2025 MOHD Imputation Hackathon

Welcome to the 2025 MOHD Imputation Hackathon!

The immediate goal of this hackathon is to prototype and compare methods for multi-omic data imputation using CCLE data. In the long term, we hope to build a community interested in writing a Nature Methods Registered Report. Preparing such a Registered Report would benefit from dividing tasks across a group, and what we do here (creating baselines and designing a benchmarking pipeline) is a toy version of what that larger project could look like.

---

Getting Started

1. Clone the repo

2. Set up your environment

Dependencies right now are quite minimal: numpy, pandas, scipy, scikit-learn.

---

Data

The data splits are provided as pickled dictionaries.
Each split file is a dictionary of {modality_name: pandas.DataFrame} where:

• Rows = samples
• Columns = features

Available splits:

• Data/ccle_split_train.pkl
• Data/ccle_split_val.pkl
• Data/ccle_split_test.pkl

👉 Data for the hackathon can be found at this link

---

Code Structure

• base_imputer.py — defines the BaseImputer API (all models must inherit or follow this interface).
• global_mean.py — baseline model that predicts feature-wise global mean.
• random_copy.py — baseline model that copies feature values from a random training sample.
• metrics.py — evaluation metrics (MAE, RMSE, R², Spearman, Pearson).
• Testing_model.ipynb — notebook to test your model. (Note: please use this notebook locally to check your code, but do not push changes to it in the repo.)

---

Writing Your Own Model

To add a new model:

1. Create a new file, e.g. my_model.py.
2. Define a class that implements the same API as the baselines:

from base_imputer import BaseImputer

class MyModel(BaseImputer):
    name = "my_model"

    def fit(self, train, input_modalities, target_modalities):
        # train is dict[modality -> DataFrame]
        # store anything you need here
        return self

    def predict(self, inputs, target_modalities):
        # return dict[target_modality -> DataFrame]
        return {target_modality: preds_df}

3. Import it in the notebook and run with the shared evaluate_model function.

---

Evaluation

We provide a common evaluation function (evaluate_model) to compare models consistently.
It will:

• Fit your model on the training split.
• Predict the target modality on the validation/test split.
• Compute metrics: MAE, RMSE, R², Spearman, Pearson.
• Return both the results dictionary and the predictions DataFrame.

---

Example

from random_copy import RandomTrainingSampleImputer
from metrics import METRICS
from Testing_model import evaluate_model, load_split_dict

train = load_split_dict("Data/ccle_split_train.pkl")
test  = load_split_dict("Data/ccle_split_test.pkl")

target = list(train.keys())[0]
inputs = [m for m in train if m != target]

model = RandomTrainingSampleImputer(seed=42)
res, preds = evaluate_model(model, train, test, inputs, target)
print(res)

---

Hackathon Schedule

Day 1

12:20–12:35 — Kickoff

• Explain structure (3 groups).
• Explain CCLE data that is already processed.
• Show repo + input/output schema for metrics.

12:35–1:30 — Work

• Group A: start coding metrics and plots for given output structure.
• Group B: implement KNN imputation.
• Group C: implement LASSO imputation.

1:30–1:40 — Sync

• Group A shows sample metrics working on CCLE data + toy model.
• Group B and C share preliminary results from imputation methods.

1:40–2:00 — Consortium update

• Present early progress: evaluation pipeline in progress + baseline methods.

---

Day 2

10:45–10:50 — Quick regroup

10:50–11:50 — Work

• Group A: finalize metrics and make demo plots on existing predictions.
• Group B and C: perform some hyperparameter optimization
  (Nothing too complicated — just a small grid search).

11:50–12:15 — Final sync & consortium prep

• Group A runs metrics on top predictors from Group B and C.
• Combine outputs into single deck:
  • Slide 1: Hackathon goals
  • Slide 2: Evaluation metrics & demo results
  • Slide 3: Baseline/imputer comparisons
  • Slide 4: Next steps

---