Seamann, A. | Search bioRxiv

bioRxiv2025

Constrained Diffusion for Protein Design with Hard Structural Constraints

AO_SCPLOWBSTRACTC_SCPLOWDiffusion models offer a powerful means of capturing the manifold of realistic protein structures, enabling rapid design for protein engineering tasks. However, existing approaches observe critical failure modes when precise constraints are necessary for functional design. To this end, we present a constrained diffusion framework for structure-guided protein design, ensuring strict adherence to functional requirements while maintaining precise stereochemical and geometric feasibility. The approach integrates proximal feasibility updates with ADMM decomposition into the generative process, scaling effectively to the complex constraint sets of this domain. We evaluate on challenging protein design tasks, including motif scaffolding and vacancy-constrained pocket design, while introducing a novel curated benchmark dataset for motif scaffolding in the PDZ domain. Our approach achieves state-of-the-art, providing perfect satisfaction of bonding and geometric constraints with no degradation in structural diversity.

molecular biology↗

bioRxiv2025

Form Follows Function: A New Paradigm of Pancreatic Cancer Progression

Pancreatic ductal adenocarcinoma (PDAC) exhibits a distinctive propensity to invade nearby organs and infiltrate large blood vessels, even in the absence of distant metastasis. While the genetic and transcriptomic drivers of PDAC progression have been well studied, the mechanisms by which these molecular changes translate into functional, invasive behavior remain largely unknown. Here, we uncover a striking level of tissue organization, characterized by previously unrecognized spatial and geometric properties within and among tumor structures. Leveraging the first large-scale, AI-assisted, human-curated PDAC atlas from hematoxylin and eosin (H&E) images, we annotated, classified, and characterized 144,474 malignant and normal structures from treatment-naive (n=118) and neoadjuvant-treated PDAC patients (n=50). Additionally, we developed a new computational tool, SHAPE, to investigate PDAC aggressiveness through a comprehensive "geometrization" of cancer progression. Using traditional H&E-stained slides and three-dimensional (3D) tissue reconstruction experiments, we observed that invading tumor structures display an eccentric morphology with pronounced local angular coherence. These geometric and spatial properties revealed coherent architectural patterns, with invasive structures closely tracking vessels and nerves as they infiltrate surrounding tissue. Mechanistically, integration of morphological features from 39,045 annotated tumor structures with whole-genome and RNA sequencing data revealed that PDACs with numerous eccentric structures exhibit increased copy number alterations (CNAs), loss of heterozygosity (LoH) on the p-arm of chromosome 17, and a quasi-mesenchymal/basal-like molecular subtype. Spatial transcriptomic analysis of 1,650 tumor structures from six additional PDAC patients further confirmed upregulation of invasive cellular programs within highly eccentric structures, such as epithelial-to-mesenchymal transition (EMT), angiogenesis, coagulation, and complement pathways, underscoring their infiltrative nature. Finally, cross-validation of our AI-based method enabled a fully automated, highly interpretable computational approach to assist pathologists and clinicians in evaluating neoadjuvant chemotherapy response, predicting patient survival, and guiding chemotherapy in adjuvant settings. Collectively, these findings deepen our understanding of PDAC progression, identify a new hallmark of tumor architecture, and pave the way for full integration of AI-driven morphology-based approaches into clinical workflows to improve the management of PDAC patients.

cancer biology↗