Protein Folding Prediction

A protein must fold into a useful shape, but it often works by meeting another molecule. An antibody must fit its antigen; a DNA-binding protein must approach the right surface of the double helix. Predicting each partner separately cannot tell us whether they assemble correctly.

FoldBench tests one shared model in nine settings: isolated proteins, RNA and DNA, plus six kinds of molecular partnership. The challenge is to learn common rules of molecular geometry without neglecting the less familiar partners.

The training pool contains 100,000 molecular records. Protein sequences also arrive with fixed features from a protein language model: numerical descriptions learned from sequence patterns. RNA, DNA and small molecules require their own descriptions. The model must turn this mixed information into positions for individual atoms.

Evaluation structures are visible for diagnosis, but cannot be used for training. Repeatedly choosing models using these results makes the benchmark part of development, so it is not an untouched test of generalization. The interactive structures above are schematic.

The starter first represents individual atoms and relationships between molecular positions. These relationships let information cross a chain boundary: the location of one partner can influence the predicted shape of another.

A diffusion module then turns noisy coordinates into a proposed structure. Think of gradually resolving an initially blurred arrangement of atoms while keeping track of their identities. The folding model is trained from scratch; fixed protein-language features provide useful inputs without supplying a trained structure predictor.

Training asks the model to recover structures after their coordinates have been disturbed, and to predict how far pairs of atoms should be apart. These two signals teach complementary aspects of geometry.

Neither signal guarantees chemistry. Atoms can occupy approximately the right region while bonds have impossible angles or handedness. That is why a visually convincing fold must still pass atomic checks.

For isolated molecules, lDDT measures agreement in local atomic distances. For complexes, success depends on how accurately the partners meet. Each of the nine settings contributes equally to the final score; a failed prediction remains in the denominator.

This distinction is visible in the starter: its protein folds have a moderate overall shape score of 0.53, yet their scored local accuracy is zero after stereochemical checks. A recognizable silhouette is insufficient if the atomic structure is invalid.

Across 1,522 requests, the starter averages 0.037 out of one. A pretrained ESMFold2-Fast reference reaches 0.545, but it uses a different training history and is not an eligible from-scratch competitor. The comparison shows room for improvement; it does not identify which architectural change would close the gap.

The starter occasionally succeeds at interfaces, especially protein–peptide binding, while all three isolated-molecule scores are zero. The most useful next step is to repair atomic geometry while checking that partner placement also improves.

Compare a geometry-aware model with the starter, measuring both chemical violations and structural accuracy. Then ask whether a more balanced mixture of molecular partners improves RNA, DNA and ligand cases without harming proteins. All nine results matter: a higher average can conceal a newly neglected task.