Zero-Shot Generation of Protein Conformational Ensembles Through AlphaFold Latent Flooding

Despite transformative advances in protein structure prediction, generating conformational ensembles directly from sequence in an efficient and accessible manner remains a central challenge. We introduce a heuristic importance sampling framework for the zero-shot prediction of functionally relevant, conformationally diverse structures, bypassing the need for extensive physics-based modeling or prior domain knowledge. Guided by an analysis of massive activations in AlphaFold (AF), we develop a latent flooding algorithm that enables adversarial exploration of the latent space while preserving local geometric integrity. We show that the AF latent flooding (AFLF) method robustly recovers experimental structural fluctuations, captures function-relevant conformational states, and reveals cryptic binding sites across diverse protein systems. These results suggest that AF latent features implicitly encode the biophysical principles of protein thermodynamics, which AFLF exploits. Finally, we show that AFLF is computationally accessible and interoperable, offering a test-time generalization of AF for investigating protein dynamics and accelerating structure-based ligand discovery.