Course materials for the M2 teaching unit, Master 2 Chimie — Sciences analytiques (3 ECTS, 15 h lectures / 6 h practicals).
Coordinator: Lennart Martens. Instructors: Lennart Martens, Robbin Bouwmeester and Jérôme Pansanel.
- All materials and links to them are below
- Practicals are done in the form of notebooks; we recommend you can run them on google colab (requires a google account) or locally (we can help you with the installation, but this is much more involved)
- Answers for all practicals are available
- Complete the Foundations quiz after day 2, the Evaluation quiz after day 3 and the Applications quiz after day 4. Each quiz follows the lecture and both corresponding practicals.
- Based on timing we might change when specific subjects are discussed
The three quizzes together count for 40% of the final mark. The written exam counts for the remaining 60%. The LLM-assisted project is part of the day 5 practical.
These notes support the lectures and practicals, with consistent terminology and explanations for students who do not program.
| Topic | Notes |
|---|---|
| Foundations (days 1 and 2) | Foundations notes |
| Evaluation | Evaluation notes |
| Applications | Applications notes |
| LLM tools and the peptide project | LLM Tools notes |
| Day | Main question | Materials | Quiz |
|---|---|---|---|
| 1. Foundations | How does a model learn to predict from data? | Lectures and practicals | Foundations quiz follows day 2 |
| 2. Foundations and Data | How do neural networks learn, and what data and infrastructure support an AI analysis? | Lectures and practicals | Foundations quiz |
| 3. Evaluation | Which model should we trust on new examples? | Lectures and practicals | Evaluation quiz |
| 4. Applications | Which measurements or compounds need a closer look? | Lectures and practicals | Applications quiz |
| 5. LLM tools | How can LLMs support research while we check their outputs? | Lectures and practical | — |
Grading · Lecture notes · Datasets
Introduction, decision trees, regression and Foundations practicals
| Topic | Slides |
|---|---|
| Introduction to machine learning | PowerPoint · PDF |
| Decision trees: bits, entropy and information gain | PowerPoint · PDF |
| Regression | PowerPoint · PDF |
| Logistic regression | PowerPoint · PDF |
Begin with the data, features and small decision tree in the Strawberry notebook, working through the steps together with the instructor. Then complete the corresponding steps in the Tox21 notebook independently. Continue both notebooks on day 2 for neural networks and molecular generation.
| Practical and format | Student notebook | Run the practical | Answers |
|---|---|---|---|
| Strawberry (guided): is this strawberry purée? | Notebook |  |
Notebook with answers · Colab |
| Tox21 (independent): predict activity and propose molecules | Notebook | |
Notebook with answers · Colab |
Foundations continued and Data. The Foundations quiz follows the second day.
Neural networks, Data, continued practicals and Foundations quiz
| Topic | Materials |
|---|---|
| Neural networks | PowerPoint · PDF |
| Convolutional neural networks | PowerPoint · PDF |
| Data and computation: quality, governance, software, storage and compute | Lecture materials are not included in this repository. |
The Data lecture covers data quality and lifecycle, FAIR principles and trusted repositories, governance and ownership, software reproducibility, storage, computing needs and costs, environmental impacts, and infrastructure access and selection. The practicals provide the examples of preparing model inputs.
Continue both notebooks using your saved answers from day 1. Together with the instructor, compare a small neural network with the tree in the Strawberry practical. Independently complete the Tox21 practical, using a pretrained autoencoder to propose related molecules and predict their assay classes.
| Practical and format | Student notebook | Continue in Colab | Answers |
|---|---|---|---|
| Strawberry (guided) | Notebook | |
Notebook with answers · Colab |
| Tox21 (independent) | Notebook | |
Notebook with answers · Colab |
Lecture slides, practicals and quiz
| Topic | Slides |
|---|---|
| Regularisation and failure modes: underfitting, overfitting, bias, out of distribution and information leakage | PowerPoint · PDF |
| Decision trees: Gini impurity and split quality | PowerPoint · PDF |
| Ensemble learning | PowerPoint · PDF |
Compare models using training and validation results, then check the fixed choice on test records. Work through the Strawberry practical together with the instructor, then complete the Tox21 practical independently. Tree-depth plots illustrate overfitting. The Strawberry practical also compares a small CNN using the full spectra.
| Practical and format | Student notebook | Run the practical | Answers |
|---|---|---|---|
| Strawberry (guided): which model should we trust? | Notebook | |
Notebook with answers · Colab |
| Tox21 (independent): compare models and inspect mistakes | Notebook | |
Notebook with answers · Colab |
Lecture slides, practicals and quiz
| Topic | Slides |
|---|---|
| Applications: proteomics, NMR, Raman and powder diffraction | PowerPoint · PDF |
Use an isolation forest to flag unusual measurements or descriptor patterns. Work through the Strawberry practical together with the instructor, then complete the Tox21 practical independently. Interpret the flags and decide what laboratory follow-up is needed. A flag requests review; its presence or absence does not establish purity or safety.
| Practical and format | Student notebook | Run the practical | Answers |
|---|---|---|---|
| Strawberry (guided): which purées need a closer look? | Notebook | |
Notebook with answers · Colab |
| Tox21 (independent): which compounds look unusual? | Notebook | |
Notebook with answers · Colab |
Lecture slides and practical
Use LLMs to support research while documenting prompts, checking outputs and keeping responsibility for the analysis.
In the peptide practical, build your own retention-index model with an LLM. The notebook supplies the data and an open project brief. Choose your features and model, compare with a baseline, and try an improvement using validation data before the final test. Explain your decisions and document your prompts and checks.
| Topic | Slides |
|---|---|
| Recurrent neural networks and transformers | PowerPoint · PDF |
| Foundation models | PowerPoint · PDF |
| LLMs as research tools: overview | PowerPoint · PDF |
| Practical | Student notebook | Run the practical | Answers |
|---|---|---|---|
| Peptide retention-time prediction: prompt, build and verify | Notebook | |
Worked example · Colab |
| Dataset | Used for | Documentation and download |
|---|---|---|
| Strawberry fruit purées | Spectral classification and outlier detection | Background, preparation and licence · CSV download (.gz) |
| DeepChem Tox21 | Assay prediction, molecular generation and descriptor outliers | Background, preparation and provenance · CSV download (.gz) |
| Peptide retention indices (DLOmix / ProteomeTools) | LLM-assisted regression project in Part 5 | Background, preparation and provenance · Classroom CSV backup (.gz) |
Each practical loads its data automatically and can be started independently. Within the Strawberry and Tox21 series, the three notebooks reuse preparation and test records for teaching; this reuse is not new independent validation. Tox21 assay predictions do not establish overall toxicity or human safety. Dataset licences are documented separately from the course repository licence.