Objectives

[t-kimber]

Objectives

maxsmi code:

• Provides confidence in the prediction using ensemble learning
• Provides augmentation strategies to be used e.g. in benchmarks
• Provides a guide to best practices in SMILES augmentation
• Provides a systematic way of generating SMILES augmentation

maxsmi research:

• Does augmentation improve performance independently of the model ?
• Does the performance (only/strictly) increase with augmentation or does it reach a plateau ?
• What is the best augmentation strategy ?
• Can the model learn the inherent symmetry of the molecules using controlled duplication ?
• (Show that in the string-encoding of molecules, SMILES augmentation works better than SMILES, SELFIES and DeepSMILES.)

Datasets

Size / Dataset	Data before processing	Data after processing	Train set (before augmenting) 80%	Test set (before augmenting) 20%
ESOL	1128	1128	902	226
ESOL_small	1128	1068	854	214
free solv	642	642	513	129
lipophilicity	4200	4199	3359	840
ChEMBL28	6026	5849	4679	1170

Results

To answer the questions from the objectives, #28 is created.

Note:

• Delaney corresponds to ESOL
• SAMPL corresponds to free solv

MoleculeNet results:

• http://moleculenet.ai/latest-results

[t-kimber]

MoleculeNet R2 results

[t-kimber]

MoleculeNet RMSE results

[t-kimber]

CNF RMSE results