Machine Learning for Chemical Transition State Search

34 concepts25 papers

To understand why a chemical reaction happens—and how fast—you must find its transition state (TS). The TS is the highest-energy point along the minimum energy path connecting reactants to products. It represents a fleeting, unstable configuration where old bonds are half-broken and new ones are half-formed. Because the TS dictates the activation energy barrier, locating it is fundamental to predicting reaction kinetics, calculating yields, and designing new catalysts.

However, finding a TS is notoriously difficult. Mathematically, it is a first-order saddle point on a high-dimensional potential energy surface (PES)—a point that is a local maximum along exactly one direction (the reaction coordinate) but a local minimum along all others. Traditional quantum chemistry methods, such as Density Functional Theory (DFT), rely on iterative geometry optimization algorithms to locate these points. As Pascal Friederich et al. note, these automated workflows frequently suffer from numerical convergence failures, requiring highly accurate initial structural guesses to succeed.

Furthermore, verifying that a structure is truly a TS requires calculating the Hessian matrix (the second derivatives of energy with respect to atomic positions) to confirm the presence of exactly one imaginary frequency. Calculating the full Hessian is computationally prohibitive for large systems. To mitigate this, Markus Reiher introduced a mode-tracking algorithm (MTSEARCH) that selectively calculates only the necessary vibrational modes rather than the entire Hessian. Yet, even with such algorithmic optimizations, traditional PES exploration remains a massive computational bottleneck that prevents high-throughput screening.

From Iterative Optimization to Generative Modeling

Instead of slowly climbing the energy landscape step-by-step using expensive quantum mechanical forces, a rapidly growing frontier in computational chemistry asks: What if we could predict the 3D geometry of the transition state directly?

By framing transition state search as a conditional generation problem—where a machine learning model generates the TS geometry conditioned on the reactants and products—researchers can bypass hours or days of quantum calculations.

Traditional:  [Reactants/Products] ---> Iterative DFT Gradient Ascent (Hours/Days) ---> [Transition State]
Generative:   [Reactants/Products] ---> ML Generative Model (Seconds)              ---> [Transition State]

Diffusion models have emerged as a powerful tool for this task. For instance, Yuanqi Du et al. developed OA-ReactDiff, an object-aware equivariant diffusion model that generates 3D TS structures in seconds. By enforcing physical symmetries (such as rotational and translational equivariance) over multi-molecule systems, the model learns the spatial relationships inherent to reacting species without requiring tedious manual atom mapping. Similarly, Seonghwan Kim et al. introduced TSDiff, which generates 3D TS geometries directly from simple 2D molecular graphs, bypassing the need for complex 3D alignments of the starting materials.

While diffusion models are highly effective, their stochastic nature means they require sampling multiple noisy steps to arrive at a final structure. To introduce determinism and further accelerate generation, Chenru Duan et al. proposed React-OT, an optimal transport-based framework. React-OT maps the transformation from reactants to products along a mathematically optimal trajectory, generating accurate transition states deterministically in a single forward pass. Similarly, Ping Tuo et al. utilized flow matching in their MolGEN framework to efficiently generate continuous reaction pathways, offering a fast, deterministic alternative to stochastic sampling.

Mapping Continuous Paths and Escaping Local Minima

Transition states do not exist in isolation; they are part of a continuous path. Traditional methods often discretize this path into a series of images (like the Nudged Elastic Band method), which can get stuck in local minima or fail if the path is highly curved.

To solve this, researchers are leveraging continuous neural representations. Kalyan Ramakrishnan et al. introduced the use of Implicit Neural Representations (INRs) for chemical reaction paths. Instead of optimizing discrete molecular configurations, an INR trains a neural network to represent the entire reaction path as a continuous function. This allows the model to robustly smooth out the energy landscape, escape local minima, and accurately pinpoint the exact saddle point.

When the collective variables (the specific geometric parameters that change during a reaction) are unknown, sampling transition paths becomes even harder. Kiyoung Seong et al. addressed this by developing TPS-GFN, which uses Generative Flow Networks (GFlowNets) to sample diverse transition paths proportional to their energy rewards without requiring predefined collective variables. This ensures that the model explores multiple hidden reaction mechanisms that traditional methods might miss.

Scaling Up to Large Molecules

As molecules grow larger, the dimensionality of the potential energy surface explodes, making direct generative modeling or path sampling increasingly difficult. To scale these methods to real-world systems, researchers must exploit the locality of chemical reactions.

Ron Shprints et al. developed FragmentFlow, which employs a divide-and-conquer strategy. Recognizing that chemical reactions are localized phenomena, FragmentFlow focuses its generative power on the "reactive core"—the specific atoms undergoing bond breaking and making—while treating the rest of the large molecule as a rigid or semi-rigid framework. This localization significantly reduces the complexity of the generation task, allowing the model to accurately predict TS geometries for large, complex structures.

For alternative high-throughput generation of transition-state conformer ensembles, Stefan P. Schmid et al. introduced racerTS, which leverages constrained distance geometry to rapidly sample valid conformers, providing the necessary structural diversity required for accurate machine learning data generation.

Surrogates and Interatomic Potentials

In some applications, explicitly generating the 3D coordinates of a transition state is not strictly necessary; instead, we only need a representation of its stabilizing effects or its topological structure.

For example, in enzyme screening, enzymes are evolutionarily optimized to stabilize transition states. Peter Mikhael et al. developed CLIPZyme, which uses contrastive learning to align 3D enzyme structures with a "pseudo-transition state" representation of a reaction, enabling rapid virtual screening of millions of proteins. Similarly, for reaction yield prediction, Ankit Ghosh et al. created YIELDNET, which learns a "Condensed Graph of Reaction" as a topological surrogate for the transition state, bypassing geometric calculations entirely to predict chemical efficiency.

Explicit TS Generation:  [Reactants] ---> [3D Coordinates of Saddle Point] ---> [Activation Energy]
Surrogate Modeling:      [Reactants] ---> [Condensed Graph / Pseudo-TS]    ---> [Yield / Enzyme Fit]

When explicit energy evaluations are required, Machine Learning Interatomic Potentials (MLIPs) act as fast surrogates for DFT. However, standard MLIPs often struggle to model the high-energy regions around transition states because they lack training data for those unstable regions. To fix this, Taoyong Cui et al. compiled HORM, a large-scale molecular Hessian database specifically designed to train reactive MLIPs, ensuring they capture the correct curvature around saddle points.

To understand when to use which tool, Qiyuan Zhao et al. established a standardized framework to benchmark these emerging ML approaches. Their findings highlight that while pre-trained MLIPs often require extensive fine-tuning to find transition states reliably, direct generative models (like React-OT) consistently show superior performance in discovering valid TS structures, cementing generative AI as a cornerstone for the future of automated chemical discovery.

Key Foundational Work

Mode-tracking based stationary-point optimization (2015) — contributed 2 concept elements to this direction.
Machine learning dihydrogen activation in the chemical space surrounding Vaska's complex (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_computational_cost_of_transition_state_s_c3a090a6,
  title = {computational cost of transition state search},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/computational-cost-of-transition-state-search/rdr_c3a090a6b0f04303853adb9ed491e82b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_computational_cost_of_transition_state_s_c3a090a6_v1,
  title = {computational cost of transition state search},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/computational-cost-of-transition-state-search/rdr_c3a090a6b0f04303853adb9ed491e82b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake