From shape to fate: making bacterial swarming expansion predictable

TL;DR

This work reframes bacterial swarming expansion as a geometry-driven forecasting problem, transforming descriptive front morphology into a predictive dynamical system. It introduces SwarmEvo, a boundary-resolved time-lapse dataset, TexPol--Net for stable boundary recovery, and Morpher with Morphon memory to forecast front evolution in morphology space. Morpher consistently outperforms leading video-prediction models in maintaining front localization and anisotropic fingering, with segmentation quality directly impacting forecast stability. The framework enables quantitative interrogation and potential control of microbial collectives on mucosal surfaces, offering a pathway toward predictive mucosal repair and gut ecosystem engineering.

Abstract

Microbial swarming on mucosal surfaces reshapes microbial communities and influences mucosal healing and antibiotic tolerance. Yet even with time-lapse microscopy and deep learning, analyses of swarming colonies remain descriptive and cannot forecast how their fronts reorganize in time. This limitation is significant because the advancing edge determines access to nutrients, host tissue and competing microbes. We recast the expansion of Enterobacter sp. SM3 swarms as a problem of morphological forecasting, and assemble SwarmEvo, a time-lapse dataset represented as boundary-resolved segmentations. TexPol--Net, a texture- and geometry-aware segmentation model, sharpens diffuse edges and preserves fingered fronts, creating a stable substrate for dynamics. On this representation, we develop Morpher, an autoregressive forecasting network with a ``Morphon'' memory that links local curvature to long-range temporal dependencies. Morpher outperforms leading video-prediction models in maintaining front localization and anisotropic branching, and modest segmentation improvements yield noticeably more stable forecasts. Ablations across sequence models, inference strategies and observation ratios show that attention-based architectures with structural memory best preserve dense-finger propagation. By uniting geometry-aware segmentation with morphology-level forecasting, this framework turns swarming expansion into a predictive dynamical system, enabling quantitative interrogation and potential control of microbial collectives during mucosal repair and gut ecosystem engineering.

Figures (5)

• Figure 1: From swarming dynamics to predictive guidance. In a mouse model of colitis, Enterobacter sp. SM3 swarms along the inflamed mucosal surface. Under pharmacological, environmental, or microbiota-based modulation, forecasting the future position of the active swarming front may enable spatially targeted and temporally coordinated interventions. Such predictive capability could be coupled with experimental strategies that locally modulate oxygen availability or microbial competition at the advancing front, with potential effects on anaerobic niche formation and downstream community composition, including Muribaculaceae expansion. Time-lapse swarming assays are converted into boundary-resolved morphological masks by TexPol--Net, with Texture-Edge and Polar--Context attention, shared prototypes and mask assembly, defining observed morphological states. Morpher models their temporal evolution with a sequence backbone and Morphon memory to predict future morphologies for quantitative evaluation. Panels I–IV depict an application-oriented biological and experimental pipeline, whereas the modeling framework follows a two-stage formulation of boundary measurement and morphology-level forecasting.
• Figure 3: TexPol--Net improves colony-front segmentation by coupling texture-sensitive boundary encoding with a geometry-aligned context prior.a, Texture-Polar Network (TexPol--Net) within a prototype-based instance segmentation pipeline. Texture-Edge Attention (TEA) is embedded in the backbone to preserve high-frequency boundary cues, while Polar--Context Attention (PCA) is interleaved in the PANet-style bidirectional neck to maintain polar consistency during multi-scale fusion. Dense heads at $P3$--$P5$ predict class scores, boxes and mask coefficients; a lightweight Protonet generates $k{=}32$ shared prototypes that are linearly combined into instance masks and post-processed by cropping and thresholding. b, Qualitative comparison on two representative swarming regimes. In the dense--finger phase (top), YOLOv11 khanam2024yolov11 captures the coarse outline but smooths the corrugated front, shortening narrow fingers and reducing anisotropy. SAM kirillov2023segmentanything and SAM2 ravi2024sam2 favor the colony core, producing a boundary that approaches a near-circular shape. TexPol--Net preserves finger heterogeneity and radial bias, with YOLOv12 tian2025yolov12 as the closest competing backbone. In the near--concentric ring regime (bottom), all methods recover the circular structure, while TexPol--Net more precisely localizes the advancing boundary. c, Average precision as a function of IoU matching threshold, highlighting boundary-stringent performance at high IoU. d, Image-wise overlap statistics for IoU and Dice, summarizing accuracy and variability across images.
• Figure 4: Texture-Edge Attention (TEA) and Polar--Context Attention (PCA) modules.a, The TEA block enhances fine-scale texture fidelity and boundary sharpness through three cooperative branches: a local depthwise path for intra-channel spatial preservation, multi-dilated convolutions for scale-robust texture encoding, and an edge-sensitive Laplacian path that injects a high-pass prior. Channel and spatial gating further refine feature fusion, producing an edge-aware, redundancy-suppressed representation. b, The PCA block embeds a polar-aware geometric prior aligned with the radial growth of swarming colonies. Input features are first compressed and then processed by a local branch, a large-kernel Cartesian branch, and a polar-warped branch operating in $(\rho,\theta)$ coordinates. Depthwise dilated filters extract context along radial and angular axes, and subsequent channel-- and spatial--attention gates yield a geometry-aligned output.
• Figure :
• Figure :