A Machine-Learned Symbolic Committor for a Chemical Reaction: Retinal Isomerization
Kai Töpfer, Gianmarco Lazzeri, Vittoria Ossanna, Florian Renner, Gianluca Lattanzi, Roberto Covino, Bettina G. Keller · 2026
Read A Machine-Learned Symbolic Committor for a Chemical Reaction: Retinal Isomerization by Kai Töpfer, Gianmarco Lazzeri, Vittoria Ossanna, Florian Renner, Gianluca Lattanzi, Roberto Covino, Bettina G. Keller (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-24245.txt.
Introduction
Töpfer and six colleagues learn an interpretable committor for retinal isomerization, the cis-trans flip around the C13-C14 bond. They apply AI molecular mechanism discovery to N-retinylidene-lysine in vacuum, learning from unbiased trajectories by two-way shooting. Using the logit rather than the committor resolves the reaction path across the transition region. Four dihedrals around the reactive bond carry the signal. Symbolic regression shows their nonlinear coupling is required for the S-shaped stepwise pathway in transition paths, a shape the minimum-energy path misses.
Catalog record
| CREATOR | Kai Töpfer, Gianmarco Lazzeri, Vittoria Ossanna, Florian Renner, Gianluca Lattanzi, Roberto Covino, Bettina G. Keller |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61356 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.24245
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Machine-Learned Symbolic Committor for a Chemical Reaction: Retinal Isomerization Kai Töpfer, Gianmarco Lazzeri, Vittoria Ossanna, Florian Renner, Gianluca Lattanzi, Roberto Covino, Bettina G. Keller Abstract The thermal cis – trans isomerization around the C 13 =C 14 double bond of retinal is a prototypical high-barrier reaction whose mechanism hinges on subtle out-of-plane bending motions. We apply Artificial Intelligence for Molecular Mechanism Discovery (AIMMD) to N-retinylidene-lysine in vacuum, learning the committor from unbiased molecular dynamics trajectories generated by two-way shooting. Parametrizing the logit of the committor, rather than the committor itself, allows the neural network to resolve the reaction coordinate across the full transition region, not only at the isocommittor surface p B ( 𝐱 ) = 0.5 p_{B}(\mathbf{x})=0.5 . Holdback input randomization identifies four proper dihedrals around the reactive bond as the informative coordinates, while the improper dihedrals at C 13 and C 14 prove unsuitable because reactant, transition, and product states share the same values. Symbolic regression then distills the network into compact analytical expressions and shows that a nonlinear coupling of all four dihedrals is required to reproduce the S-shaped, stepwise pathway seen in the transition path ensemble. This S-shape is absent from the minimum-free-energy path: it arises from the non-equilibrium dynamics of the short ( ∼ 0.13 \sim 0.13 ps) transition events combined with the mass asymmetry between heavy-atom and hydrogen-bearing dihedrals. An interpretable, machine-learned committor thus exposes dynamical features of the mechanism to which the free-energy surface is blind. The workflow requires no prior assumptions about the reaction coordinate and extends naturally to other isomerizations and to chemical reactions more broadly. † † affiliation: Department of Biology, Chemistry and Pharmacy, Freie Universität Berlin, 14195 Berlin, Germany. † † affiliation: These authors contributed equally to this work. † † affiliation: Institute of Computer Science, Goethe University Frankfurt, Frankfurt am Main, Germany. † † affiliation: Frankfurt Institute for Advanced Studies, Frankfurt am Main, Germany. † † affiliation: These authors contributed equally to this work. † † affiliation: Department of Cellular, Computational and Integrative Biology, University of Trento, Trento, Italy. † † affiliation: Frankfurt Institute for Advanced Studies, Frankfurt am Main, Germany. † † affiliation: Department of Biology, Chemistry and Pharmacy, Freie Universität Berlin, 14195 Berlin, Germany. † † affiliation: Deparment of Physics, University of Trento, Trento, Italy. † † affiliation: Institute of Computer Science, Goethe University Frankfurt, Frankfurt am Main, Germany. † † affiliation: Frankfurt Institute for Advanced Studies, Frankfurt am Main, Germany. † † affiliation: Department of Biology, Chemistry and Pharmacy, Freie Universität Berlin, 14195 Berlin, Germany. † † email: [email protected], [email protected] † † suppinfo: 1 Introduction A rare event 55 ; 21 is a transition across an energy barrier that occurs infrequently because the barrier is much higher than the average thermal fluctuations. Chemical reactions are a prototypical but also a particularly challenging class of rare events 6 ; 54 . To rationalize their mechanism and to calculate reaction rates, one introduces reaction coordinates, i.e. functions of the conformational space that monotonically scale between the two energy basins in either side of the barrier. While it is usually possible to define a naive reaction coordinate by simply interpolating between reactant and product state, such a reaction coordinate poorly captures the mechanism and often yields highly inaccurate rates. Finding optimal reaction coordinates is therefore a long-standing but still very active field of research 10 ; 24 ; 5 . Minimum energy paths 47 ; 52 ; 15 , traditionally used as models for the optimal reaction coordinate, identify the transition state (TS) as a saddle point in the PES but neglect entropic contributions and dynamical effects. Alternatively, the committor function 3 ; 12 ; 34 characterizes the actual dynamical flow of reactive trajectories. This offers a complementary perspective, in which the TS is defined as the isocommittor surface at which trajectories are equally likely to reach reactant and product. The committor thus yields the mechanism at finite temperature and accurate rate predictions in complex, high-dimensional systems. Committor functions can be parametrized from molecular dynamics (MD) trajectories, but this is notoriously difficult 6 ; 17 ; 49 , because the committor is a non-linear function of the high-dimensional molecular state space. Neural networks have been transformative in this context 44 ; 11 ; 35 , as they can efficiently represent high-dimensional, non-linear functions. Some of us recently developed “Artificial Intelligence for Molecular Mechanism Discovery” (AIMMD), 33 ; 41 a neural network that learns the committor of a rare event from short dynamically unbiased trajectories, which we generate using molecular simulations. A low-dimensional approximation of the committor is obtained by symbolic regression, a genetic algorithm that searches both functional form and parameters to best fit a data set. This provides a human-interpretable reaction mechanism expressed in terms of a few internal degrees of freedom, making complex reaction dynamics easier to understand. Figure 1: (a.) Lewis structure of retinal covalently linked to a lysine via a protonated Schiff base (N-retinylidene-lysine). The reaction center around the C 13 =C 14 double-bond is shown in red with atom labeling. (b.) Cis - trans isomerization around the C 13 =C 14 double-bond, with \theta_{1} as naive reaction coordinate marked in orange. (c.) Minimum energy path (black) and logarithmic transition path ensemble (green) of reactive trajectories projected into the space of the naive reaction coordinate \theta_{1} and the improper dihedral angles, \chi_{1} and \chi_{2} , centered at C 13 and C 14 , respectively. AIMMD has been successfully used to model a variety of diffusive molecular self-organization processes. 33 ; 41 ; 32 ; 57 ; 29 However, it remains unclear whether learning the committor for chemical reactions is possible or practical, for two reasons: 1: For the high free-energy barriers that characterize chemical reactions, the committor is essentially a step function that switches between 0 (reactant) and 1 (product) with only a narrow region around the TS yielding intermediate values. A model that accurately predicts only the isocommittor surface p_{B}(\mathbf{x})=0.5 will yield highly accurate predictions for most data points in a test data set. However, information on the reaction mechanism hinges on the how accurately the few data points in the intermediate committor region are represented. A model that accurately predicts only the isocommittor surface p_{B}(\mathbf{x})=0.5 misses crucial mechanistic information in the vast majority of the reactive region, as it does not capture twists and turns of the reactive channels. Learning away from the transition state becomes progressively more demanding, since ordering configurations along the reaction coordinate is increasingly difficult near the reactant and product basins, where the committor gradient vanishes. 7 2: In transitions with high energy barriers the transition event is very short. As a result, the assumption of instantaneous momentum relaxation may not hold during barrier crossing, and accurate description of the transition-path ensemble might require inertial dynamics. Consequently, a committor function that depends only on positions and neglects momenta might provide an incomplete description of the reaction. As a challenging test case for AIMMD, we model the thermal cis – trans isomerization around the C 13 =C 14 double bond in retinal. Specifically, we consider the protonated retinal Schiff base covalently linked to a lysine residue, i.e. N-retinylidene-lysine, in vacuum (Fig. 1 a and Fig. 6 ).