Skip to content

Fingerprint Models

Workbench supports training models on Morgan count fingerprints — a powerful molecular representation that captures substructure occurrence counts. Fingerprint models work with any framework (XGBoost, PyTorch) and can be combined with RDKit descriptors.

Pipeline Overview

Fingerprint models use a two-step pipeline:

  1. Generate fingerprints — A transformer endpoint computes Morgan count fingerprints from SMILES
  2. Train a model — Use the "fingerprint" column as a feature (stored as a compressed column)

Create the Fingerprint Endpoint

Note: This 'feature endpoint' only needs to be created once. It can then be reused across multiple models and FeatureSets. Please see our blog post Feature Endpoints: Reusable Data Transformations for more details.

from workbench.api import FeatureSet, ModelType
from workbench.utils.model_utils import get_custom_script_path

tags = ["smiles", "morgan fingerprints"]
script_path = get_custom_script_path("chem_info", "morgan_fingerprints.py")
feature_set = FeatureSet("aqsol_features")
model = feature_set.to_model(
    name="smiles-to-fingerprints-v0",
    model_type=ModelType.TRANSFORMER,
    feature_list=["smiles"],
    description="Smiles to Morgan Fingerprints",
    tags=tags,
    custom_script=script_path,
)

# Create the endpoint for the model
end = model.to_endpoint(tags=tags)
end.auto_inference()

Step 1: Compute Fingerprints and Create a FeatureSet

Run your data through the fingerprint endpoint, then create a FeatureSet and mark the fingerprint column as compressed:

from workbench.api import DataSource, Endpoint
from workbench.core.transforms.pandas_transforms import PandasToFeatures

ds = DataSource("aqsol_data")
df = ds.pull_dataframe()

# Run the data through our Smiles to Fingerprints Endpoint
fp_end = Endpoint("smiles-to-fingerprints-v0")
df_with_fp = fp_end.inference(df)

# Create a Feature Set
to_features = PandasToFeatures("aqsol_fingerprints")
to_features.set_input(df_with_fp, id_column="id")
to_features.set_output_tags(["aqsol", "fingerprints"])
to_features.transform()

# Set our compressed features for this FeatureSet
fs = FeatureSet("aqsol_fingerprints")
fs.set_compressed_features(["fingerprint"])

Step 2: Train a Model on Fingerprints

from workbench.api import FeatureSet, ModelType

fs = FeatureSet("aqsol_fingerprints")

model = fs.to_model(
    name="aqsol-fingerprint-reg-v0",
    model_type=ModelType.UQ_REGRESSOR,
    target_column="solubility",
    feature_list=["fingerprint"],
    description="Model for Aqueous Solubility using Morgan Fingerprints",
    tags=["aqsol", "fingerprints", "regression"],
)
end = model.to_endpoint(tags=["aqsol", "fingerprints", "regression"])
end.auto_inference()

Fingerprints + Descriptors

Combine fingerprints with RDKit molecular descriptors for richer feature sets:

# Grab features from an existing descriptor model and add fingerprints
descriptor_features = Model("sol-xgb-reg").features()
combined_features = descriptor_features + ["fingerprint"]

model = fs.to_model(
    name="sol-fingerprints-plus",
    model_type=ModelType.REGRESSOR,
    target_column="solubility",
    feature_list=combined_features,
    description="Fingerprints + Molecular Descriptors",
    tags=["fingerprints", "descriptors"],
)

Using PyTorch with Fingerprints

from workbench.api import FeatureSet, ModelType, ModelFramework

fs = FeatureSet("aqsol_fingerprints")

model = fs.to_model(
    name="sol-fp-pytorch",
    model_type=ModelType.UQ_REGRESSOR,
    model_framework=ModelFramework.PYTORCH,
    target_column="solubility",
    feature_list=["fingerprint"],
    description="PyTorch model on Morgan Fingerprints",
    tags=["fingerprints", "pytorch"],
)

How Count Fingerprints Work

Morgan count fingerprints (ECFP4 equivalent) encode substructure occurrence counts rather than binary presence/absence:

  • Radius 2 (ECFP4) — captures local chemical environments up to 2 bonds from each atom
  • 2048 bits — hashed into a fixed-length vector
  • Count values (0–255) — how many times each substructure occurs, providing richer information than binary fingerprints

Examples

Full code listing: examples/models/smiles_to_fingerprints.py

Questions?

The SuperCowPowers team is happy to answer any questions you may have about AWS® and Workbench.