Accurate modeling of antibody–antigen (Ab–Ag) complexes is central to biologic development, yet the reliability and failures of modern Ab–Ag folding pipelines remain poorly characterized. Single-chain variable fragments (scFvs) are thera-peutically important antibodies, but large-scale evaluations of structure prediction models on scFv–Ag complexes are largely lacking. We introduce a scalable bench-marking pipeline that generates large ensembles of scFv–Ag structure predictions by cofolding a curated subset of 3,800 Ab–Ag complexes from SAbDab using multiple state-of-the-art models under diverse inference-time settings. The resulting dataset, SCALE (scFv–Ag CompLex Ensembles), includes standardized scFv–Ag sequences and around 200,000 predicted complexes spanning different models, sampling strategies, and auxiliary inputs. Using SCALE, we evaluate model performance in recovering correct scFv–Ag interfaces and assess the ability of existing confidence metrics to select the best structure from prediction ensembles. We find that while confidence scores effectively distinguish easy from hard scFv–Ag complexes, they often fail to identify the highest-quality interface for a given target. Further analysis shows that near-correct interfaces typically appear in ensembles but at low frequency, and inference-time choices such as sampling, recycling, and using evolutionary or structural information are crucial for accurate scFv–Ag complex predictions. Dataset and analysis code are available at https://huggingface.co/datasets/ravishah1/SCALE
Ravi K Shah, Jeffrey Ouyang-Zhang, Zachary Cohen et al.· bioRxiv· 0 citations
Vision--Language Models (VLMs) are increasingly deployed through a model supply chain in which pretrained checkpoints, architecture definitions, text encoders, and exported computation graphs are distributed by third parties and reused across downstream services. This reuse model creates a security-critical trust boundary: VLM deployments inherit not only learned parameters but also executable behavior encoded in shared model artifacts. In this paper, we show that a malicious provider can exploit this trust boundary by embedding architectural backdoors into VLM supply chains through representation steering. Our attack introduces dormant steering logic into the model architecture through a trigger-gated additive modification of an intermediate representation, without poisoning training data, controlling downstream fine-tuning, or modifying prompts at deployment time. When the trigger is absent, the modification reduces to zero and the model follows its normal computation, preserving clean utility. When the trigger is present, a steering direction shifts the internal representation toward an attacker-defined objective. We evaluate the attack across multiple VLM families and downstream tasks, including visual question answering, text-to-image generation, retrieval, and semantic response biasing. The results show that the proposed architectural steering backdoor compromises integrity, safety enforcement, and ranking fairness while preserving normal behavior on clean inputs. We further show that shared VLM artifacts can carry dormant steering logic against downstream services, and we propose an auditing defense that inspects the executable logic distributed with model artifacts rather than only their learned weights.
M. Briglia, Igor Maljkovic, A. E. Cinà et al.· 0 citations