Right into the Saddle: Stereochemistry-Aware Generation of Molecular Transition States

Filipp Nikitin^1,2   ·   Dylan M. Anstine³   ·   Olexandr Isayev^1,2*
1Ray and Stephanie Lane Computational Biology Department, Carnegie Mellon University, Pittsburgh, PA, USA
²Department of Chemistry, Carnegie Mellon University, Pittsburgh, PA, USA
³Department of Chemical Engineering and Materials Science, Michigan State University, East Lansing, MI, USA

📄 Paper   ·   📖 Citation   ·   ⚙️ Setup   ·   🔗 GitHub

^*Corresponding author: olexandr@olexandrisayev.com

---

Overview

Abstract

Determination of transition state structures is fundamental to the mechanistic understanding of chemical reactivity and selectivity, yet conventional approaches require extensive expert intervention for geometry preparation and saddle-point optimization, severely limiting the throughput of computational reaction analysis. We present RitS, a generative model trained on a curated dataset of approximately 1.7 million transition states computed at the GFN2-xTB level of theory. RitS generates three-dimensional transition-state geometries directly from reactant–product bond connectivity without requiring pre-optimized molecular structures as input. The model accommodates H, B, C, N, O, F, Si, P, S, Cl, Br, I elements, systems containing up to 51 total atoms, and both neutral and charged species, substantially extending the chemical scope of prior generative approaches. Benchmarking on the same TSDiff/Grambow test set demonstrates that RitS, both after benchmark fine-tuning and when trained from scratch on the benchmark, generates structures closer to the reference geometries than TSDiff and GoFlow. A key capability of RitS is stereochemistry-aware generation: encoding product chirality in the input graph enables selective generation of competing diastereomeric transition states, as demonstrated for endo versus exo Diels–Alder cycloadditions and E versus Z elimination pathways. The model further extends to the multistep Hajos–Parrish–Eder–Sauer–Wiechert organocatalytic cycle, capturing each elementary step including the enantiodetermining intramolecular aldol transition state. Comparison of competing transition states refined with UMA also reproduces the experimentally favored products in seven stereoselective and regioselective reactions. These results establish RitS as a tool for automated, high-throughput transition-state generation applicable to mechanistic studies, selectivity analysis, and large-scale reaction network exploration.

Key Features

• Direct transition-state generation: Generates 3D transition-state geometries directly from reactant and product connectivity without requiring hand-crafted initial guesses.
• Stereochemistry-aware generation: Supports explicit stereochemical control, enabling generation of transition states corresponding to competing pathways such as endo/exo and E/Z configurations.
• Flow-matching generative model: Uses a flow-matching objective with Kabsch-aligned optimal transport for stable and efficient molecular geometry generation.
• Large-scale training dataset Trained on a dataset of ~1.7 million transition-state reactions constructed through mechanistic enumeration and active learning.

---

Setup

Installation will usually take up to 20 minutes.

System and Hardware Requirements

• OS tested by authors:
  • Ubuntu 24.04 LTS
• Other platforms:
  • Expected to work, but if installation is not out-of-the-box, use the PyTorch Geometric installation guide for your exact Python/PyTorch/CUDA combination: https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html
• Tested inference hardware:
  • GPU: NVIDIA RTX 3090 (24 GB VRAM)
  • CPU: AMD Ryzen 9 5950X
• Recommended GPU memory:
  • 16-24 GB VRAM for comfortable inference/evaluation with larger molecules and higher batch sizes
• Minimum practical GPU memory:
  • 8 GB VRAM can run inference, but requires reduced batch sizes
• CPU-only:
  • Supported; the checkpoint instructions include tested CPU inference commands

OOM mitigation for larger molecules:

• reduce inference batch size (--batch_size in sampling, or data.inference_batch_size in config)

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (recommended for training)
• Conda or Mamba (recommended)

Environment Setup

# Clone the repository
git clone https://github.com/isayevlab/rits.git
cd RitS

# Create and activate conda environment
conda create -n rits python=3.10 -y
conda activate rits

# Install dependencies
pip install -r requirements.txt
pip install -e .

If you prefer a fully conda-based setup (recommended for RDKit), you can install RDKit via conda-forge before running pip install -r requirements.txt.

The pinned dependencies include h5py, numcodecs, and the Zarr v2 API used by the HDF5 preprocessing pipeline.

For the interactive Streamlit app and IRC post-processing, also install:

pip install -r app/requirements.txt
conda install -c conda-forge xtb

Data Setup

Release folder: RitS on Google Drive. The files are in its data subfolder.

Material	Download	Description
Training dataset	rgd1_chembl_extended.h5	Approximately 1.7 million GFN2-xTB transition states
Original RitS checkpoint	rits.ckpt	GFN2-xTB-trained model used for the broader reaction and selectivity examples
Fine-tuned RitS checkpoint	rits_finetuned.ckpt	GFN2-xTB pretraining followed by TSDiff/Grambow benchmark fine-tuning
RitS-TSDiff checkpoint	rits_tsdiff.ckpt	Trained from scratch on the TSDiff/Grambow benchmark
Reaction examples	reactions	Mapped inputs and a portable runner for the diverse reactions, Hajos cycle and selectivity channels
Selectivity results	selectivity	Compressed UMA geometries, IRC trajectories and thermochemistry
Benchmark source data	Transition1x	TSDiff/Grambow source files; legacy folder name retained
Processed benchmark data	ts1x	Preprocessed TSDiff/Grambow tensors; legacy folder name retained
Checksums	SHA256SUMS	SHA-256 checksums for the release files

For inference, download the desired checkpoint into data/; the training datasets are not required. See Running the released checkpoints for each model's matching configuration, preprocessing flags, runnable commands and SHA-256 checksum. The original and fine-tuned checkpoints use rits.yaml with --kekulize --add_stereo; rits_tsdiff.ckpt uses ts1x_rits.yaml without these flags. The older ts1x_rits.ckpt is not the selected rits_tsdiff.ckpt.

The TSDiff/Grambow benchmark and Transition1x are distinct datasets. Existing Transition1x and ts1x file/configuration names are retained for compatibility.

Expected local layout after downloading the selected files and preparing the datasets:

data/
├── rits.ckpt
├── rits_finetuned.ckpt
├── rits_tsdiff.ckpt
├── rgd1_chembl_extended.h5
├── reactions/
├── selectivity/
│   └── uma_results.tar.gz
├── rits_dataset/
│   └── processed/
│       ├── train_h.pt
│       ├── val_h.pt
│       └── test_h.pt
├── Transition1x/
│   └── wb97xd3/
│       ├── feat_dict.pkl
│       └── raw_data/
│           ├── wb97xd3_ts.xyz
│           └── wb97xd3_fwd_rev_chemprop.csv
└── ts1x/
    └── processed/
        ├── train_h.pt
        ├── val_h.pt
        └── test_h.pt

The rits_dataset/processed/ directory is generated by preprocessing the HDF5 file. Download the processed benchmark tensors or regenerate ts1x/processed/ from the TSDiff/Grambow source files using the commands below.

RGD1-ChEMBL-extended

Prepare the large dataset used by RitS:

python data_processing/prepare_rgd1_chembl_extended.py \
    --h5_file data/rgd1_chembl_extended.h5 \
    --save_data_folder data/rits_dataset

This stage creates random 96%/2%/2% train/validation/test splits. By default it kekulizes aromatic systems into explicit single/double bonds and adds the model's stereochemical pseudo-edges. Training stereochemistry is assigned from the reactant and product 3D coordinates in the HDF5 file. During inference, the same edge encoding is constructed from stereochemistry specified in the input SMARTS. These defaults match the released rits.ckpt; they can be disabled for experiments with --no-kekulize or --no-add-stereo.

TSDiff/Grambow benchmark

Preprocess the benchmark source files for training (legacy paths shown below):

python data_processing/prepare_ts1x_for_training.py \
    --ts_data data/Transition1x/wb97xd3/raw_data/wb97xd3_ts.xyz \
    --rxn_smarts_file data/Transition1x/wb97xd3/raw_data/wb97xd3_fwd_rev_chemprop.csv \
    --save_dir data/ts1x

---

Usage

Lightweight runnable examples are available in examples/, with one folder per reaction family (reaction scheme image(s) + cmd.sh).

Model Training

# Train RitS from scratch on preprocessed RGD1-ChEMBL-extended
python scripts/train.py \
    --config-path=conf \
    --config-name rits \
    train.gpus=1 \
    train.seed=28 \
    run_name=test_train \
    outdir="../test_runs" \
    data.dataset_root="./data/rits_dataset"

# Train RitS-TSDiff on the TSDiff/Grambow benchmark instead
python scripts/train.py \
    --config-path=conf \
    --config-name ts1x_rits \
    train.gpus=1 \
    train.seed=28 \
    run_name=ts1x_train \
    outdir="../test_runs" \
    data.dataset_root="./data/ts1x"

# Resume training from checkpoint
python scripts/train.py \
    --config-path=conf \
    --config-name rits \
    train.gpus=1 \
    train.seed=28 \
    run_name=test_train \
    outdir="../test_runs" \
    resume="./data/rits.ckpt"

# Customize training parameters
python scripts/train.py \
    --config-path=conf \
    --config-name rits \
    outdir=./outputs \
    train.gpus=2 \
    train.n_epochs=800 \
    train.seed=42 \
    data.batch_size=150 \
    optimizer.lr=0.0001

Model Inference and Sampling

Transition States Generation

Use atom-mapped SMILES for the released checkpoint commands so their preprocessing matches training.

reactant='[C:1]([c:2]1[n:3][o:4][n:5][n:6]1)([H:7])([H:8])[H:9]'
product='[C:1]1([H:7])([H:8])/[C:2](=[N:3]\[H:9])[N:6]1[N:5]=[O:4]'

# Generate a transition-state guess from atom-mapped SMILES
python scripts/sample_transition_state.py \
    --reactant_smi "$reactant" --product_smi "$product" \
    --config scripts/conf/rits.yaml \
    --ckpt data/rits.ckpt \
    --kekulize --add_stereo \
    --output output.xyz \
    --n_samples 1 \
    --batch_size 32

# Generate multiple samples for the same mapped reaction
python scripts/sample_transition_state.py \
    --reactant_smi "$reactant" --product_smi "$product" \
    --config scripts/conf/rits.yaml \
    --ckpt data/rits.ckpt \
    --kekulize --add_stereo \
    --output ts_samples.xyz \
    --n_samples 10 \
    --batch_size 32

Input formats:

• SMILES: Atom-mapped SMILES with explicit hydrogens (e.g., [C:1][H:2])
• XYZ: Standard XYZ coordinate files (bonds will be inferred using OpenBabel)

Notes:

• SMILES must have explicit hydrogens; atom maps should identify the same atoms on both sides
• Reactant and product must contain the same atoms; XYZ inputs must use matching atom order
• Output is saved as XYZ file(s) with transition state coordinates

Available Configurations

Transition State Configs:

• rits.yaml - RitS stereochemistry-aware transition-state model trained on our GFN2-xTB dataset; also used by the fine-tuned checkpoint
• ts1x_rits.yaml - RitS-TSDiff configuration for TSDiff/Grambow training/evaluation

Interactive App

We also provide an interactive Streamlit web application to make it easier to use RitS without writing any code. The app allows you to select from built-in example reactions or enter custom reaction SMARTS, generate transition states, visualize 2D reaction schemes and 3D transition-state geometries, and optionally run IRC post-processing with pysisyphus.

# Install app dependencies (if not already done)
pip install -r app/requirements.txt
conda install -c conda-forge xtb

# Run the app
streamlit run app/app.py

See app/README.md for more details.

Training Configuration

You can easily override configuration parameters:

# Example with custom parameters
python scripts/train.py \
    --config-path=conf \
    --config-name rits \
    outdir=./my_training \
    run_name=my_experiment \
    train.gpus=4 \
    train.n_epochs=500 \
    data.batch_size=64 \
    data.dataset_root="./data/rits_dataset" \
    wandb_params.mode=online

---

Citation

RitS preprint citation:

@article{
rits,
author = {Filipp Nikitin  and Dylan M. Anstine  and Olexandr Isayev },
title = {Right into the Saddle: Stereochemistry-Aware Generation of Molecular Transition States},
journal = {ChemRxiv},
volume = {2026},
number = {0406},
pages = {},
year = {2026},
doi = {10.26434/chemrxiv.15001681/v1},
URL = {https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15001681/v1},
eprint = {https://chemrxiv.org/doi/pdf/10.26434/chemrxiv.15001681/v1},
}

You may also find useful our previous LoQI paper:

@article{loqi,
author = {Filipp Nikitin  and Dylan M. Anstine  and Roman Zubatyuk  and Saee Gopal Paliwal  and Olexandr Isayev },
title = {Scalable Low-Energy Molecular Conformer Generation with Quantum Mechanical Accuracy},
journal = {ChemRxiv},
volume = {2025},
number = {0820},
pages = {},
year = {2025},
doi = {10.26434/chemrxiv-2025-k4h7v},
URL = {https://chemrxiv.org/doi/abs/10.26434/chemrxiv-2025-k4h7v},
eprint = {https://chemrxiv.org/doi/pdf/10.26434/chemrxiv-2025-k4h7v}}

This work builds upon the Megalodon architecture. If you use the underlying architecture, please also cite:

@article{reidenbach2026applications,
  title={Applications of modular co-design for de novo 3d molecule generation},
  author={Reidenbach, Danny and Nikitin, Filipp and Isayev, Olexandr and Paliwal, Saee Gopal},
  journal={Digital Discovery},
  year={2026},
  publisher={Royal Society of Chemistry}
}