Submission

Submit your notebook here

Deliverables

Submit your completed notebook (.ipynb) with:

1. Results from at least 3 clustering methods
2. Comparison table (clusters, sizes, runtime)
3. Train/test accuracy for each clustering approach
4. Reflection answers in markdown cells

---

Exercise 4: Reflection

Goal: Synthesize what you learned.

Answer in your notebook (2-3 sentences each):

1. Which clustering method gave the most/fewest clusters? Why do you think that is?

2. Did the choice of clustering method affect your model's accuracy much? What does that tell you?

3. If you were starting a new protein ML project, which clustering approach would you use and why?

🧗 Flash to the top!

---

Exercise 3: Compare Everything

Goal: See if clustering method matters for model performance.

For each clustering method you tried:

1. Do a cluster-aware train/test split
2. Train a classifier (Random Forest works well)
3. Record the accuracy

Create a comparison table:

Questions:

• Does the clustering method significantly affect accuracy?
• Which method is fastest?
• Which gives the most "balanced" clusters?

🤸 Campus to next exercise

---

Exercise 2: Try Other Methods

Goal: Explore different clustering algorithms.

You've already done threshold + connected components in R019. Now try these:

Option A: Agglomerative Clustering

from sklearn.cluster import AgglomerativeClustering

# Convert similarity to distance
distance_matrix = 1 - similarity_matrix

clustering = AgglomerativeClustering(
    n_clusters=None,
    distance_threshold=0.2,  # adjust this
    metric='precomputed',
    linkage='average'
)
labels = clustering.fit_predict(distance_matrix)

Option B: DBSCAN

from sklearn.cluster import DBSCAN

distance_matrix = 1 - similarity_matrix

clustering = DBSCAN(
    eps=0.2,  # adjust this
    min_samples=2,
    metric='precomputed'
)
labels = clustering.fit_predict(distance_matrix)

Try at least two of these methods. Experiment with the thresholds.

Questions to explore:

• How many clusters does each method produce?
• How does changing the threshold affect the results?
• Do some methods create more "singleton" clusters (clusters with just one protein)?

🦵 Toe hook to next exercise

---

Exercise 1: Recap Your R019 Approach

Goal: Establish a baseline with the method you already know.

In R019, you clustered proteins using:

1. Cosine similarity matrix
2. Threshold (e.g., 0.8)
3. Connected components

Recreate this clustering (or load your results from R019). Record:

• Number of clusters
• Size of the largest cluster
• How long it took

This is your baseline for comparison.

🧱 Stem to next exercise

---

Exercise 0: Setup

Goal: Load your similarity matrix from R019.

You should already have:

• Mtb protein embeddings loaded
• Similarity matrix computed
• Labels for localization prediction

If you saved your similarity matrix from R019, load it. Otherwise, recompute it:

from sklearn.metrics.pairwise import cosine_similarity
similarity_matrix = cosine_similarity(embeddings)

Success check:

• You have your similarity matrix ready
• You have your labels ready

🦶 Rockover to next exercise

---

Why this route exists

In R019, you learned one way to cluster proteins: threshold + connected components. It works, but it's not the only option.

Different clustering methods make different assumptions and produce different results. Some create many small clusters, others create fewer large ones. Some are fast, others are slow. Some handle outliers well, others don't.

In this route, you'll try multiple methods and see how the choice affects your train/test splits and model performance.

What you'll be able to do after this route

By the end, you can:

• Apply multiple clustering algorithms to protein data
• Compare clustering results (size, count, runtime)
• Understand trade-offs between different methods
• Make informed choices about clustering in your projects

Key definitions

Agglomerative Clustering Bottom-up approach: start with each point as its own cluster, then merge the closest pairs until a stopping criterion is met.

DBSCAN Density-based clustering: finds clusters as dense regions separated by sparse regions. Good at finding outliers.

Linkage In agglomerative clustering, how you measure distance between clusters. "Average" uses mean distance between all pairs.

💪 Dyno to first exercise

---

Route 020: Clustering Methods Showdown

• RouteID: 020
• Wall: The Machine Learning Offwidth (W06)
• Grade: 5.9
• Routesetter: Adrian
• Time: ~30 minutes
• Dataset: Same as R019 (Mtb embeddings + similarity matrix)

🧗 Start climbing!