Learning the origins of protein energy landscapes from large-scale experiments

Every protein sequence has its own energy landscape, defined as its set of conformational states and their populations in a given environment. These energy landscapes vary widely from protein to protein, but the sequence and structural determinants of these landscapes are poorly understood. In part, this reflects the difficulty of experimentally characterizing energy landscapes. Characterizing the diverse folded, partially folded, and unfolded states in a protein’s energy landscape typically requires applying biophysical approaches that can only examine one protein at a time. To overcome this limitation, we developed a new approach to quantify energy landscapes for ~1,000 protein domains in parallel using pooled top-down hydrogen-deuterium exchange mass spectrometry. These experiments quantify both global folding stability and the energies of conformational fluctuations for each domain. Analyzing >5,000 domains revealed a wide range of different landscapes, with large differences in fluctuations even for homologous domains with similar global stability. Our large-scale data enabled us to statistically analyze the structural origins of the varied landscapes, revealing how different interactions modulate stability and fluctuations. Machine learning based on biophysical features and protein language models enabled us to predict specific residues that influenced the energies of conformational fluctuations, and we tested these predictions using large-scale measurements of mutant proteins.