Steering Conformational Sampling in Boltz-2 via Pair Representation Scaling

Deep learning has transformed protein structure prediction; however, most systems predominantly return a single dominant conformation with limited control over alternative states. We introduce Boltz-sample as a systematic steering strategy for Boltz-2 that modulates conformational sampling by uniformly rescaling the latent pair representation. In contrast to heuristic input perturbations such as MSA subsampling that rely on trial-and-error to induce variability, this operation systematically modulates the effective strength of pairwise couplings while preserving the input sequence and alignment. Across diverse benchmarks including membrane transporters and a curated panel of 15 multi-state targets, Boltz-sample significantly improves the recovery of both alternative states and ensemble coverage compared to standard inference. Crucially, our analysis of sequence-only inference reveals that the method does not rely solely on coevolutionary signals but actively unlocks the structural priors internalized by the model to recover alternative states even without MSA information. From a practical standpoint, the sign of the scaling parameter defines distinct search directions in the latent space and enables the efficient retrieval of diverse ensembles through simple confidence-based selection. Boltz-sample thus offers a transparent framework to navigate the latent conformational landscape and shifts the paradigm from unpredictable input variation to systematic latent space steering.