Business Context

NovaThera runs a high-throughput biology platform that executes ~40,000 wet-lab experiments per month. The R&D leadership team wants a model that predicts whether an assay run will succeed, but the senior experimentalist approving deployment is skeptical of black-box AI and requires a clear, evidence-based explanation of the model.

Dataset

You are given historical assay-run data collected over 24 months.

• Size: 96K assay runs, 54 features
• Target: Binary — assay success (1) vs failure/repeat required (0)
• Class balance: 68% success, 32% failure
• Missing data: 9% missing in telemetry due to sensor dropouts, 6% missing in sample metadata for external partners

Success Criteria

A good solution should achieve strong predictive performance while producing explanations a senior experimentalist can validate against domain knowledge. The model should reach ROC-AUC >= 0.86, PR-AUC >= 0.74, and provide both global and per-run explanations.

Constraints

• Explanations must be understandable to non-ML stakeholders
• Batch inference must score 40K runs in under 10 minutes
• Retraining is allowed monthly, not daily
• The final recommendation must compare the neural network to a simpler baseline

Deliverables

1. Train a baseline interpretable model and a neural network classifier
2. Evaluate both models on a held-out test set with appropriate metrics
3. Produce feature-level and example-level explanations for the neural network
4. Describe how you would explain the architecture, predictions, and limitations to a skeptical senior experimentalist
5. Recommend whether the neural network is suitable for production given performance and interpretability tradeoffs