Scientific ML: RNA 3D structure prediction

The agent predicts five candidate backbone-coordinate structures for each RNA sequence in the evaluation set. Offline training and inference; submissions through a sealed leaderboard gateway.

RNA function follows RNA structure, and experimental structure determination is slow enough that computational candidates guide most of the work. Protein folding has its landmark models; RNA is meaningfully behind, with less training data and floppier molecules, which leaves genuine headroom for methodology rather than scale.

Each target is scored by the best TM-score among the five candidates, averaged over targets, so the task rewards structural diversity in the candidate set, not just one good guess. Scored submissions are quota-limited and return public scores only; the reward is the floor-normalized private score of the best-public submission.

The wide spread here traces to a single fork. One side folds structures from scratch with a neural network predicting distance maps. The other retrieves candidates from known folds and selects among them.

The strongest run started on the neural side too, with a distogram model. It crossed quickly to template retrieval, deepened the search, and finished with a learned judge choosing which candidates fill the five slots. The middle run landed on the same side through an archetype library with shortlist-plus-consensus selection. It tested the neural route and rejected it when its decoder plateaued well below the shortlist. The weakest run spent its entire budget on that neural route: positional encodings, pair-type features, a three-model ensemble. Its final submission scored roughly where the strongest run's opening distogram attempt had started.