Appendix G. Chapter Exercises

The chapter presents the exercises as a flexible learning tool that can be used for review, self-study, discussion, or research preparation, with each item tagged by level, type, time, format, and supporting material so readers can choose what fits their goals and background. It also cautions that datasets, checkpoints, and APIs evolve over time, so learners should treat external resources as moving targets and adapt code to current library versions.

Across the book, the exercises are designed to reinforce the main methods chapter by chapter, starting with molecular representations and drug-discovery pipelines, then moving through virtual screening, property prediction, calibration, unsupervised learning, deep learning, structure-based design, generative models, graph neural networks, protein language models, and finally multimodal integration. The tasks progress from beginner coding and theory prompts to advanced, open-ended projects that ask readers to compare representations, tune models, evaluate metrics appropriately, interpret results critically, and connect different techniques into larger workflows.

A recurring theme is that good modeling in drug discovery depends not just on building a predictor, but on choosing the right representation, validating carefully, and understanding failure modes such as class imbalance, overfitting, leakage, calibration errors, noisy or incomplete oracles, and compounding errors in multi-stage pipelines. Many exercises push readers to test ideas on real datasets, compare alternatives, and think like practitioners by balancing accuracy, uncertainty, cost, interpretability, and experimental feasibility in realistic discovery settings.

1. End-to-end mini-campaign. Choose a biological target that interests you and assemble a dataset from a public repository (chapter 1 and Appendix B). Build one or more ADMET or toxicity filters (chapters 3–5), screen a vendor or generated library for candidate binders (chapters 2, 8, and 9), and optionally propose new molecules with a generative model (chapter 10). Deliver a short report: a ranked shortlist of at most 25 compounds, the filters and models you applied in what order, and an honest account of where your pipeline is weakest. The point is the documented reasoning, not the ranking.
Level: Challenge Type: Synthesis + Coding Time: multi-day project Format: Individual or group Reinforces: chapters 1–10; Appendix B
Objective: run a realistic discovery funnel end to end and defend every design decision in writing.
Hint: Freeze your evaluation plan (splits, metrics, applicability-domain checks) before you look at results, so you cannot unconsciously tune toward a nicer-looking shortlist.
Extension: Hand your report to a peer and have them try to break your strongest claim; revise accordingly.
2. Reproducibility audit. Pick any one chapter’s notebook and write a two-to-three-page reproducibility report. Document the data provenance, the random seeds, the train/validation/test split strategy, the metric definitions, and the applicability-domain assumptions. Then identify the single change that would most improve the reproducibility or the honesty of the evaluation, and justify it. This exercise builds the habit of reading your own and others’ pipelines the way a reviewer would.
Level: Advanced Type: Critical thinking Time: half-day Format: Individual or group Reinforces: chapters 3–9 (evaluation, splitting, applicability domain), esp. 3.4.4, 4.3.1, 6.2.2
Objective: audit a real ML pipeline for reproducibility and evaluation integrity.
Hint: The usual culprits are an undocumented split, a metric computed on resampled data, and a missing applicability-domain check — look for those first.
Extension: Re-run the notebook after your proposed change and report whether the headline numbers move.
3. A theory of representations. Write a short essay (two to three pages) comparing the molecular and protein representations used across the book: hashed fingerprints (chapters 2–8), SMILES sequences (chapter 10), molecular graphs (chapter 11), 3-D conformer descriptors (chapter 6), and learned protein-language-model embeddings (chapter 12). For each, state what it encodes, what it discards, how much labeled data it typically needs, and one task where it is the natural choice. Close with a decision guide a newcomer could follow.
Level: Advanced Type: Theory + Synthesis Time: half-day Format: Individual Reinforces: chapters 2, 6, 8, 10, 11, 12
Objective: consolidate the book’s representations into a single comparative framework you can reason from.
Extension: Turn your decision guide into a one-page flowchart and test it against three tasks drawn from different chapters.
4. From paper to plan. Find a recent paper (last two or three years) that builds on any technique from this book — an active-learning screen, a generative model, a GNN affinity predictor, or a protein-language-model application. Summarize what the authors did and why, then write a concrete reproduction or extension plan: what data and compute you would need, which parts you could run on a single GPU, and what you would check first to know whether you were on track. Chapter 5’s multi-instance-learning exercise promised that by the book’s end you could return to a research paper and understand it in full; this capstone is where you cash that promise in.
Level: Challenge Type: Literature + Synthesis Time: multi-day project Format: Individual or group Reinforces: the entire book; connects explicitly to the chapter 5 challenge exercise
Objective: bridge from reading current literature to a credible implementation plan of your own.
Hint: Scope ruthlessly: pick the one figure or result you would try to reproduce first, and design the smallest experiment that would confirm or refute it.
Extension: Actually attempt the smallest experiment and write up what happened, including what did not work.

FAQ