betaFeedback

WorksChemistry

TS Second Opinion: catching, repairing and triaging mirror-image errors in machine-learned transition states

Abstract

Generative models now predict a reaction’s transition state (TS) in milliseconds, but they cannot say when a guess is wrong. On the RDB7 benchmark, 17.7% of the TS guesses from GoFlow, a flow-matching TS generator, have a stereocentre the wrong way round, and 85% of those still count as accurate by the usual distance-based score. 90% of these errors are whole mirror images, which have the same energy and vibrational frequencies as the correct TS, so they also pass a frequency check. We present TS Second Opinion, a model-agnostic layer that runs after generation. Chemists state checks in plain English; a large language model (Claude) writes each check as Python, and a gate keeps it only if it accepts at least 99% of 9,540 reference DFT transition states, never crashes, and still catches mirrored or distorted ones. The accepted checks then test every guess, repair whole mirror images by reflection, and assign each reaction a verdict (trust, check with DFT, or re-generate); conformal selection turns the ranking into a set of trusted reactions with a stated average error rate. On 1,194 held-out RDB7 reactions with eight GoFlow models, two LLM-written stereo checks catch 93% of wrong-handed guesses at 0.2% false alarms, wrong-handed final answers fall from 17.7% to 2.3%, and 80% of reactions can be trusted at a 5% average geometry-error rate. New ωB97X/def2-SVP transition-state optimisations on 16 reactions support both the symmetry argument and the verdicts: an unrepaired mirror image converges to a first-order saddle point 0.0006 kcal/mol from the repaired one, all six trusted guesses reach the reference TS, and none of the six re-generate guesses do, although five of them still pass a DFT frequency check. The results are mostly retrospective, on one dataset and one model family; code, data and a browser app are openly available.

Introduction

Transition states set reaction rates, and finding them with density functional theory (DFT) is one of the more expensive routine steps in computational chemistry. Generative models such as GoFlow [1] propose a TS geometry from the reactant and product in milliseconds. They are usually scored by the mean absolute error of interatomic distances (D-MAE) or by RMSD after optimal alignment, and a guess with D-MAE ≤0.2\le0.2 Å is commonly counted as a success.

Neither score sees handedness. A reflection σ\sigma preserves every interatomic distance, so a mirror image of the reference TS has D-MAE exactly zero, and the RMSD code shared by several TS generators is called with the mirror image allowed. Because the electronic Hamiltonian is achiral, E(σR)=E(R)E(\sigma\mathbf{R}) = E(\mathbf{R}) and the Hessians are related by the same reflection, so a mirror image also has identical vibrational frequencies: an optimiser and a frequency check accept it as a genuine first-order saddle point. It is, however, the TS of the mirror-image reaction. FK-Flow [2] first measured these chirality errors in GoFlow and reduced them by Feynman–Kac steering during sampling, which requires access to the model. RitS [4] makes generation stereo-aware by encoding product chirality in the input. Both act inside one model.

This work asks a complementary question: given any generator’s existing output, can we detect and repair handedness errors after the fact, and say which guesses are safe to use without DFT? TS Second Opinion

Opinion (Figure 1) does this with checks that chemists write in plain English and that an LLM turns into code, gated on real reference data.

A near-perfect TS guess with the wrong handedness

Figure 1. A near-perfect GoFlow guess (RDB7 reaction 3509) that is the mirror image of the reference DFT transition state. Distance-based scores rate it as almost exact; only its handedness is wrong.

Method

Data. We use GoFlow’s released test-set predictions for the 1,194 reactions of RDB7’s random split [1, 3], from eight models trained with different seeds. All checks were written, tuned and gated using only the 9,540 RDB7 training reactions. The reference TS is used for evaluation only, never for a decision.

From a sentence to a gated check. A chemist writes a rule as one sentence (for example, “every stereocentre in the reactant that does not take part in the reaction keeps its handedness in the TS”). Claude writes a function check(rxn, xyz) that sees only what is known before any quantum chemistry: the reactant and product geometries, bonds and stereo tags. Code with imports, file access or double-underscore names is rejected. A gate then runs the check on all 9,540 reference TSs and on mirrored and noise-distorted copies. It is accepted only if it passes at least 99% of real TSs, never crashes, and flags at least 1% of the distorted ones; otherwise the failing examples go back to the model for a rewrite. Six checks were accepted, covering reacting bonds, stereocentres in the reactant and product, steric clashes, bond lengths and planarity; in the final run five passed first time and one needed two rewrites.

Repair and triage. For each reaction, every accepted check is run on all eight guesses, and the final answer is the consensus among guesses that pass. If a stereo check still fails, the mirror image is tried; it keeps every interatomic distance and flips every stereocentre, and it is used only if it then passes every check. Each reaction gets a verdict: trust; check with DFT for the 20% of reactions with the highest disagreement between models (mean D-MAE between the chosen guess and the other seven); or re-generate if no guess, original or mirrored, passes the checks.

Certified trust. To turn the ranking into a statement about error rates, we apply conformal selection [6]. Given calibration reactions with known outcomes, each new reaction receives a conformal pp-value from its trust score, and the Benjamini–Hochberg procedure selects a set SS such that, if new reactions are exchangeable with the calibration ones,

Measure (1,194 held-out RDB7 reactions, 8 GoFlow models)Result
GoFlow guesses with a stereocentre the wrong way round17.7%
of which still “accurate” by D-MAE ≤0.2 A˚\le0.2\,\text{Å}85%
of which whole mirror images90%
Wrong-handed guesses caught by two LLM-written stereo checks93% (0.2% false alarms)
Wrong-handed final answers: one model → after checks and repair17.7% → 2.3%
Final answers with a geometry error: one model → after checks14.1% → 11.7%
Geometry errors among trust / check with DFT / re-generate4.4% / 40% / 67%
Reactions trusted at a 5% average geometry-error target (observed rate)80% (4.9%)
Success with correct handedness: GoFlow → with TS Second Opinion70.8% → 86.5%

Table 1. Main retrospective results. Certified-trust figures are averaged over 2,000 random calibration/test splits.

E[∣Swrong∣max⁡(1,∣S∣)]≤α.\mathbb{E}\left[\frac{\lvert S_{\mathrm{wrong}}\rvert}{\max(1,\lvert S\rvert)}\right] \le\alpha.

Error definitions. We count two kinds of error against the reference TS separately: a geometry error (D-MAE >0.2 A˚> 0.2\,\text{Å}), which affects the barrier, and a handedness error (any checked stereocentre the wrong way round), which usually leaves the energy unchanged.

Results

Retrospective results on RDB7

Table 1 summarises the main results. Of the 1,690 wrong-handed guesses, 1,514 (90%) flip every stereocentre and are therefore whole mirror images; the rest flip only some stereocentres and correspond to a diastereomer, which reflection cannot fix. The checks repair handedness but not geometry (Figure 2); geometry errors are concentrated in the reactions sent to DFT (40% of check-with-DFT reactions versus 4.4% of trusted ones).

Share of final answers that are wrong, split by kind

Figure 2. Share of final answers that are wrong, split by kind. The checks and mirror repair reduce handedness errors; geometry errors are flagged for DFT rather than fixed.

The blind spot in the usual metrics. For the 447 test reactions with stereocentres, the mirror image of the reference TS scores D-MAE 0.000 Å and RMSD 0.000 Å with the mirror allowed (median proper-rotation RMSD 1.50 Å). GoFlow’s reported success rate (D-MAE ≤0.2\le0.2 Å) falls from 85.9% to 70.8% once correct

Certified trust by conformal selection: share of reactions trusted and observed error rate among them, for a range of targets.

caption: Certified trust by conformal selection: share of reactions trusted and observed error rate among them, for a range of targets.

handedness is also required. The effect is the same for every released GoFlow model size (82–85% of guesses correctly handed), so scaling the model does not remove it.

Certified trust. Over 2,000 random splits of the test reactions into calibration and new halves, the observed error rate among trusted reactions stayed at the target (Figure 3): 80% of reactions were trusted at a 5% geometry-error target (observed 4.9%) and 63% when both kinds of error count (observed 4.9%). The guarantee is an average; for a 5% target, 95% of individual draws stayed at or below 7.4%.

Generalisation. On RDB7’s reaction-core split, where test reaction types are not seen in training, the stereo checks catch 91% of wrong-handed guesses at 0.1% false alarms; on the barrier-height split, 89%. These splits are not fully independent of the gate, since the random-split training set used for gating contains about 80% of their test reactions. An independent 2026 study on Transition1x [5] reports that GoFlow, MolGen and TSDiff all lose 0.15–0.35 Å of RMSD once mirror images are no longer allowed, so the blind spot is not specific to one model or dataset.

Does the LLM matter? For the same six rules, checks written by Claude from the sentence and drafts written directly as code during development both passed the gate and performed about the same (geometry-wrong final answers 11.7% versus 11.6%; handedness 2.3% versus 1.4%). The LLM’s contribution is access, a tested check from one sentence for chemists who do not write code; what makes a check trustworthy is the gate. Asked to propose eight rules of its own, Claude produced plausible ones that all passed the gate, but none improved triage on held-out reactions, so none were kept.

A data audit. The same consistency logic, applied to the dataset itself, flags 51 of the 10,734 RDB7 reactions used (0.5%, always in the product geometry) whose bonding contradicts the atom-mapped product SMILES, for example a hydrogen bonded to a different carbon.

New DFT and GFN2-xTB calculations

To test the two central claims with new calculations, we optimised guesses to saddle points with DFT (ω\omegaB97X/def2-SVP in PySCF with density fitting, Sella [7] in internal coordinates, fmax⁡=0.05f_{\max} = 0.05 eV/Å, at most 80 steps) and computed analytic Hessians and harmonic frequencies. RDB7’s references are ω\omegaB97X-D3/def2-TZVP, so this is the same functional family with a smaller basis and no dispersion. A guess found the reference TS if it reached a first-order saddle point within 0.2 Å (D-MAE) of the RDB7 TS with the correct handedness.

A mirror image passes a DFT frequency check. For reaction 3509, whose best guess is a whole mirror image, both the unrepaired and the repaired guess converged to first-order saddle points (654i and 656i cm−1^{-1}),

0.0006 kcal/mol apart and both 0.018 Å (D-MAE) from the reference. The unrepaired one kept the wrong handedness; the repaired one is the reference TS.

The verdicts hold at DFT level. We optimised the recommended guess for the first six trusted reactions by id (excluding mirror-repaired ones), the first three check-with-DFT reactions, and all six re-generate reactions in the test set (Table 2, Figure 4). Every trusted guess converged in 5–14 steps to the reference TS. No re-generate guess did, yet five of the six still converged to a first-order saddle point, all with a stereocentre the wrong way round, so a frequency check would have accepted them. The check-with-DFT guesses sat in between: all three became first-order saddle points with the correct handedness, one reached the reference TS and two ended just outside the cut-off (0.206 and 0.231 Å).

VerdictFound the reference TSFirst-order saddle pointWrong handedness after optimisationMedian D-MAE to the RDB7 TS
Trust6 of 66 of 60 of 60.018 Å
Check with DFT1 of 33 of 30 of 30.206 Å
Re-generate0 of 65 of 65 of 60.258 Å

Table 2. New ω\omegaB97X/def2-SVP transition-state optimisations from the recommended guess, by verdict.

New DFT calculations showing verdicts and handedness

Figure 4. New DFT calculations. Each point is one optimised guess; position shows its distance to the RDB7 reference TS after optimisation, and the marker shows whether its handedness is correct.

A larger sample with GFN2-xTB. With GFN2-xTB (tblite) and the same protocol on 106 reactions, both the unrepaired and repaired guess reached first-order saddle points with the same energy (within 0.01 kcal/mol) and imaginary frequency (within 6 cm−1^{-1}) in 34 of 40 mirror-repaired reactions, and the unrepaired guess kept the wrong handedness in all 40. Recommended guesses found the reference TS for 77% of trusted, 60% of check-with-DFT and 0% of re-generate reactions (30, 30 and 6 reactions). A quick xTB optimisation rescued only 1 of 13 guesses that started with a geometry error, which supports sending flagged geometry to DFT.

Implementation and availability

TS Second Opinion is an MIT-licensed Python package with a command-line interface, a Model Context Protocol (MCP) server so that LLM agents can triage guesses, certify trust and propose new checks, and a marimo app that also runs entirely in the browser via WebAssembly. Users can check their own reactions and guesses from any generator and with any elements, score their own model’s test set with a stereo-aware metric, test whether a metric sees handedness, and audit their own dataset. Literature for this work was searched with the Valency MCP Bond server.

  • Code and data: https://github.com/tammy357/ts-second-opinion

  • Browser app: https://tammy357.github.io/ts-second-opinion/app/

  • Every number in this paper is reproduced by the scripts listed in the repository README.

Limitations

The headline results are retrospective: “right” means close to RDB7’s reference geometry. The DFT check is small (16 reactions, chosen as the first by id within each verdict rather than at random) and uses a smaller basis than RDB7’s references; it supports, rather than proves, the symmetry argument. All results use one dataset of small gas-phase H/C/N/O molecules and one model family; other generators were not run, and the checks should be re-gated on reference TSs from other chemistry before use there. Whether a mirror image counts as wrong depends on the use: for a barrier height alone it gives the right number. GoFlow’s released predictions were aligned to the reference before saving, which may make models look slightly more consistent with each other. Some design choices, including the product-stereocentre check and the 0.2 Å threshold, were made after seeing test results, although every check’s code and thresholds were fixed using training references only. The trust guarantee is an average that assumes new reactions resemble the calibration set; it should be recalibrated whenever the generator, data or checks change. Finally, LLM-written checks can be wrong in ways a gate cannot detect, so every check is kept as plain, readable Python.

Conclusion and outlook

Mirror-image transition states are a common failure of current generators that distance scores, chirality-blind RMSD and frequency checks all miss. A small set of gated, plain-English checks catches most of them, repairs whole mirror images at no cost, and turns a model ensemble into verdicts with a stated error rate, which new DFT calculations support. Natural next steps are a larger, random DFT sample at RDB7’s level of theory; other generators and datasets such as Transition1x; and using the same gated checks as steering potentials during generation, since Feynman–Kac steering does not require differentiable potentials. We also suggest that TS benchmarks report stereo correctness alongside D-MAE and RMSD.

Use of AI

This is AI-assisted research. Claude (Anthropic) wrote the gated checks from plain-English rules, as described above, and was used as a coding and writing assistant for the software, analyses and this manuscript. The author designed the study, chose the rules and checks to keep, and is responsible for all content and any errors.

Data and code availability

All code, data derived from GoFlow and RDB7, calculation outputs and figures are openly available at https://github.com/tammy357/ts-second-opinion under the MIT licence.

Competing interests

The author declares no competing interests.

Acknowledgements

The idea grew out of a question at Dr Esther Heid’s AIchemy webinar (October 2026); this work builds directly on the openly released GoFlow models and predictions and on FK-Flow from her group at TU Wien. I thank the organisers of the Open Scientific Intelligence Hackathon 2026, where this work was developed; marimo, whose open-source notebooks power the interactive and in-browser app; and Valency for access to MCP Bond. This work also relies on open-source software, including RDKit, PySCF, tblite (GFN2-xTB), ASE, Sella and 3Dmol.js.

References

  1. [1]L. Galustian, K. Mark, J. Karwounopoulos, M. P.-P. Kovar and E. Heid, GoFlow: efficient transition state geometry prediction with flow matching and E(3)-equivariant neural networks, Digital Discovery, 2025, doi:10.1039/D5DD00283D.DOI
  2. [2]K. Mark, L. Galustian, M. P.-P. Kovar and E. Heid, Feynman–Kac–Flow: inference steering of conditional flow matching to an energy-tilted posterior, arXiv:2509.01543, 2025.arxiv.org/abs/2509.01543
  3. [3]K. Spiekermann, L. Pattanaik and W. H. Green, High accuracy barrier heights, enthalpies, and rate coefficients for chemical reactions, Scientific Data, 2022, doi:10.1038/s41597-022-01529-6.DOI
  4. [4]F. Nikitin, D. M. Anstine and O. Isayev, Right into the Saddle: stereochemistry-aware generation of molecular transition states, ChemRxiv, 2026, doi:10.26434/chemrxiv.15001681.DOI
  5. [5]S. Baek, M. Gim and J. Kang, ReCurveflow: a flow matching framework that learns curved reaction trajectories to predict transition state geometries, arXiv:2608.20869, 2026.arxiv.org/abs/2608.20869
  6. [6]Y. Jin and E. J. Candès, Selection by prediction with conformal p-values, Journal of Machine Learning Research 24(244), 1–41, 2023.arxiv.org/abs/2210.01408
  7. [7]E. D. Hermes, K. Sargsyan, H. N. Najm and J. Zádor, Sella, an open-source automation-friendly molecular saddle point optimizer, Journal of Chemical Theory and Computation 18(11), 2022, doi:10.1021/acs.jctc.2c00395.DOI

Paper details

vlnc:W_g2uzaenx
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 15

downloads · 3

A view counts once per browser per day, a download each time the PDF is requested.

  • Chemistry
  • Advanced Physical and Chemical Molecular Interactions
  • Chemical Reaction Mechanisms
  • Chemistry and Stereochemistry Studies

Claude Opus (multiple versions)

View statement

Sisodiya, T. (2026). TS Second Opinion: catching, repairing and triaging mirror-image errors in machine-learned transition states. Valency, vlnc:W_g2uzaenx, v1. https://hub.valency.io/works/W_g2uzaenx

Looking for related papers…
Contents