Digital Alchemy: Predicting Absorption Maxima from SMILES

Project Overview

This project focuses on predicting the absorption maxima (λₘₐₓ) of organic chromophores based solely on their molecular structure. Using a supervised machine learning approach, I developed a complete pipeline that takes SMILES representations as input and outputs predicted absorption wavelengths in nanometers. The model is trained on a large, experimentally measured dataset and built around a Random Forest regressor, which has proven effective for structured, high-dimensional input like molecular fingerprints.

The pipeline begins by preprocessing the dataset: cleaning missing values, canonicalizing SMILES, and generating 1024-bit Morgan fingerprints to numerically encode molecular structure. A stratified train/test split ensures balanced distribution of outliers and avoids data leakage. After training, the model’s performance is evaluated using standard regression metrics, including Mean Squared Error (MSE) and R² score. An additional outlier detection step is integrated to identify predictions with unusually high error, helping improve the model’s robustness and interpretation.

Overall, this project demonstrates how machine learning can offer a fast, low-cost alternative to experimental UV–Vis measurements. While inspired by open-source examples, all code and pipeline logic were developed independently and adapted to suit the specific dataset and goals of this work.

Data Processing Pipeline

1. Data Preprocessing (data_preprocessing.py)

• Loads and cleans raw dataset
• Canonicalizes SMILES using RDKit
• Converts spectral peaks into a consistent format
• Flags outliers based on IQR
• Outputs cleaned dataset ready for modeling

2. Model Training (ML_model_training.py)

• Generates 1024-bit Morgan fingerprints
• Splits data into train/test while preserving outlier distribution
• Trains a Random Forest regressor with tuned hyperparameters
• Evaluates performance and saves metrics

---

Model Details

• Model: Random Forest Regressor
• Features: Morgan fingerprints (radius=2, nBits=1024)
• Target: Absorption maximum (λₘₐₓ in nm)

🔧 Hyperparameters

• n_estimators: 134
• max_depth: 78
• max_features: 0.3007
• min_samples_split: 2
• bootstrap: True

Evaluation Metrics

• Mean Squared Error (MSE)
• Coefficient of Determination (R²)
• Parity plot and error histogram (saved in /figures/)

---

Results Summary

• Test R²: 0.9339
• Test MSE: 748.52 nm²
• Improved results after integrating outlier flagging logic

---

How to Use

1. Set up Dataset

Download the experimental chromophore dataset from: DB for Chromophore – Figshare

Save the .csv file inside: /dataset/raw_dataset/

2. Run Pipeline

python script/run_pipeline.py


This will process the data, train the model, and save outputs to processed_dataset/ and figures/.