TAckling Chemistry through ImpliciT negative learning
Nick Wierich*, Niklas Dobberstein*, Frank Glorius†
Organisch-Chemisches-Institut, Universität Münster, 48149 Münster, Germany
* These authors contributed equally. · † Corresponding author.
Machine learning models for reaction product prediction have largely been trained and evaluated on datasets of successful reactions. While this setting allows reaching high accuracy scores, these models generally struggle to discriminate between successful and unsuccessful transformations due to the lack of negative training data. This critically limits their practical use, as reliably recognizing unproductive substrate combinations in silico is required to accelerate experimental campaigns and save resources in the laboratory. While the availability of negative reaction data remains limited to date, we show that even purely positive datasets implicitly contain a significant amount of information about cases in which no reaction occurs. To exploit this, we introduce TACIT (TAckling Chemistry through ImpliciT negative learning). While trained on positive-only data, TACIT learns to distinguish between productive and unproductive transformations. On two published high-throughput experimentation datasets, it outperforms all evaluated reaction outcome prediction models in separating successful from failed reactions. Crucially, TACIT's accuracy on the popular USPTO-MIT benchmark remains competitive, despite its added ability to predict negative reaction outcomes.
Install, download the checkpoints and data, and evaluate TACIT on the USPTO-MIT test set:
git clone https://github.com/GloriusGroup/tacit.git && cd tacit
uv sync
uv run python download_zenodo.py
uv run python src/bond_match/evaluate_ranking.py --config runs/tacit/config.yaml --checkpoint runs/tacit/checkpoints/tacit.pt --data data/uspto_mit/test.jsonl --all-entries --save-csv tacit_uspto.csv| Model | USPTO-MIT top-1 [%] | Chemlex ROC-AUC / PR-AUC | AIChemEco ROC-AUC / PR-AUC |
|---|---|---|---|
| TACIT | 77.8 | 0.80 / 0.56 | 0.61 / 0.69 |
| MEGANa | 80.5 | 0.69 / 0.38 | 0.47 / 0.56 |
| FlowERb | 42.8 | 0.50 / 0.26 | 0.51 / 0.60 |
| Molecular Transformera,d | 81.7 | 0.68 / 0.42 | 0.51 / 0.61 |
| LocalTransforma,d | 80.4 | 0.64 / 0.37 | 0.51 / 0.59 |
| ReactionT5a,c,d | 95.1e | 0.69 / 0.41 | 0.47 / 0.58 |
USPTO-MIT: all 40,000 test reactions. HTE datasets: all 11,669 (Chemlex) and 47,015 (AIChemEco) wells. A well is positive if its conversion or yield is at least 20%, and negative otherwise.
a Retrained on the blockspace-filtered USPTO-MIT training set used for TACIT.
b Published checkpoint. Not retrained, since a mechanistic dataset is required for training.
c Pretrained on the Open Reaction Database.
d No stop action. The confidence of the top-1 prediction is used as the score.
e 34,552 of the 40,000 test reactions (86.4%) are contained in the ORD pretraining data. The top-1 accuracy is 97.5% on these and 80.4% on the remaining 5,448 reactions.
Requires Python 3.11 and a CUDA 12.4 environment. The project uses uv.
git clone https://github.com/GloriusGroup/tacit.git && cd tacit
uv syncThe checkpoints and the dataset are archived on Zenodo. Download them into the right folders with
uv run python download_zenodo.pyor place them manually:
| File | Content | Place at |
|---|---|---|
tacit.pt |
TACIT weights (base model) | runs/tacit/checkpoints/ |
tacit_no-stop.pt |
TACITno-stop weights (trained without the auxiliary stop loss) | runs/tacit_no-stop/checkpoints/ |
uspto_mit_blockspace.zip |
USPTO-MIT train, validation and test splits with blockspace decompositions | data/uspto_mit/ (and unzip) |
The data is computed from the original USPTO-MIT splits of Jin et al. (2017).
wget https://github.com/wengong-jin/nips17-rexgen/raw/master/USPTO/data.zip -O uspto.zip
unzip -j uspto.zip -d uspto && (echo rxn_smiles; cut -d' ' -f1 uspto/train.txt) > uspto/train.csv
uv run python src/bond_match/blockspace/batch.py uspto/train.csv -c rxn_smiles -o uspto/train.jsonl --strict --mixed --remove_stereo --clean-coo -w 32 -t 10The Zenodo record also contains the retrained baseline models (MEGAN, Molecular Transformer, LocalTransform and ReactionT5), each with its upstream license and a note on how it was trained.
uv run python src/bond_match/evaluate_ranking.py --config runs/tacit/config.yaml --checkpoint runs/tacit/checkpoints/tacit.pt --data data/uspto_mit/test.jsonl --all-entries --save-csv tacit_uspto.csvThe HTE datasets are not redistributed here. Download them from the original
authors and convert each into a wells.csv:
| Dataset | Source file | License |
|---|---|---|
| Chemlex (Zhong et al., Nat. Commun. 2025) | Chemlex_Acidamine_Wetlab_Data.xlsx (Zenodo, version of Nov 2025) |
CC BY-NC 4.0 |
| AIChemEco (Zhang et al., Chem. Sci. 2025) | all_HTE_with_condition.csv (GitHub, commit fa29d55) |
MIT |
wget https://zenodo.org/records/17596563/files/Chemlex_Acidamine_Wetlab_Data.xlsx
wget https://github.com/aichemeco/amide_coupling/raw/fa29d55/data/all_HTE_with_condition.csv
uv run python src/bond_match/hte_wells.py chemlex Chemlex_Acidamine_Wetlab_Data.xlsx chemlex.csv
uv run python src/bond_match/hte_wells.py aichemeco all_HTE_with_condition.csv aichemeco.csvThen score the wells:
uv run python src/bond_match/score_hte.py --config runs/tacit/config.yaml --checkpoint runs/tacit/checkpoints/tacit.pt --wells chemlex.csv --label-column conversion_pct --out tacit_chemlex.csv
uv run python src/bond_match/score_hte.py --config runs/tacit/config.yaml --checkpoint runs/tacit/checkpoints/tacit.pt --wells aichemeco.csv --label-column yield_pct --out tacit_aichemeco.csvuv run python src/bond_match/evaluate_ranking.py --config runs/tacit/config.yaml --checkpoint runs/tacit/checkpoints/tacit.pt=== Ranking Model Eval Menu ===
1. Interactive SMILES Prediction (Top 5)
2. Evaluate on Test Set (test.jsonl)
3. Random Sample from Test Set (Visual)
q. Quit
Select option: 3
Chosen Reaction: [Br:1][c:2]1[cH:3][cH:4][c:5]([F:22])[c:6]([CH:8]([C:9]([CH2:10][O:11][Si:12]([CH3:13])([CH3:14])[C:15]([CH3:16])([CH3:17])[CH3:18])([F:19])[F:20])[OH:21])[cH:7]1.[Cl:44][CH2:45][Cl:46].[Cr:23]([O:24][Cr:25]([O-:26])(=[O:27])=[O:28])([O-:29])(=[O:30])=[O:31].[nH+:32]1[cH:33][cH:34][cH:35][cH:36][cH:37]1.[nH+:38]1[cH:39][cH:40][cH:41][cH:42][cH:43]1>>[Br:1][c:2]1[cH:3][cH:4][c:5]([F:22])[c:6]([C:8]([C:9]([CH2:10][O:11][Si:12]([CH3:13])([CH3:14])[C:15]([CH3:16])([CH3:17])[CH3:18])([F:19])[F:20])=[O:21])[cH:7]1
--- Initial Step Analysis for: CC(C)(C)[Si](C)(C)OCC(F)(F)C(O)c1cc(Br)ccc1F.ClCCl.[O]=[Cr](=[O])([O-])[O][Cr](=[O])(=[O])[O-].c1cc[nH+]cc1.c1cc[nH+]cc1 ---
Rank Score Prob Action Match?
----------------------------------------------------------------------
1 25.2775 99.91 H#0-O:21 | H#29-C:8 ✅ 0 (0: 96%, 1: 4%)
2 18.2333 0.09 is_over ❌
3 1.7181 0.00 H#0-O:21 | H#18-C:16 ❌ 0 (0: 89%, 1: 11%)
4 0.8446 0.00 H#0-O:21 | O:21-C:8 ❌ 0 (0: 90%, 1: 10%)
5 0.7310 0.00 H#0-O:21 | C:8-C:9 ❌ 0 (0: 82%, 1: 18%)
>>> Continuing to step 2 from current product: [H][H].[H][n+:32]1[c:33]([H])[c:34]([H])[c:35]([H])[c:36]([H])[c:37]1[H].[H][n+:38]1[c:39]([H])[c:40]([H])[c:41]([H])[c:42]([H])[c:43]1[H].[H][c:3]1[c:2]([Br:1])[c:7]([H])[c:6]([C:8]([C:9]([C:10]([H])([H])[O:11][Si:12]([C:13]([H])([H])[H])([C:14]([H])([H])[H])[C:15]([C:16]([H])([H])[H])([C:17]([H])([H])[H])[C:18]([H])([H])[H])([F:19])[F:20])=[O:21])[c:5]([F:22])[c:4]1[H].[H][C:45]([H])([Cl:44])[Cl:46].[Cr:23]([O:24][Cr:25]([O-:26])(=[O:27])=[O:28])([O-:29])(=[O:30])=[O:31]
--- Iterative Step 2 ---
Rank Score Prob Action Match?
----------------------------------------------------------------------
1 20.3534 100.00 is_over ✅
2 0.0574 0.00 O:11-Si:12 | Cl:44-LP ❌ 1 (0: 39%, 1: 61%)
3 0.0528 0.00 H#30-C:13 | Cl:44-LP ✅ 0 (0: 55%, 1: 45%)
4 0.0422 0.00 H#14-N:32 | Cl:44-LP ✅ 0 (0: 52%, 1: 48%)
5 0.0421 0.00 H#0-H#29 | Cl:44-LP ✅ 1 (0: 44%, 1: 56%)
>>> Stopping: Model selected is_over
--- Final Result ---
Steps applied: 1
Stop reason: Model selected is_over stop action
Pred Product: [Cr:23]([O:24][Cr:25]([O-:26])(=[O:27])=[O:28])([O-:29])(=[O:30])=[O:31].[H][C:45]([H])([Cl:44])[Cl:46].[H][c:3]1[c:2]([Br:1])[c:7]([H])[c:6]([C:8]([H])([C:9]([C:10]([H])([H])[O:11][Si:12]([C:13]([H])([H])[H])([C:14]([H])([H])[H])[C:15]([C:16]([H])([H])[H])([C:17]([H])([H])[H])[C:18]([H])([H])[H])([F:19])[F:20])[O:21][H])[c:5]([F:22])[c:4]1[H].[H][n+:32]1[c:33]([H])[c:34]([H])[c:35]([H])[c:36]([H])[c:37]1[H].[H][n+:38]1[c:39]([H])[c:40]([H])[c:41]([H])[c:42]([H])[c:43]1[H]>>[H][H].[H][n+:32]1[c:33]([H])[c:34]([H])[c:35]([H])[c:36]([H])[c:37]1[H].[H][n+:38]1[c:39]([H])[c:40]([H])[c:41]([H])[c:42]([H])[c:43]1[H].[H][c:3]1[c:2]([Br:1])[c:7]([H])[c:6]([C:8]([C:9]([C:10]([H])([H])[O:11][Si:12]([C:13]([H])([H])[H])([C:14]([H])([H])[H])[C:15]([C:16]([H])([H])[H])([C:17]([H])([H])[H])[C:18]([H])([H])[H])([F:19])[F:20])=[O:21])[c:5]([F:22])[c:4]1[H].[H][C:45]([H])([Cl:44])[Cl:46].[Cr:23]([O:24][Cr:25]([O-:26])(=[O:27])=[O:28])([O-:29])(=[O:30])=[O:31]
Top-1 Correct? Trueuv run bash run_train_multigpu.sh runs/tacit/config.yaml 4src/bond_match/
├── train_ranking.py # training entry point (training/ package, DDP and single GPU)
├── evaluate_ranking.py # evaluation entry point (evaluation/ package, batch and interactive)
├── hte_wells.py # converts the published HTE files into wells.csv
├── score_hte.py # HTE well scoring and metrics
├── training/ # trainer, EMA, DDP, loss and optimizer helpers
├── evaluation/ # evaluator, BE-step operations, inference data, SMILES utilities
├── model/ # BondGAT encoder, Transformer, heads, scorers
├── modules/ # MoE, attention residuals, LoRA, Muon optimizer, mHC
├── losses/ # pair-ranking and energy-based losses
├── blockspace/ # Bond-Electron matrix algebra, reaction decomposition, batch.py
├── dataset/ # sharded dataset: molecular graphs, reaction processing, collate
├── config.py # RankingConfig schema and YAML loader
└── sampler.py # shard-aware and weighted sampling
runs/ # config.yaml per model (checkpoints/ from Zenodo)
data/uspto_mit/ # dataset files (from Zenodo)
docs/ # project page (GitHub Pages)
@article{tacit,
title = {Implicit Learning: Reaction Outcome Prediction from Positive-Only Data},
author = {Wierich, Nick and Dobberstein, Niklas and Glorius, Frank},
journal = {ChemRxiv},
year = {2026},
doi = {10.26434/chemrxiv.15011216/v1},
note = {TACIT: TAckling Chemistry through ImpliciT negative learning}
}The code, the TACIT checkpoints and the data on Zenodo are released under the
MIT License (see LICENSE). The USPTO-MIT splits are by
Jin et al. (2017) and were
released under the MIT License. The retrained baseline models keep their
upstream licenses.