Applied Machine Learning - Bioinformatics Case Study

This repository contains the code, data processing pipeline, and exploratory analysis for the Applied Machine Learning coursework case study (2026).

The project is structured as a small Python package using a src/ layout, with data loading and preprocessing logic separated from notebooks and analysis code.

---

Objective

The dataset reflects the operational state of a large-scale biotechnological process as it was 5 years ago. Product yield across batches is variable, and the aim of this case study is to identify the process variables that contribute to this loss of productivity across the highest-yielding batches.

Sensor data are captured every 5–15 minutes across 17 process parameters for 22 batches (the "4000 series operating data"), and laboratory data are captured at ~4-hour intervals for 21 of those batches.

The goal is to use the 21 "training" batches to identify the key process variables that influence product yield, and to build a purely data-driven model to predict the yield of the 22nd ("held-out") batch.

A video describing the approach and findings is available here.

An infographic of the project is available in Infographic Applied ML.pdf.

---

Results

Several ML models were evaluated using two feature representations: aggregated batch statistics (B) and time-series-derived features (D). Models were ranked by RMSE (lower is better):

Model	Feature Set	RMSE
HistGB	Batch Statistics	26.01
Ridge	Time-Series	23.56
Ridge	Batch Statistics	22.23
ElasticNet	Time-Series	22.21
Lasso	Time-Series	22.11
RandomForest	Time-Series	17.48
RandomForest	Batch Statistics	16.29
XGBoost	Batch Statistics	13.72
XGBoost	Time-Series	9.54
ElasticNet	Batch Statistics	8.93
HistGB	Time-Series	8.89
Lasso	Batch Statistics	7.49

Lasso regression on batch statistics performed best, achieving the lowest RMSE of 7.49. Using this model, the predicted yield for the held-out 22nd batch was 539.13 kg/hr.

---

Setup Instructions

1. Create and activate a virtual environment

python -m venv .venv

On Windows:

.venv\Scripts\activate

On macOS/Linux:

source .venv/bin/activate

---

2. Install dependencies

pip install -r requirements.txt

---

3. Install the project package (editable mode)

pip install -e .

This makes the applied package importable in notebooks and scripts.

---

Usage

After installation, project modules can be imported as:

from applied.data_processing import build_features_and_target

Raw data are loaded from the data/ directory, and notebooks are intended to be run from the notebooks/ folder.

---

Notes

• Tested with Python 3.10+
• The src/ layout is used to avoid import ambiguity and ensure clean separation between source code and analysis notebooks.
• Parallel sensor channels (e.g. liquid inflow streams) are aggregated based on physical interpretation, as described in the accompanying analysis.

---