This submission uses a retrained phase-2 ensemble built from the established model family selection, followed by a small posthoc pruning step and a final calibration pass. The goal was to keep the strongest parts of the ensemble pipeline while reducing redundant or harmful members and corr ecting residual prediction compression.
The underlying predictor is a weighted ensemble of diverse tabular and molecular modeling families trained on the full available training data fo r phase 2. The ensemble mixes complementary model classes, including:
- kernel and tree-based tabular regressors
- graph/message-passing style molecular models
- pretrained tabular/foundation-style predictors
The feature space spans multiple learned and engineered molecular representations. Representative examples include:
- MOE-style descriptor sets
- continuous learned molecular embeddings
- high-throughput screening-informed learned representations
The ensemble weights were determined from internal validation predictions using simplex-style convex weighting so that the final prediction is a non-negative weighted combination of member outputs.
After selecting the ensemble, we applied a leave-one-member-out style pruning step to remove a member that consistently hurt internal validation performance when combined with the rest of the ensemble. This was intended as a conservative refinement rather than a broad model rewrite:
- keep the main ensemble structure
- remove one underperforming or redundant contributor
- renormalize the remaining ensemble weights
This pruning step improved the validation profile of the ensemble and was carried forward into the final full-data retrain workflow.
The raw ensemble predictions showed the usual regression compression pattern, with predictions pulled too strongly toward the center of the activity range. To address this, we applied a posthoc calibration model fitted on internal validation predictions.
For the pruned variant, the chosen calibration family was a piecewise linear mapping. This was selected because it gave the strongest internal validation behavior among the calibration methods compared on the pruned ensemble variant. In practical terms, the calibration step:
- preserves the overall ranking signal from the ensemble
- stretches predictions away from the mean where needed
- improves agreement across low, middle, and high activity regimes better than a single global linear correction
The final submission was produced in the following sequence:
1. retrain the selected run2 ensemble on the full phase-2 training set 2. apply the same posthoc single-member prune used in the earlier run2 refinement
3. recompute the ensemble prediction from the remaining weighted members 4. apply piecewise linear calibration to the pruned ensemble prediction
5. export calibrated predictions for the full blind test set
This approach was chosen because it is a high-confidence extension of the strongest existing pipeline rather than a speculative redesign. The mai n ingredients were all motivated by prior validation evidence:
- diverse ensemble members improved robustness over any single model family - simplex weighting provided a stable way to combine heterogeneous predictors
- light posthoc pruning reduced redundancy and removed a harmful contributor
- calibration corrected residual compression and improved the final prediction surface