#18 · ML / AI methods

Compact protein representations probe structure inference from sparse or degraded inputs

The dominant thread is whether compact learned representations recover biologically meaningful structure rather than merely fitting outputs. The frequency-space preprint encodes proteins as harmonic graphs of MD vibrational modes and predicts GO molecular function without sequence on SwissProt. The ESM-2 contact-graph study projects residue embeddings onto graphs and applies SoftBlobGIN with Gumbel-softmax substructure pooling to probe interpretability. The BINN preprint shows networks can match input-output mappings while failing to recover gene-to-pathway weights and pathway activations, a direct test of representational fidelity. A sparse autoencoder analysis of ESM-2 and ESM-3 reports 78% feature convergence across architectures and finds that structure tokens sharpen a shared feature vocabulary. What is new is explicit scrutiny of whether internal representations are faithful, not only predictive. The items differ in task and data type, so this is a dominant methodological trend rather than a shared result.