SMILES formats, featurizer selection, model configuration, and training parameters for DeepChem.

You provide molecules as SMILES — entered directly, uploaded as CSV/TSV, or imported from a previous job. The choice that most affects your result is the featurizer paired with the model: graph featurizers suit GNN architectures, while fingerprint or descriptor featurizers suit classical and tree-based models.

Webapp input methods

The .bloom* webapp provides several input methods for DeepChem property prediction.

Input type selection

Input Type	Description	Example
SMILES	Direct SMILES string entry	`CC(=O)Oc1ccccc1C(=O)O`
CSV/TSV Upload	Tabular file with SMILES column	Upload `compounds.csv`
From Previous Job	Import from GenMol, PubChem, etc.	Select from job outputs

SMILES input requirements

Valid SMILES

Standard notation: Canonical or isomeric SMILES (e.g., c1ccccc1 for benzene)
Validation: Real-time RDKit parsing; invalid SMILES rejected before submission
Stereochemistry: Supported for chiral centers and E/Z bonds
Salts and mixtures: Use a single connected component; remove counter-ions with . separator

Batch input

Multi-line SMILES: One SMILES per line for direct entry
CSV format: Must contain a column named smiles or SMILES
TSV format: Tab-separated with SMILES column
Maximum batch size: 10,000 molecules per prediction job
Additional columns: Preserved in output for result matching

Featurizer selection

Featurizers convert SMILES strings into numerical representations for the neural network.

Featurizer	Type	Description	Best For
MolGraphConv	Graph	Atom/bond feature graph	GCN, GAT, MPNN models
ECFP	Fingerprint	Extended connectivity circular fingerprint (Morgan)	Classical ML, similarity search
MACCS	Fingerprint	166-bit MACCS structural keys	Quick substructure-based models
Mol2Vec	Embedding	Learned molecular embeddings	Transfer learning tasks
CoulombMatrix	3D	Coulomb matrix from 3D coordinates	3D-aware models
RDKitDescriptors	Descriptors	200+ physicochemical descriptors	XGBoost, Random Forest

Choosing a featurizer

Graph-based (MolGraphConv): Best for GNN models; captures full molecular topology
Fingerprint-based (ECFP, MACCS): Best for classical ML or when interpretability matters
Descriptor-based (RDKitDescriptors): Best for tree-based models with tabular features

Model parameters

Pre-trained model selection

Endpoint	Model	Training Data	Metric
Solubility (logS)	MPNN	AqSolDB	RMSE: 0.82
Lipophilicity (logP)	GCN	Lipophilicity dataset	RMSE: 0.55
hERG inhibition	AttentiveFP	ChEMBL hERG	AUC: 0.87
AMES mutagenicity	DMPNN	Hansen AMES	AUC: 0.85
BBB permeability	MPNN	BBBP dataset	AUC: 0.89
CYP inhibition	GCN	ToxCast CYP	AUC: 0.83

Architecture parameters

Parameter	Type	Default	Description
model	string	`mpnn`	Model architecture (mpnn, gcn, gat, attentivefp, dmpnn, chemberta)
endpoint	string	-	Property endpoint to predict
featurizer	string	`mol_graph`	Molecular featurizer
uncertainty	string	`none`	Uncertainty method (none, ensemble, mc_dropout)
num_ensemble	integer	5	Number of ensemble members (if uncertainty=ensemble)
dropout_samples	integer	20	MC-dropout forward passes (if uncertainty=mc_dropout)

Training parameters (custom models)

Parameter	Type	Default	Description
task_type	string	`regression`	Task type (regression or classification)
epochs	integer	100	Training epochs
learning_rate	float	0.001	Initial learning rate
batch_size	integer	64	Training batch size
hidden_dim	integer	128	Hidden layer dimension
num_layers	integer	3	Number of GNN layers
split_ratio	string	`0.8/0.1/0.1`	Train/validation/test split
early_stopping	boolean	true	Stop training when validation loss plateaus

Previous job integration

Compatible tools

GenMol: Generated molecules for property screening
PubChem: Compound SMILES from database lookups
Fragment library: Fragment property profiling
ChEMBL: Known bioactive compounds

Data handling

SMILES extraction: Automatic extraction from multi-format job outputs
Deduplication: Optional duplicate removal before prediction
Provenance tracking: Maintains connection to original data source

Custom training data format

For training custom models, provide a CSV file with:

smiles column: SMILES strings for each molecule
label column(s): Experimental property values (numeric for regression, 0/1 for classification)
Optional: Additional descriptor columns, compound identifiers
Missing values: Use NaN or empty cells for missing labels in multi-task datasets

Data quality requirements

Minimum 50 molecules for meaningful model training (500+ recommended)
Balanced classes for classification tasks (oversample minority class if imbalanced)
Remove duplicates and conflicting labels
Standardize SMILES to canonical form for consistency

DeepChem inputs and parameters