Reproduce “Neural spectroscopy of AlphaFold2 reveals encoded protein conformational landscapes” on OpenResearch

About this paper

Kaustav Mehta

This study introduces "neural spectroscopy" to probe AlphaFold2's Evoformer weights, demonstrating they encode protein conformational landscapes beyond single static structures. The systematic perturbation method reveals a physically coherent unfolding pathway for ubiquitin, showing strong correlation with molecular dynamics simulations and experimental flexibility, and highlights the model's distinct responses for fold-switching and intrinsically disordered proteins.

Key insights

• AlphaFold2's Evoformer weights encode a structured, data-driven representation of protein conformational space, functioning as a "scientific object" that internalizes aspects of protein folding dynamics.
• The systematic perturbation reveals that AlphaFold2's learned representations can produce physically plausible conformational transitions, such as unfolding pathways consistent with experimental stability ordering and molecular dynamics.
• The consistency of the encoded landscape for well-characterized proteins (ubiquitin) across independently trained models, alongside structured divergence for intrinsically disordered proteins, maps the boundaries and reliability of AlphaFold2's learned "knowledge."

Summary by alphaXiv.