Data Understanding

In the Business Understanding section, we learned that our goal is to predict molecular properties from SMILES strings. The three properties we want to predict are HOMO, LUMO, and Stability — these are the outputs of our machine learning models. The inputs are the SMILES strings, which describe molecules in text form.

From SMILES to Predictions

However, SMILES strings cannot be used directly as input to a machine learning model. They need to be transformed into numerical features first. To do this, we use RDKit, a Python library that converts SMILES strings into molecular descriptors.

The process starts with a simple idea:

flowchart LR
    A["SMILES String"] --> B["Machine Learning Model"]
    B --> C["HOMO\nLUMO\nStability"]
    style A fill:#f47a20,color:#fff,stroke:none
    style B fill:#00558c,color:#fff,stroke:none
    style C fill:#6c757d,color:#fff,stroke:none

But to make this work, SMILES strings need an intermediate step — conversion to numerical descriptors using RDKit:

flowchart LR
    A["SMILES String"] --> B["RDKit"]
    B --> C["Molecular\nDescriptors"]
    C --> D["Machine Learning\nModel"]
    D --> E["HOMO\nLUMO\nStability"]
    style A fill:#f47a20,color:#fff,stroke:none
    style B fill:#adb5bd,color:#000,stroke:none
    style C fill:#f47a20,color:#fff,stroke:none
    style D fill:#00558c,color:#fff,stroke:none
    style E fill:#6c757d,color:#fff,stroke:none

Descriptors vs. Properties

It is important to understand the difference between descriptors (input) and properties (output):

A molecular descriptor is an input: a numerical or symbolic feature that encodes information about a molecule’s structure.
A molecular property is an output: a measurable characteristic that describes how the molecule behaves.

Think of it like cars. Engine size, weight, and tire width are descriptors. Top speed and fuel efficiency are properties. In machine learning terms, descriptors are the features we feed into the model, while properties are the labels we try to predict.

In an ideal world, a dataset for this task would look like this:

SMILES	Descriptor 1	Descriptor 2	…	HOMO	LUMO	Stability
…	…	…	…	…	…	…

This would be perfect, but data science rarely starts this way. Instead, we often need to work with multiple raw datasets from different sources and shape them into the format we need by cleaning, transforming, and combining them.

Datasets

In this course, you will work with datasets from multiple sources:

Name	Molecules	Files	Contains
Starter	80	`starter_dataset.xlsx`	SMILES, Molecular Descriptors, Targets
HOPV	350	`HOPV_homolumo.data`, `HOPV_stability.csv`	SMILES, Targets
Lab database	232	`lab_measurements.db`	SMILES, Targets
Kaggle	200	`classification_sample_submission.csv`, `regression_sample_submission.csv`	SMILES

The starter dataset is a subset of the HOPV dataset. You use the starter dataset in Iteration 1, HOPV from Iteration 2, and the lab database and the Kaggle molecules in Iteration 3+.

Important

To ensure a fair competition, the target values (HOMO, LUMO, and stability) for all datasets, including the Kaggle set, have been synthetically generated based on the real molecular properties. While they follow realistic chemical distributions, they will not match publicly available data for these specific molecules. Your goal is to model the relationships within the datasets we provide to you.

The Starter Dataset

This dataset is modelling-ready and is prepared to help you begin quickly. It contains 80 molecules randomly selected from the 350 molecules in the HOPV dataset.

Column A: The SMILES string for each molecule.
Columns B–HJ: Molecular descriptors computed using RDKit.
Columns HK–HM: The target variables — HOMO, LUMO, and Stability.

This dataset will only be used in Iteration 1. From Iteration 2 onward, you are expected to process the raw data yourself and bring it into the required format.

The Harvard Organic Photovoltaics (HOPV) Dataset

This dataset was created by researchers at Harvard University’s Department of Chemistry and Chemical Biology. It contains SMILES strings as well as HOMO and LUMO values for 350 molecules.

The file HOPV_homolumo.data does not contain stability information. This is provided as a separate file, HOPV_stability.csv.

The Kaggle Dataset

The HOPV molecules, and from Iteration 3+ the lab database, form the development set for your classification and regression models. Once your models are ready, you will submit predictions for 200 SMILES molecules on Kaggle. The true values for these molecules are hidden, but after submission, you will see your results on the platform.

Two files are provided — one for the classification competition and one for the regression competition. These files contain 200 SMILES molecules each, with placeholder values in the target columns. Your task is to replace these placeholders with your model predictions and submit the files in the same format.