Scientific ML: Single-cell modality prediction

The agent gets paired single-cell measurements, one modality provided and one held out per cell, and builds a model predicting the held-out modality. It trains offline and submits through a sealed leaderboard gateway.

Single-cell assays measure complementary views of the same cell, but collecting every modality for every cell is expensive, and some measurements destroy the sample. If one modality predicts another well, the missing measurements come for free, and the quality of that prediction is itself a statement about how tightly the two biological layers are coupled.

The metric is the mean per-cell Pearson correlation between predicted and measured values. Scored submissions are quota-limited and return public scores only; the reward is the floor-normalized private score of the best-public submission.

The graded runs are nearly interchangeable, and their experiment logs are too. The same shallow playbook appears throughout: dimensionality-reduced features feeding ridge regression, small MLPs, and nearest-neighbor blends. Each run diversified its models with noise augmentation and seed averaging, then re-weighted the blend round after round.

The endgames were grinding. One run logged five consecutive experiments that left its local validation score unchanged to the fifth decimal: recency weighting, adaptive weights, ensemble-member selection. Another closed with a blend-grid expansion whose own verdict was no change. Nothing beyond that shallow family appears in any log, so the shared plateau reads as the ceiling of this recipe. Whether the task has more to give, none of these runs brought a tool that could tell.