Predicting Melting Points from MACCS Keys

Import data, compute molecular fingerprints, obtain a melting point predictor function and evaluate performance

Import a melting point dataset from Kaggle:

In[1]:=

meltingdata=Import["https://www.kaggle.com/api/v1/datasets/download/aliffaagnur/melting-point-chemical-dataset/BradleyDoublePlusGoodMeltingPointDataset.csv","Tabular"]

Out[1]=

Tabular

Row count: 3041

Column count: 11



MACCS keys are used to compute molecular similarity for computational drug design and database matching.

Compute the MACCS keys for each of the molecules in the dataset (this can take several minutes):

In[2]:=

maccsData=ResourceFunction["MACCSKeys"]@meltingdata[All,"smiles"]//Quiet;Short[maccsData]

Out[2]//Short=

Arrange the data into a format convenient for listing features and ground truth:

In[3]:=

processedData=InsertColumns[meltingdata[All,{"name","smiles","mpC"}],Thread[Table["MACCS_"<>ToString[n],{n,166}]Transpose@Normal@maccsData]]

Out[3]=

Tabular

Row count: 3041

Column count: 169



Split the data into training and test sets:

In[4]:=

{train,test}=ResourceFunction["TrainTestSplit"][KeyDrop[processedData,{"name","smiles"}],"TrainingSetSize"Scaled[.9]]

Out[4]=



Tabular

Row count: 2737

Column count: 167



Tabular

Row count: 304

Column count: 167





Use gradient boosted trees to obtain a predictor function for melting point:

In[5]:=

pf=Predict[train"mpC",Method"GradientBoostedTrees"]

Out[5]=

PredictorFunction

Input type:

Mixed (number: 166)

Method: GradientBoostedTrees

Data not saved. Save now



Check overall performance of the predictor:

In[6]:=

results=PredictorMeasurements[pf,test"mpC"]

Out[6]=

Predictor Measurements

Predictor method	GradientBoostedTrees
Number of test examples	304
Standard deviation	53.7 ± 2.8
Standard deviation baseline	98.5 ± 3.7
R-squared	0.702 ± 0.038
Mean cross entropy	6.08 ± 0.19
Single evaluation time	39. ms/example
Batch evaluation speed	884. examples/s

MACCS keys alone are no longer state of the art features to extract for this task, but are often combined with other descriptors. With MACCS keys alone, 2025 methods (Y. Zushi) have achieved root-mean-square error (RMSE) of ~50 °C.

In[7]:=

RootMeanSquare@results["Residuals"]

Out[7]=

47.7562

Check

and the mean absolute error (MAE):

In[8]:=

results[{"RSquared","MeanDeviation"}]

Out[8]=

{0.702369,39.6046}

Source Metadata

Citation:
- Zushi, Y. Mining Structural Information from Gas Chromatography-Electron-Impact Ionization-Mass Spectrometry Data for Analytical-Descriptor-Based Quantitative Structure–Activity Relationship. J. Xenobiot. 2025, 15, 177. https://doi.org/10.3390/jox15060177

Publisher Information

Contributed by: John McNally

License Information

This work is licensed under a Creative Commons Attribution 4.0 International License

Wolfram Language Example Repository

Predicting Melting Points from MACCS Keys

Source Metadata

See Also

Publisher Information

License Information

Predicting Melting Points from MACCS Keys

Source Metadata

See Also

Related Symbols

Publisher Information

License Information