Synergy Effect between Convolutional Neural Networks and the Multiplicity of SMILES for Improvement of Molecular Prediction

• Molecular structures are complex, and machine learning requires numerical representations; fingerprints encode molecular information and can capture substructures.The motivation for substructure extraction rests on the belief that molecules with similar substructures have similar properties.Source: §1 · p. 1
• Cycle, branch, and stereochemistry breaks in linearized SMILES can create ambiguities, especially when training neural networks.Source: §1 · p. 2
• Several valid SMILES can represent one molecule, and canonicalization is toolkit-dependent; similar molecules need not have similar canonical SMILES.Source: §1 · p. 2
• Multiple SMILES versions are proposed to expose different views of a molecular graph and support a more global understanding of its structure.Source: §1 · p. 2
• The proposed approach combines CNN feature extraction with multiple SMILES as data augmentation for modeling chemical structures.The paper describes its CNN architecture as combining linear and hierarchical layers.Source: §1 · p. 2

• SMILES-based RNNs were already used for molecular generation and chemical-property prediction, alongside other machine-learning applications to SMILES.The passage cites Segler et al. for generation and Goh et al. [10] for property prediction.Source: §2 · p. 2
• Prior work proposed increasing dataset size by considering several SMILES for one molecule as data augmentation.The proposal is attributed to Bjerrum.Source: §2 · p. 2
• Neural fingerprints and ConvGraph apply convolutions directly to 2D molecular graphs, while string-vector CNNs have also been used for sentence classification.Source: §2 · p. 2
• The paper combines CNNs with SMILES multiplicity, building on CNN sequence processing and data augmentation with multiple SMILES per molecule.Source: §2 · p. 2

• The implementation applies convolutions to SMILES rather than directly to molecular graphs, which the passage describes as computationally inefficient for the graph-convolution approach.SMILES are single-line characters encoding molecular-graph information, and string-vector convolutions have been used in sentence classification.Source: §3 · p. 3
• CNF applies convolutional layers to one-hot SMILES and follows them with matrix multiplication using an embedding matrix, linking neighborhood features to hashed features.The passage relates this operation to the Morgan fingerprint of Rogers and Hahn.Source: §3 · p. 3
• The model uses flat and hierarchical convolutional layers, resembling the ResNet architecture.Source: §3 · p. 3
• Pooling sums the rows of each convolution-and-multiplication result, and concatenating the pooled vectors produces the neural fingerprint.Source: §3 · p. 3
• The neural fingerprint maps sparse one-hot SMILES into a dense vector that can feed standard machine-learning algorithms.The passage names neural networks, support vector machines, and random forests as examples.Source: §3 · p. 3

• SMILES augmentation is framed as increasing training observations, while data augmentation is described as a regularization technique that can prevent overfitting.Source: §4 · p. 3
• Augmentation generates multiple valid SMILES by varying the enumeration start atom and path through the same molecular graph.The four illustrated ethylcyclopropane strings are all valid representations of that molecule.Source: §4 · pp. 3, 5, Figure 2 · p. 5
• The CNF neural fingerprint concatenates vectors produced by pooling convolution-and-hashing layers.The illustrated flat-layer design sums columns after each layer before concatenating the resulting vectors.Source: §4 · p. 4
• Different SMILES spellings expose multiple views of a molecule, addressing how linearization can obscure cycles and branches from a neural network.Source: §4 · p. 5
• The study distinguishes training-time augmentation from inference-time augmentation and evaluates their combination, which earlier cited studies had not examined together.Source: §4 · p. 5
• At testing, outputs from multiple SMILES are aggregated as an ensemble: averaged for regression and majority-voted for classification.Source: §4 · p. 5
• SMILES augmentation during both training and testing improves predictive performance.Source: §4 · p. 5

• A new RDKit feature was developed to generate all possible valid SMILES, addressing restrictions in the previously implemented enumeration function.Source: §5 · p. 5
• Evaluation used 5-fold cross-validation with splits made before augmentation, spanning regression scored by RMSE and classification scored by AUC.Source: §5 · p. 6
• Training-only augmentation almost always improves over SMILES 1/1, while testing-only almost always worsens; Figure 3 reports improved classification accuracy for both-side augmentation.Source: §5 · p. 6, Figure 3 · p. 7
• The best results occur with augmentation during both training and testing, combining training-set data augmentation with test-set ensemble learning.Source: §5 · p. 6
• In most cases, CNF performs at least as well as the best DeepChem models.Source: §5 · p. 6
• In most cases, default-parameter CNF provides better accuracy than models built using the best descriptors and computational methods.Source: §5 · p. 7

6 Conclusion and Future Work

• The authors identify quantum-chemistry atomic targets, reaction tasks, and augmentation with TextCNN as future directions or ongoing work.Molecular QSAR/QSPR prediction has been investigated; TextCNN augmentation is under investigation.Source: §6 · pp. 7, 10
• For regression, enumeration during both training and testing (SMILES 10/10 or 10/25) improves CNF accuracy.The conclusion also states that augmentation improves predictions overall by approximately 15%, and by approximately 25% in the best cases.Source: §6 · p. 7, Figure 4 · p. 8
• Using multiple valid SMILES as data augmentation improves CNF predictions by approximately 15% (approximately 25% in the best cases), reaching conventional human-descriptor performance on difficult targets such as molecular melting point.Source: §6 · p. 7
• Using testing-time augmentation as an ensemble may provide an applicability-domain metric, but this approach has not been fully investigated.The mean and variance may provide information about model quality.Source: §6 · p. 10

7 Supplementary Material

• The analysis proposes that fingerprints from generated SMILES of the same initial molecule should be no farther apart than fingerprints from different molecules.The stated distance criterion compares Euclidean distances between neural fingerprints.Source: §7 · p. 13
• Training-and-testing augmentation (SMILES 10/10 or 10/25) is reported to improve model accuracy across regression and classification simulations.Table 1 summarizes target, initial dataset size, augmentation type, and the RMSE or AUC measure.Source: Table 1 · p. 14
• As the CNF model trains, neural fingerprints generated for the same molecule become closer than fingerprints generated for different molecules.The comparison uses distance matrices after one and ten epochs.Source: §7 · pp. 13, 15
• CNF usually performs better than the best Deepchem models in the reported regression and classification comparisons.Table 2 compares the best CNF results with Deepchem results reported by Wu et al. [44].Source: Table 2 · p. 16
• OCHEM comparisons report that CNF provides better accuracy than models using the best descriptors and computational methods across classification and regression tasks.Table 3 covers nine classification tasks using AUC; Table 4 covers nine regression tasks using RMSE.Source: Table 3 · p. 17, Table 4 · p. 18