Protein Secondary Structure Prediction

The task is residue-level classification, not whole-protein prediction.

We predict an eight-state secondary-structure label for every amino acid in a protein sequence. The model first learns a token embedding, then uses parallel 1D convolutions to extract local motifs. Three bidirectional recurrent layers add sequence-wide context, while skip connections preserve earlier representations. Two dense layers and a time-distributed softmax generate one label per position.

The 3.45m-parameter configuration uses Adam at a learning rate of 0.0004 and 50% recurrent dropout in the later recurrent layers. Ten recorded validation runs average 64.05%, ranging from 59.52% to 65.80%. Because the training workflow reused one network instance across the runs, the mean is a recorded benchmark rather than an independent cross-validation estimate.

1. The architecture matches the structure of the problem. Convolutions capture short motifs; bidirectional recurrence captures upstream and downstream context.
2. The result stabilizes after the first run. Nine of ten recorded runs fall between 63.06% and 65.80%.
3. The evaluation is directional, not definitive. Reusing one trained model across runs makes the later validation scores interdependent.