Boltz-Perturb: Probing Generalization in Co-Folding Models via Inference-Time Perturbation

AlphaFold3-style co-folding models are powerful deep generative models (DGMs) for biomolecular structure prediction, yet they exhibit low-diversity outputs that frequently miss correct small molecule binding modes. We present Boltz-Perturb, a training-free perturbation framework that injects time-annealed noise into the conditioning signals of Boltz-2's denoising module during inference. Through true-coordinate injection experiments, we establish that the model's learned energy landscape already encodes correct binding-mode basins, reframing the problem as a sampling deficiency. We introduce Token Bias Perturbation (TBP) and Token Conditioning Perturbation (TCP), targeting attention biases and token embeddings, respectively. Notably, TCP raised oracle success rate from 17.7\% to 30.6\% using three-fold fewer samples across 62 targets. Our results support that co-folding models learn more than they express, but redundant, sample-identical conditioning masks the sampling paths toward correct structures. To our knowledge, this is the first perturbation analysis of an AF3-style co-folding architecture for small-molecule binding mode diversity, demonstrating that conditioning perturbation can unlock latent capacity for drug discovery without retraining.