2S (%)	Transformer (%)
Step prediction (ID)	98.98 ± 0.12	98.00 ± 0.14	93.11 ± 0.27
CRM prediction (ID)	95.94 ± 0.21	93.52 ± 0.67	75.27 ± 1.28
CRM (OOD: Amine+Acid Halide)	93.55 ± 0.72	60.59 ± 4.77	16.41 ± 8.02

• Top-k performance for elementary steps and full mechanism recovery is highest for DeepMech.
• Robustness across OOD classes is observed, with DeepMech substantially outperforming G2S and Transformer baselines for several mechanistic families (e.g., sulfonyl halides, Ni-catalyzed coupling).
• Byproduct/side-product enumeration validated on SNAr, Appel, and Suzuki reactions.

Dataset/Template-Generalization

• Mechanistic template models remain limited by the scope of template classes exhibited during training.
• Global (product-only) models achieve slightly higher product prediction on unseen classes, but fail to enumerate mechanisms, intermediates, or byproducts.

6. Interpretability and Case Studies

Simulation Visualizations (Somogyi et al., 2017)

• OpenGL-based overlays of scalar chemical fields (concentration per particle).
• Deformable meshes for cell boundaries, colored by local mechanical or chemical state.
• Direct mapping from mechanistic parameters to observable morphologies or spatial dynamics.

Attention-Based Mechanistic Insight (Das et al., 19 Sep 2025)

• Atom-level attention $\alpha_{uv}$ highlights known reactive centers (e.g., transition metals, functional groups).
• Bond-level attention $\beta_{uv}$ ranks bond-formation/breakage events, matching chemical intuition.
• High attention sites correlate with experimentally validated stepwise reactivity, guiding both mechanism elucidation and design.

Case analyses:

• Prebiotic synthesis of serine and aldopentose demonstrates plausible, ab initio pathway recovery.
• Major/side-product differentiation in SNAr and Appel test sets.
• Byproduct prediction (e.g., triphenylphosphine oxide in Appel) for impurity-aware modeling.

7. Practical Implications and Limitations

Application Domains

• Mechanism elucidation (chemistry): End-to-end, interpretable prediction of intermediate states and reaction pathways, supporting hypothesis generation and automated arrow-pushing analogs (Joung et al., 2024, Das et al., 19 Sep 2025).
• Impurity screening: Enumeration of pathway byproducts is embedded in the data and inference, aiding process development.
• Biological simulation: Unified treatment of reaction-diffusion dynamics and mechanical cell processes (adhesion, division, migration) using declarative model scripts (Somogyi et al., 2017).

Outstanding Limitations

• Mechanistic data coverage remains a bottleneck: template/operation classes not seen during training limit OOD generalization.
• Mass/charge conservation is not always enforced in transformer/SMILES-sequence models, causing hallucinations.
• Missing reagents/stoichiometric information can impede mechanistic completeness for some reaction classes.
• Multi-step inference (beam search) may accumulate small errors, reducing full-pathway accuracy.

Ongoing work focuses on extending reaction-class coverage, refining mechanistic template extraction, enforcing explicit conservation laws, and further enhancing performance for both simulation and predictive mechanistic frameworks.