Predicting the products of chemical reactions is a task of combinatorial complexity. Traditional DFT calculations for a single conformation take hours, and enumerating all possible synthesis paths takes weeks. We solve this problem with ML models: we predict products in seconds with up to 92% accuracy and build force fields with MAE less than 50 meV/Å, accelerating calculations tens of thousands of times. Savings on DFT calculations reach 70%, significantly reducing operational costs at typical modeling volumes.
How ML Predicts Reaction Products
Formulation via atom mapping. A reaction is a redistribution of atoms: some bonds break, others form. The task is to predict the reaction center. Molecular Transformer (Schwaller et al.) — SMILES of reactants are concatenated with a > separator, and the Transformer (like T5) generates SMILES of products. On USPTO-480k: Top-1 accuracy ~90.4%.
Limitation: the Transformer may output an invalid molecule. Alternative — WLDN and LocalTransform: graph-oriented models that predict changes in the graph. They guarantee valid products. Top-1 accuracy of graph models reaches 92%.
Regioselectivity. Aromatic nitration of naphthalene yields products at positions 1 or 2. The model must predict not only the product but also the isomer ratio. This is solved by integrating QM descriptors (Fukui coefficients, electrostatic potential maps) into GNNs via Coulomb Matrix or SchNet-encoder.
Why ML Potentials Are Faster Than DFT
DFT molecular dynamics (MD) requires 10^5–10^6 steps with a step of 0.5–1 fs — calculating each step with DFT is computationally unfeasible. ML potentials (MLFF) are trained on DFT data and predict energies/forces 1000–10000× faster. Speedup reaches 8000× with error ΔG‡ ~1.2 kcal/mol relative to CCSD(T).
Modern architectures: NequIP (equivariant GNN), MACE (faster than NequIP, good coverage of the periodic table), M3GNet (for solids). Process: DFT calculations (ORCA, VASP) → dataset (energies + gradients) → MLFF training → validation (force MAE <50 meV/Å).
| Architecture | Speed | Element Coverage | Force MAE |
|---|---|---|---|
| MACE | 10^4× faster than DFT | up to 89 elements | <30 meV/Å |
| NequIP | 10^3× faster than DFT | up to 54 elements | <40 meV/Å |
| M3GNet | 10^3× faster than DFT | solids | <50 meV/Å |
In practice: when modeling the acylation reaction in acetonitrile, MACE-MLFF reproduced ΔG‡ with an error of ~1.2 kcal/mol relative to CCSD(T) at a speedup of 8000×. A typical mistake when training MLFF is insufficient coverage of conformations in the training dataset, leading to high MAE (>50 meV/Å). The solution is active learning with query-by-committee to select representative conformations.
What Retrosynthetic Planning Includes
AiZynthFinder (MIT/AstraZeneca) uses MCTS to search for paths, applying reaction templates or Molecular Transformer in reverse mode. Template-free alternatives — GraphRetro, Retrosim: they predict synthons directly as graph transformations. Coverage is broader but interpretability is lower.
Workflow: target molecule → AiZynthFinder (5–50 paths in seconds) → filter by commercial availability (Sigma-Aldrich/Enamine API) → score by length, reagent cost, predicted yield.
What the Work Includes
- Problem analysis: choose model type (reaction prediction, retrosynthesis, MLFF), collect/generate dataset.
- Training and validation: configure architecture, hyperparameters, evaluate on holdout set.
- Deployment: inference via REST API (FastAPI) with batch support, documentation (model card).
- Integration: connect to your pipeline (Python SDK, Docker).
- Team training: handover of model, documentation, 2 weeks support.
Work Process
Analysis → architecture design → data collection (DFT calculations, datasets) → training → test (MAE/accuracy) → deployment (Triton, SageMaker). Each stage is documented in a report with metrics and recommendations.
Tech Stack
| Task | Tools |
|---|---|
| Reaction prediction | Molecular Transformer, LocalTransform, RXNMapper |
| Retrosynthesis | AiZynthFinder, ASKCOS, RXN4Chemistry |
| MLFF | NequIP, MACE, TorchANI, M3GNet |
| MD engine | OpenMM, LAMMPS, ASE |
| QM reference | ORCA, Psi4, xTB (semi-empirical) |
| Molecular graph | PyTorch Geometric, DGL-LifeSci |
Timelines
- MVP for product prediction on an existing dataset: 8–12 weeks.
- Custom MLFF with DFT calculations: 3–6 months.
- End-to-end retrosynthesis integration: 4–8 weeks.
Contact us — we will assess your task, choose the architecture and accuracy for your budget. Over 5 years of experience in chemical AI, 20+ projects. Get a consultation on model selection for your dataset. Order model development — we will prepare a technical specification in 2 days.