t-SMILES: A Scalable Fragment-based Molecular Representation Framework

When using advanced NLP methodologies to solve chemical problems, two fundamental questions arise: 1) What are 'chemical words'? and 2) How can they be encoded as 'chemical sentences’?

This study introduces a scalable, fragment-based, multiscale molecular representation algorithm called t-SMILES (tree-based SMILES) to address the second question. It describes molecules using SMILES-type strings obtained by performing a breadth-first search on a full binary tree formed from a fragmented molecular graph.

For more details, please refer to the paper.

Systematic evaluations using JTVAE, BRICS, MMPA, and Scaffold show that:

1. It can build a multi-code molecular description system, where various descriptions complement each other, enhancing the overall performance. Under this framework, classical SMILES can be unified as a special case of t-SMILES to achieve better balanced performance using hybrid decomposition algorithms.

2. It exhibits impressive performance on low-resource datasets JNK3 and AID1706, whether the model is original, data augmented, or pre-training fine-tuned;

3. It significantly outperforms classical SMILES, DeepSMILES, SELFIES and baseline models in goal-directed tasks.

4. It outperforms previous fragment-based models being competitive with classical SMILES and graph-based methods on Zinc, QM9, and ChEMBL.

To support the t-SMILES algorithm, we introduce a new character, '&', to act as a tree node when the node is not a real fragment in FBT. Additionally, we introduce another new character, '^', to separate two adjacent substructure segments in t-SMILES string, similar to the blank space in English sentences that separates two words.

Three coding algorithms are presented in this study:

1. TSSA: t-SMILES with shared atom.

2. TSDY: t-SMILES with dummy atom but without ID.

3. TSID: t-SMILES with ID and dummy atom.

For example, the three t-SMILES codes of Celecoxib are:

TSID_M:

• [1*]C&[1*]C1=CC=C([2*])C=C1&[2*]C1=CC([3*])=NN1[5*]&[3*]C([4*])(F)F&[4*]F^[5*]C1=CC=C([6*])C=C1&&[6*]S(N)(=O)=O&&&

TSDY_M (replace [n*] with *):

• *C&*C1=CC=C(*)C=C1&*C1=CC(*)=NN1*&*C(*)(F)F&*F^*C1=CC=C(*)C=C1&&*S(N)(=O)=O&&&

TSSA_M:

• CC&C1=CC=CC=C1&CC&C1=C[NH]N=C1&CN&C1=CC=CC=C1^CC^CS&C&N[SH]=O&CF&&&&FCF&&

Here we provide the source code of our method.

Dependencies

We recommend Anaconda to manage the version of Python and installed packages.

Please make sure the following packages are installed:

1. Python(version >= 3.7)

2. PyTorch (version == 1.7)

3. RDKit (version >= 2020.03)

4. Networkx(version >= 2.4)

5. Numpy (version >= 1.19)

6. Pandas (version >= 1.2.2)

7. Matplotlib (version >= 2.0)

8. Scipy(version >= 1.4.1)

Usage

1. DataSet/Graph/CNJTMol.py

encode_single ()

It contained a preprocess function to generate t-SMILES from data set.

1. DataSet/Graph/CNJMolAssembler.py

decode_single()

It reconstructs molecules form t-SMILES to generate classical SMILES.

In this study, GPT and RNN generative models are used for evaluation.

Acknowledgement

We thank the following Git repositories that gave me a lot of inspirations:

1. MolGPT : https://github.com/devalab/molgpt

2. MGM: https://github.com/nyu-dl/dl4chem-mgm

3. JTVAE: https://github.com/wengon-jin/icml18-jtnn

4. hgraph2graph: https://github.com/wengong-jin/hgraph2graph

5. DeepSmiles: https://github.com/baoilleach/deepsmiles

6. SELFIES: https://github.com/aspuru-guzik-group/selfies

7. FragDGM: https://github.com/marcopodda/fragment-based-dgm

8. CReM: https://github.com/DrrDom/crem

9. AttentiveFP: https://github.com/OpenDrugAI/AttentiveFP

10. Guacamol: https://github.com/BenevolentAI/guacamol\_baselines

11. MOSES: https://github.com/molecularsets/moses

12. GPT2: https://github.com/samwisegamjeee/pytorch-transformers

13. Datamol: https://github.com/datamol-io/datamol