View PDF Version

Open Access Article
This Open Access Article is licensed under a
Creative Commons Attribution 3.0 Unported Licence
DOI: 10.1039/D5CC02641E (Communication) Chem. Commun., 2025, Advance Article

Enhancing deep chemical reaction prediction with advanced chirality and fragment representation

Fabrizio Mastroloritoab, Fulvio Ciriacoc, Orazio Nicolottib and Francesca Grisoni*a
aDepartment of Biomedical Engineering, Institute for Complex Molecular Systems (ICMS) & Eindhoven AI Systems Institute (EAISI), Eindhoven University of Technology, Eindhoven, The Netherlands. E-mail: f.grisoni@tue.nl
bDipartimento di Farmacia-Scienze del Farmaco, Università degli Studi di Bari Aldo Moro, Bari, Italy
cDipartimento di Chimica, Università degli Studi di Bari Aldo Moro, Bari, Italy

Received 13th May 2025 , Accepted 25th August 2025

First published on 26th August 2025

Abstract

This work focuses on organic reaction prediction with deep learning, with the recently introduced fragSMILES representation – which encodes molecular substructures and chirality, enabling compact and expressive molecular representation in a textual form. In a systematic comparison with well-established molecular notations – simplified molecular input line entry system (SMILES), self-referencing embedded strings (SELFIES), sequential attachment-based fragment embedding (SAFE) and tree-based SMILES (t-SMILES) – fragSMILES achieved the highest performance across forward- and retro-synthesis prediction, with superior recognition of stereochemical reaction information. Moreover, fragSMILES enhances the capacity to capture stereochemical complexity – a key challenge in synthesis planning. Our results demonstrate that chirality-aware and fragment-level representations can advance current computer-assisted synthesis planning efforts.

Since time immemorial, operating a chemical laboratory has required patience and meticulous attention to detail, often resulting in long timelines and inconclusive outcomes. In the last decades, artificial intelligence has increasingly supported chemists in expediting their experiments, through machine learning algorithms for process and molecule optimization1–3 and robotics-assisted laboratories that streamline the execution.4,5 Among these advances, computer-assisted synthesis planning has been particularly transformed by the advent of deep learning,6,7 which has demonstrated high accuracy and has significantly reduced the time and resources required compared to traditional trial-and-error approaches.7–9

Methods based on string representations of chemicals and organic reactions have gained particular traction,10 thanks to their ability to leverage natural language processing techniques.11,12 In particular, reactants (or product) molecules are represented as strings, to subsequently predict the product (or reactants) molecules using machine translation models.9,13 Popular string notations for synthesis planning13–17 include the simplified molecule input line entry system (SMILES18) strings, self-referencing embedded strings (SELFIES19), sequential attachment-based fragment embedding (SAFE20) and tree-based SMILES (t-SMILES21).

As chemical reactions involve local molecular changes (leading to a significant overlap of reactants and products), several methods have focused on substructure-based reasoning – for example, extracting preserved molecular fragments to guide decoding,22 refining precursor structures through targeted string editing,23 or assembling molecules around conserved cores.24 Moreover, substructure-based string representations have recently emerged20,25,26 to enhance the expressiveness and interpretability of molecular notations, by capturing chemically meaningful fragments and their connectivity. FragSMILES was recently developed for de novo molecule design,26 to overcome limitations of existing string representations in capturing substructure information, by denoting the fragments independently of the connector atoms, as well as capturing chirality.27–29 The fragSMILES algorithm (Fig. 1a) operates by (1) disassembling molecules via predefined cleavage rules (exo-cyclic single bonds in this study), (2) collapsing the resulting fragments into the edges of a reduced graph, while keeping track of the atoms connecting the fragments, and (3) converting this graph into a string, whose elements (‘tokens’) represent nodes or edges.


image file: d5cc02641e-f1.tif
Fig. 1 FragSMILES notation for reaction prediction. (a) Molecules are converted into a reduced graph, obtained via exocyclic bond cleavage (Sup. Fig. 1). The resulting fragments (nodes) and their connecting bonds (edges) are then converted into elements (‘tokens’) that constitute the fragSMILES string. (b) Exemplary chemical reaction (from reagents to product), encoded as fragSMILES strings. Spaces were added to highlight token separation.

In this study, we apply fragSMILES for synthesis planning, under the hypothesis that its ability to encode substructures and advanced chirality can also enhance reaction prediction and retrosynthesis accuracy. We focused on two tasks: (1) forward reaction prediction, where the goal is to predict the products of a given set of reactants, and (2) retrosynthesis prediction, where the goal is to identify potential reactants and reagents needed to synthesize a target molecule. To this end, we used 1[thin space (1/6-em)]002[thin space (1/6-em)]602 curated chemical reactions from the USPTO database30 and represented them with different string notations. SMILES, SELFIES, SAFE, and t-SMILES were used as benchmarks. Other notable string representations exist (e.g., DeepSMILES,31 GroupSELFIES,25 and GenSMILES32), which were not considered due to their limited application to organic reaction prediction. SMILES, SELFIES, SAFE and t-SMILES were tokenized at the atom-level. FragSMILES were tokenized at the ‘chemical-word’ level, leading to remarkably more compact sequences26 (Sup. Table 1 and Sup. Fig. 3). This characteristic might help mitigate the memory usage associated with the increased complexity of word-level languages.33,34 We used the transformer architecture35 – the de facto standard for organic reaction planning36 – and framed the prediction task as a sequence-to-sequence translation (i.e., reactants to reagents, or the other way around) problem.13,14 Models were optimized and trained separately for each representation and task (Sup. Tables 2 and 3), and used to generate molecular strings via beam search37 (see SI). The transformer models were evaluated on 50[thin space (1/6-em)]234 reactions (unseen during model optimization or training) by measuring (Table 1) (a) validity, i.e., the number of ‘chemically-valid’ strings generated, including correct stereocenter assignations, and (b) accuracy, computed as the number of correct predictions over the total of considered predictions (from top-1 to top-5 sequences).

Table 1 Prediction accuracy of SMILES, SELFIES, SAFE, t-SMILES and fragSMILES, on the total set of reactions considered and on a subset of reactions involving stereocenters. Results are reported for both reaction prediction and for retrosynthesis prediction, in terms of validity (i.e., number of ‘chemically valid’ strings generated) and of top-k accuracy (50[thin space (1/6-em)]234 in total, and 8588 when considering reactions involving stereocenters). Metrics are analysed for the top-k generations (from 1 to 5) of beam search. Best (bold) and the second best (underline) metrics are highlighted
Task Metric Notation Top-1 Top-2 Top-3 Top-4 Top-5
a Computed by considering both syntactic validity (Sup. Table 4) and correct chirality annotation.
Forward synthesis Validitya SMILES 48[thin space (1/6-em)]366 (96.3%) 49[thin space (1/6-em)]470 (98.5%) 49[thin space (1/6-em)]798 (99.1%) [4 with combining low line][9 with combining low line][thin space (1/6-em)][9 with combining low line][2 with combining low line][7 with combining low line] ([9 with combining low line][9 with combining low line][. with combining low line][4 with combining low line]%) [5 with combining low line][0 with combining low line][thin space (1/6-em)][0 with combining low line][0 with combining low line]5 ([9 with combining low line][9 with combining low line][. with combining low line][5 with combining low line]%)
SELFIES 48[thin space (1/6-em)]418 (96.4%) 48[thin space (1/6-em)]857 (97.3%) 49[thin space (1/6-em)]075 (97.7%) 49[thin space (1/6-em)]213 (98.0%) 49[thin space (1/6-em)]325 (98.2%)
SAFE 46[thin space (1/6-em)]619 (92.8%) 48[thin space (1/6-em)]020 (95.6%) 48[thin space (1/6-em)]544 (96.6%) 48[thin space (1/6-em)]824 (97.2%) 49[thin space (1/6-em)]008 (97.6%)
t-SMILES 50[thin space (1/6-em)]231 (100.0%) 50[thin space (1/6-em)]234 (100.0%) 50[thin space (1/6-em)]234 (100.0%) 50[thin space (1/6-em)]234 (100.0%) 50[thin space (1/6-em)]234 (100.0%)
fragSMILES [4 with combining low line][8 with combining low line][thin space (1/6-em)][8 with combining low line][7 with combining low line]9 ([9 with combining low line][7 with combining low line][. with combining low line][3 with combining low line]%) [4 with combining low line][9 with combining low line][thin space (1/6-em)][5 with combining low line][5 with combining low line]3 ([9 with combining low line][8 with combining low line][. with combining low line][6 with combining low line]%) [4 with combining low line][9 with combining low line][thin space (1/6-em)][8 with combining low line][1 with combining low line]2 ([9 with combining low line][9 with combining low line][. with combining low line][2 with combining low line]%) 49[thin space (1/6-em)]918 (99.4%) 49[thin space (1/6-em)]989 (99.5%)
Accuracy SMILES [2 with combining low line][5 with combining low line][thin space (1/6-em)][0 with combining low line][5 with combining low line]3 ([4 with combining low line][9 with combining low line][. with combining low line][9 with combining low line]%) [2 with combining low line][9 with combining low line][thin space (1/6-em)][2 with combining low line][6 with combining low line][1 with combining low line] ([5 with combining low line][8 with combining low line][. with combining low line][2 with combining low line]%) [3 with combining low line][1 with combining low line][thin space (1/6-em)][1 with combining low line][3 with combining low line][3 with combining low line] ([6 with combining low line][2 with combining low line][. with combining low line][0 with combining low line]%) [3 with combining low line][2 with combining low line][thin space (1/6-em)][3 with combining low line][0 with combining low line][5 with combining low line] ([6 with combining low line][4 with combining low line][. with combining low line][3 with combining low line]%) [3 with combining low line][2 with combining low line][thin space (1/6-em)][9 with combining low line][8 with combining low line][8 with combining low line] ([6 with combining low line][5 with combining low line][. with combining low line][7 with combining low line]%)
SELFIES 10[thin space (1/6-em)]538 (21.0%) 13[thin space (1/6-em)]415 (26.7%) 14[thin space (1/6-em)]911 (29.7%) 15[thin space (1/6-em)]904 (31.7%) 16[thin space (1/6-em)]591 (33.0%)
SAFE 15[thin space (1/6-em)]151 (30.2%) 18[thin space (1/6-em)]758 (37.3%) 20[thin space (1/6-em)]557 (40.9%) 21[thin space (1/6-em)]609 (43.0%) 22[thin space (1/6-em)]169 (44.1%)
t-SMILES 3087 (6.1%) 4358 (8.7%) 5125 (10.2%) 5611 (11.2%) 6013 (12.0%)
fragSMILES 26[thin space (1/6-em)]826 (53.4%) 30[thin space (1/6-em)]287 (60.3%) 32[thin space (1/6-em)]026 (63.8%) 33[thin space (1/6-em)]015 (65.7%) 33[thin space (1/6-em)]692 (67.1%)
Retro-synthesis Validitya SMILES 20[thin space (1/6-em)]924 (41.7%) 28[thin space (1/6-em)]481 (56.7%) 33[thin space (1/6-em)]894 (67.5%) 37[thin space (1/6-em)]763 (75.2%) 40[thin space (1/6-em)]743 (81.1%)
SELFIES 40[thin space (1/6-em)]042 (79.7%) 45[thin space (1/6-em)]139 (89.9%) 47[thin space (1/6-em)]366 (94.3%) 48[thin space (1/6-em)]397 (96.3%) 48[thin space (1/6-em)]968 (97.5%)
SAFE 21[thin space (1/6-em)]890 (43.6%) 28[thin space (1/6-em)]193 (56.1%) 32[thin space (1/6-em)]939 (65.6%) 36[thin space (1/6-em)]289 (72.2%) 39[thin space (1/6-em)]018 (77.7%)
t-SMILES [3 with combining low line][6 with combining low line][thin space (1/6-em)][8 with combining low line][0 with combining low line][5 with combining low line] ([7 with combining low line][3 with combining low line][. with combining low line][3 with combining low line]%) [4 with combining low line][1 with combining low line][thin space (1/6-em)][1 with combining low line][8 with combining low line][8 with combining low line] ([8 with combining low line][2 with combining low line][. with combining low line][0 with combining low line]%) [4 with combining low line][4 with combining low line][thin space (1/6-em)][2 with combining low line][2 with combining low line][8 with combining low line] ([8 with combining low line][8 with combining low line][. with combining low line][0 with combining low line]%) [4 with combining low line][5 with combining low line][thin space (1/6-em)][9 with combining low line][3 with combining low line][2 with combining low line] ([9 with combining low line][1 with combining low line][. with combining low line][4 with combining low line]%) [4 with combining low line][7 with combining low line][thin space (1/6-em)][0 with combining low line][4 with combining low line][7 with combining low line] ([9 with combining low line][3 with combining low line][. with combining low line][7 with combining low line]%)
fragSMILES 28[thin space (1/6-em)]054 (55.8%) 35[thin space (1/6-em)]323 (70.3%) 39[thin space (1/6-em)]682 (79.0%) 42[thin space (1/6-em)]443 (84.5%) 44[thin space (1/6-em)]369 (88.3%)
Accuracy SMILES [4 with combining low line][0 with combining low line][3 with combining low line][1 with combining low line] ([8 with combining low line][. with combining low line][0 with combining low line]%) [5 with combining low line][6 with combining low line][0 with combining low line][2 with combining low line] ([1 with combining low line][1 with combining low line][. with combining low line][2 with combining low line]%) [6 with combining low line][7 with combining low line][0 with combining low line][9 with combining low line] ([1 with combining low line][3 with combining low line][. with combining low line][4 with combining low line]%) [7 with combining low line][5 with combining low line][9 with combining low line][0 with combining low line] ([1 with combining low line][5 with combining low line][. with combining low line][1 with combining low line]%) [8 with combining low line][3 with combining low line][0 with combining low line][2 with combining low line] ([1 with combining low line][6 with combining low line][. with combining low line][5 with combining low line]%)
SELFIES 8 (0.0%) 19 (0.0%) 29 (0.1%) 36 (0.1%) 49 (0.1%)
SAFE 3731 (7.4%) 4886 (9.7%) 5674 (11.3%) 6392 (12.7%) 6978 (13.9%)
t-SMILES 0 (0.0%) 0 (0.0%) 0 (0.0%) 0 (0.0%) 0 (0.0%)
fragSMILES 4230 (8.4%) 6129 (12.2%) 7588 (15.1%) 8905 (17.7%) 10[thin space (1/6-em)]091 (20.1%)
Forward synthesis (chiral) Validitya SMILES 8088 (94.2%) 8298 (96.6%) 8404 (97.9%) 8444 (98.3%) 8480 (98.7%)
SELFIES 6847 (79.7%) 7267 (84.6%) 7478 (87.1%) 7609 (88.6%) 7712 (89.8%)
SAFE 7814 (91.0%) 8026 (93.5%) 8099 (94.3%) 8142 (94.8%) 8182 (95.3%)
t-SMILES 8587 (100.0%) 8588 (100.0%) 8588 (100.0%) 8588 (100.0%) 8588 (100.0%)
fragSMILES [8 with combining low line][2 with combining low line][3 with combining low line][9 with combining low line] ([9 with combining low line][5 with combining low line][. with combining low line][9 with combining low line]%) [8 with combining low line][3 with combining low line][8 with combining low line][4 with combining low line] ([9 with combining low line][7 with combining low line][. with combining low line][6 with combining low line]%) [8 with combining low line][4 with combining low line][4 with combining low line][9 with combining low line] ([9 with combining low line][8 with combining low line][. with combining low line][4 with combining low line]%) [8 with combining low line][4 with combining low line][8 with combining low line][0 with combining low line] ([9 with combining low line][8 with combining low line][. with combining low line][7 with combining low line]%) [8 with combining low line][4 with combining low line][9 with combining low line][8 with combining low line] ([9 with combining low line][9 with combining low line][. with combining low line][0 with combining low line]%)
Accuracy SMILES [3 with combining low line][3 with combining low line][3 with combining low line][1 with combining low line] ([3 with combining low line][8 with combining low line][. with combining low line][8 with combining low line]%) [4 with combining low line][1 with combining low line][4 with combining low line][4 with combining low line] ([4 with combining low line][8 with combining low line][. with combining low line][3 with combining low line]%) [4 with combining low line][4 with combining low line][7 with combining low line][6 with combining low line] ([5 with combining low line][2 with combining low line][. with combining low line][1 with combining low line]%) [4 with combining low line][6 with combining low line][7 with combining low line][8 with combining low line] ([5 with combining low line][4 with combining low line][. with combining low line][5 with combining low line]%) [4 with combining low line][8 with combining low line][0 with combining low line][9 with combining low line] ([5 with combining low line][6 with combining low line][. with combining low line][0 with combining low line]%)
SELFIES 1170 (13.6%) 1548 (18.0%) 1732 (20.2%) 1859 (21.6%) 1956 (22.8%)
SAFE 1609 (18.7%) 2095 (24.4%) 2343 (27.3%) 2495 (29.1%) 2575 (30.0%)
t-SMILES 80 (0.9%) 126 (1.5%) 162 (1.9%) 177 (2.1%) 193 (2.2%)
fragSMILES 3801 (44.3%) 4345 (50.6%) 4652 (54.2%) 4825 (56.2%) 4957 (57.7%)
Retro-synthesis (chiral) Validitya SMILES 3425 (39.9%) 4576 (53.3%) 5551 (64.6%) 6255 (72.8%) 6760 (78.7%)
SELFIES 6421 (74.8%) 7356 (85.7%) 7816 (91.0%) 8029 (93.5%) 8142 (94.8%)
SAFE 3823 (44.5%) 4793 (55.8%) 5563 (64.8%) 6082 (70.8%) 6524 (76.0%)
t-SMILES [6 with combining low line][1 with combining low line][6 with combining low line][7 with combining low line] ([7 with combining low line][1 with combining low line][. with combining low line][8 with combining low line]%) [6 with combining low line][8 with combining low line][1 with combining low line][7 with combining low line] ([7 with combining low line][9 with combining low line][. with combining low line][4 with combining low line]%) [7 with combining low line][3 with combining low line][1 with combining low line][6 with combining low line] ([8 with combining low line][5 with combining low line][. with combining low line][2 with combining low line]%) [7 with combining low line][5 with combining low line][9 with combining low line][7 with combining low line] ([8 with combining low line][8 with combining low line][. with combining low line][5 with combining low line]%) [7 with combining low line][8 with combining low line][0 with combining low line][1 with combining low line] ([9 with combining low line][0 with combining low line][. with combining low line][8 with combining low line]%)
fragSMILES 4485 (52.2%) 5678 (66.1%) 6452 (75.1%) 6958 (81.0%) 7318 (85.2%)
Accuracy SMILES 669 (7.8%) 933 (10.9%) [1 with combining low line][1 with combining low line][0 with combining low line][8 with combining low line] ([1 with combining low line][2 with combining low line][. with combining low line][9 with combining low line]%) [1 with combining low line][2 with combining low line][4 with combining low line][9 with combining low line] ([1 with combining low line][4 with combining low line][. with combining low line][5 with combining low line]%) [1 with combining low line][3 with combining low line][4 with combining low line][3 with combining low line] ([1 with combining low line][5 with combining low line][. with combining low line][6 with combining low line]%)
SELFIES 8 (0.1%) 19 (0.2%) 27 (0.3%) 32 (0.4%) 43 (0.5%)
SAFE [6 with combining low line][3 with combining low line][5 with combining low line] ([7 with combining low line][. with combining low line][4 with combining low line]%) 805 (9.4%) 924 (10.8%) 1048 (12.2%) 1125 (13.1%)
t-SMILES 0 (0.0%) 0 (0.0%) 0 (0.0%) 0 (0.0%) 0 (0.0%)
fragSMILES 620 (7.2%) [9 with combining low line][1 with combining low line][9 with combining low line] ([1 with combining low line][0 with combining low line][. with combining low line][7 with combining low line]%) 1128 (13.1%) 1297 (15.1%) 1469 (17.1%)

t-SMILES consistently achieved 100% validity on forward synthesis prediction with fragSMILES achieving the second highest validity in the top-three generated candidates (Table 1). On retrosynthesis prediction, SELFIES achieved the highest validity (74.8%), with t-SMILES consistently achieving the second highest validity (73.3%). In terms of accuracy, fragSMILES always yielded the highest accuracy in both forward- and retro-synthesis prediction, with at least 204 to 1784 more correct predictions in the top-1. SMILES strings resulted in the second-best performance.

When analysing the substructure similarity between wrong predictions and the correct outcome (forward synthesis, Tanimoto coefficient on extended connectivity fingerprints38), all models exhibited comparable trends, with SELFIES and t-SMILES consistently showing lower similarity values on average (Sup. Fig. 4). Additionally, only limited overlap of correct predictions was observed among models using different notations (Sup. Fig. 5), suggesting that each representation captures distinct features of the underlying chemistry. The highest overlaps were found between SMILES and fragSMILES, ranging from 66% in top-1 to 78% in top-5 predictions, indicating some redundancy but also a degree of complementarity across models.

Moreover, we analysed the accuracy of fragSMILES on chemical reactions involving at least one stereocenter from the reactants or chemical product (8588 chemical reactions) as annotated in the original dataset (Table 1). For forward synthesis prediction, fragSMILES outperformed all tested methods, especially visible in the top-1 predictions, with differences in accuracy up to +5%. For retrosynthesis prediction, SMILES slightly outperformed fragSMILES in top-1 accuracy (+0.6%). The validity of SELFIES-generated molecules decreases when focusing on chiral compounds, highlighting the challenge of correctly capturing stereochemistry. The accuracy gap between SAFE and SELFIES further supports this observation. The overlap of accurate predictions between models is reported in Sup. Fig. 6. Neither sequence length, sampling probability nor token frequency could alone explain the general accuracy gains of fragSMILES. We analysed different subsets of reactions involving stereocenters to assess the predictive accuracy of fragSMILES. Across most subsets, fragSMILES was the top-performing representation (Sup. Table 5). The exception was stereoselective reactions, where fragSMILES ranked second (Sup. Table 5).

Finally, we examined the causes of invalid syntax (Fig. 2a)39 in forward reaction prediction. SELFIES primarily fails due to incorrect chirality assignments, while the fragment-level tokenization of fragSMILES eliminates syntax errors in cyclic structures (assigned to a single token). However, fragSMILES exhibits issues in bond assignment between fragments, as connector tokens dominate its sequences. Due to its atom-based tokenization, the SMILES language is more prone to errors involving ring closures and branches. In terms of inaccurate predictions (Fig. 2b), fragSMILES outperforms the other notations in correctly predicting cyclic substructures and scaffolds, whereas SMILES has an edge in generating acyclic substructures, reflecting the strengths of each respective representation.


image file: d5cc02641e-f2.tif
Fig. 2 Top-1 predictions (forward synthesis) per representation. (a) Syntax errors of invalid predictions; (b) correctly generated substructures among incorrectly predicted products (grouped by substructure).

This study demonstrates that the fragSMILES language represents an advancement in synthesis planning using deep learning, offering enhanced accuracy and validity over traditional string-based representations like SMILES and SELFIES. By leveraging substructure-based tokenization, fragSMILES captures the complexity of molecular stereocenters and cyclic structures, addressing key limitations in current methods. Its performance, especially in top-1 predictions, underscores its potential for enhancing reaction design and retrosynthetic planning, and becoming one of the de facto representations in the field. As AI-driven synthesis tools become more integrated into real-world applications, the ability to predict molecular transformations with high precision is critical, and fragSMILES can contribute to this evolution.

The USPTO dataset, while widely used as a benchmark, has known limitations.40,41 Incorporating more rigorous data curation, especially when dealing with stereochemistry, will further benefit the field. Future work integrating fragSMILES with more advanced machine learning techniques (e.g., large language models42) or in combination with complementary molecular representations (e.g., molecular graphs), might further push the boundaries of chemical automation.

Author contributions: Conceptualization: FM and FG. Data curation: FM and FC. Formal analysis: FM, FG, ON. Investigation: all authors. Methodology: FM and FG. Software: FM. Visualization: FM and FG. Writing – original draft: FM and FG. Writing – review and editing: all authors.

This research was co-funded by the European Union (ERC, ReMINDER, 101077879 to FG). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council.

Conflicts of interest

There are no conflicts to declare.

Data availability

Supplementary information: Materials and methods, Sup. Fig. 1–6, and Sup. Tables 1–5. See DOI: https://doi.org/10.1039/d5cc02641e

All the code and data useful to reproduce the results of this study are available on GitHub at the following URL: https://github.com/molML/fragSMILES4reaction.

References

  1. O. Engkvist, P.-O. Norrby, N. Selmi, Y.-H. Lam, Z. Peng, E. C. Sherer, W. Amberg, T. Erhard and L. A. Smyth, Drug Discovery Today, 2018, 23, 1203–1218 CrossRef CAS PubMed.
  2. A. F. De Almeida, R. Moreira and T. Rodrigues, Nat. Rev. Chem., 2019, 3, 589–604 CrossRef.
  3. D. van Tilborg, H. Brinkmann, E. Criscuolo, L. Rossen, R. Özçelik and F. Grisoni, Curr. Opin. Struct. Biol., 2024, 86, 102818 CrossRef CAS PubMed.
  4. C. W. Coley, D. A. Thomas, J. A. M. Lummiss, J. N. Jaworski, C. P. Breen, V. Schultz, T. Hart, J. S. Fishman, L. Rogers, H. Gao, R. W. Hicklin, P. P. Plehiers, J. Byington, J. S. Piotti, W. H. Green, A. J. Hart, T. F. Jamison and K. F. Jensen, Science, 2019, 365, eaax1566 CrossRef CAS PubMed.
  5. Y. Jiang, D. Salley, A. Sharma, G. Keenan, M. Mullin and L. Cronin, Sci. Adv., 2022, 8, eabo2626 CrossRef CAS PubMed.
  6. S. Johansson, A. Thakkar, T. Kogej, E. Bjerrum, S. Genheden, T. Bastys, C. Kannas, A. Schliep, H. Chen and O. Engkvist, Artif. Intell., 2019, 32–33, 65–72 Search PubMed.
  7. M. H. Segler, M. Preuss and M. P. Waller, Nature, 2018, 555, 604–610 CrossRef CAS PubMed.
  8. C. W. Coley, R. Barzilay, T. S. Jaakkola, W. H. Green and K. F. Jensen, ACS Cent. Sci., 2017, 3, 434–443 CrossRef CAS PubMed.
  9. P. Schwaller, T. Gaudin, D. Lányi, C. Bekas and T. Laino, Chem. Sci., 2018, 9, 6091–6098 RSC.
  10. H. Öztürk, A. Özgür, P. Schwaller, T. Laino and E. Ozkirimli, Drug Discovery Today, 2020, 25, 689–705 CrossRef PubMed.
  11. N. Patwardhan, S. Marrone and C. Sansone, Information, 2023, 14, 242 CrossRef.
  12. D. Alberga, N. Gambacorta, D. Trisciuzzi, F. Ciriaco, N. Amoroso and O. Nicolotti, J. Chem. Inf. Model., 2020, 60, 4582–4593 CrossRef PubMed.
  13. J. Nam and J. Kim, Linking the Neural Machine Translation and the Prediction of Organic Chemistry Reactions, arXiv, 2016, preprint, arXiv:1612.09529 DOI:10.48550/arXiv.1612.09529.
  14. P. Schwaller, T. Laino, T. Gaudin, P. Bolgar, C. A. Hunter, C. Bekas and A. A. Lee, ACS Cent. Sci., 2019, 5, 1572–1583 CrossRef CAS PubMed.
  15. H. Taniwaki and H. Kaneko, Macromol. Theory Simul., 2023, 32, 2300011 CrossRef CAS.
  16. F. Jaume-Santero, A. Bornet, A. Valery, N. Naderi, D. Vicente Alvarez, D. Proios, A. Yazdani, C. Bournez, T. Fessard and D. Teodoro, J. Chem. Inf. Model., 2023, 63, 1914–1924 CrossRef CAS PubMed.
  17. I. V. Tetko, P. Karpov, R. Van Deursen and G. Godin, Nat. Commun., 2020, 11, 5575 CrossRef CAS PubMed.
  18. D. Weininger, J. Chem. Inf. Comput. Sci., 1988, 28, 31–36 CrossRef CAS.
  19. M. Krenn, F. Häse, A. Nigam, P. Friederich and A. Aspuru-Guzik, Mach. Learn.: Sci. Technol., 2020, 1, 045024 Search PubMed.
  20. E. Noutahi, C. Gabellini, M. Craig, J. S. Lim and P. Tossou, Digital Discovery, 2024, 3, 796–804 RSC.
  21. J.-N. Wu, T. Wang, Y. Chen, L.-J. Tang, H.-L. Wu and R.-Q. Yu, Nat. Commun., 2024, 15, 4993 CrossRef CAS PubMed.
  22. L. Fang, J. Li, M. Zhao, L. Tan and J.-G. Lou, Nat. Commun., 2023, 14, 2446 CrossRef CAS PubMed.
  23. Y. Han, X. Xu, C.-Y. Hsieh, K. Ding, H. Xu, R. Xu, T. Hou, Q. Zhang and H. Chen, Nat. Commun., 2024, 15, 6404 CrossRef CAS PubMed.
  24. Y. Wang, C. Pang, Y. Wang, J. Jin, J. Zhang, X. Zeng, R. Su, Q. Zou and L. Wei, Nat. Commun., 2023, 14, 6155 CrossRef CAS PubMed.
  25. A. H. Cheng, A. Cai, S. Miret, G. Malkomes, M. Phielipp and A. Aspuru-Guzik, Digital Discovery, 2023, 2, 748–758 RSC.
  26. F. Mastrolorito, F. Ciriaco, M. V. Togo, N. Gambacorta, D. Trisciuzzi, C. D. Altomare, N. Amoroso, F. Grisoni and O. Nicolotti, Commun. Chem., 2025, 8, 26 CrossRef CAS PubMed.
  27. Y. Yoshikai, T. Mizuno, S. Nemoto and H. Kusuhara, Nat. Commun., 2024, 15, 1197 CrossRef CAS PubMed.
  28. G. Tom, E. Yu, N. Yoshikawa, K. Jorner and A. Aspuru-Guzik, Stereochemistry-aware string-based molecular generation, 2024, https://chemrxiv.org/engage/chemrxiv/article-details/6757d4eef9980725cf93c698 Search PubMed.
  29. N. Senkuttuvan, B. Komarasamy, R. Krishnamoorthy, S. Sarkar, S. Dhanasekaran and P. Anaikutti, RSC Adv., 2024, 14, 33429–33448 RSC.
  30. D. Lowe, Chemical reactions from US patents (1976–Sep 2016), 2017, https://figshare.com/articles/dataset/Chemical_reactions_from_US_patents_1976-Sep2016_/5104873 Search PubMed.
  31. N. O’Boyle and A. Dalke, An Adaptation of SMILES for Use in Machine-Learning of Chemical Structures, ChemRxiv, 2018 DOI:10.26434/chemrxiv.7097960.v1.
  32. A. S. Bhadwal, K. Kumar and N. Kumar, Knowl.-Based Syst., 2023, 268, 110429 CrossRef.
  33. C. Toraman, E. H. Yilmaz, F. Şahinuç and O. Ozcelik, ACM Trans. Asian Low-Resour. Lang. Inf. Process., 2023, 22, 1–21 CrossRef.
  34. A. Rai and S. Borah, Applications of Internet of Things: Proceedings of ICCCIOT 2020, 2021, pp. 193–200.
  35. A. Vaswani, Attention Is All You Need, arXiv, 2017, preprint, arXiv:1706.03762 DOI:10.48550/arXiv.1706.03762.
  36. M. E. Mswahili and Y.-S. Jeong, Heliyon, 2024, 10 DOI:10.1016/j.heliyon.2024.e39038.
  37. P. Koehn, Proceedings of the 2004 conference on empirical methods in natural language processing, 2004, pp. 388–395.
  38. D. Rogers and M. Hahn, J. Chem. Inf. Model., 2010, 50, 742–754 CrossRef CAS PubMed.
  39. R. Özçelik, S. de Ruiter, E. Criscuolo and F. Grisoni, Nat. Commun., 2024, 15, 6176 CrossRef PubMed.
  40. T. R. Gimadiev, A. Lin, V. A. Afonina, D. Batyrshin, R. I. Nugmanov, T. Akhmetshin, P. Sidorov, N. Duybankova, J. Verhoeven and J. Wegner, et al., Mol. Inf., 2021, 40, 2100119 CrossRef CAS PubMed.
  41. S. Szymkuć, T. Badowski and B. A. Grzybowski, Angew. Chem., 2021, 133, 26430–26436 CrossRef.
  42. K. M. Jablonka, P. Schwaller, A. Ortega-Guerrero and B. Smit, Nat. Mach. Intell., 2024, 6, 161–169 CrossRef.
This journal is © The Royal Society of Chemistry 2025
Click here to see how this site uses Cookies. View our privacy policy here.