Swarming Morphogenesis Evolution (SwarmEvo)

Forecasting the evolution of bacterial swarming colonies is a challenging task. Unlike conventional natural scenes, swarming expansion is governed by the propagation of irregular, anisotropic morphological fronts rather than by appearance continuity or pixel-level motion. These fronts undergo rapid reorganization, fingering instabilities, and curvature-driven growth, making standard image- or video-based prediction models fundamentally inadequate. Addressing this problem requires a dataset that resolves colony geometry with high fidelity and preserves its temporal evolution at the level of boundary morphology, enabling predictive rather than purely descriptive analysis.

Swarming Morphogenesis Evolution (SwarmEvo) is a curated dataset designed for studying the spatial segmentation and temporal prediction of bacterial swarming colony morphogenesis.
All data are derived from in-house experimental cultivation of Enterobacter sp. SM3 and are annotated and organized to support both instance-level segmentation and morphological forecasting.

The dataset is used throughout the accompanying paper From Shape to Fate: Making Bacterial Swarming Expansion Predictable and supports full reproducibility of the reported experiments.

---

Dataset Overview

SwarmEvo consists of two complementary components:

• Segmentation Dataset – instance-level colony segmentation on individual frames
• Prediction Dataset – grouped time-resolved image sequences for temporal forecasting

All images are experimentally acquired under controlled conditions and share a fixed spatial resolution of 1250 × 1250 pixels.

---

Data Acquisition and Experimental Origin

• All images in SwarmEvo are generated from colonies cultivated by the authors.
• Imaging is performed using time-lapse microscopy under controlled laboratory conditions.
• No external datasets or synthetic data are used.
• The dataset captures a diverse range of swarming morphologies and expansion regimes across independent experiments.

Cultivation, swarming activation, and imaging conditions

All SwarmEvo sequences were generated from in-house cultivation of Enterobacter sp. SM3 under controlled conditions. A single colony was transferred from an LB agar plate into LB broth (10 g/L tryptone, 5 g/L yeast extract, and 5 g/L NaCl) and incubated at 37 °C with shaking at 200 rpm for approximately 16 h to obtain a high-density culture. A 5–8 µL aliquot of the overnight culture was then spotted onto the center of a freshly prepared swarming agar plate (LB medium supplemented with 0.5–0.8% agar), poured to a uniform thickness of 3–4 mm to promote consistent hydration and nutrient diffusion across plates. The inoculated plates were incubated at 30 °C and approximately 90% relative humidity for 4–6 h to activate swarming motility, and then transferred to a time-lapse imaging chamber maintained between 28–37 °C and 80–92% relative humidity. Humidity was kept below the condensation threshold to prevent droplet formation on the agar surface. Surface stiffness and inoculum volume were varied across experiments to introduce controlled physical differences and generate diverse colony expansion conditions.

Imaging was performed using a vertically mounted high-resolution digital camera with uniform LED illumination to minimize specular reflection and shadowing artifacts. Frames were recorded at one-minute intervals throughout colony expansion and continued until the colony front reached the plate boundary or no further measurable growth was observed. Sequences were stored at native spatial resolution with associated timestamps. For downstream analysis, each sequence was segmented using TexPol–Net to obtain pixel-level colony masks. These mask sequences formed a morphology-consistent representation used for training and evaluating the forecasting models. Training and validation datasets were assembled from non-overlapping experimental runs collected under distinct growth conditions to assess generalization across colony expansion regimes.

Further experimental details, imaging conditions, and biological protocols are described in the Methods section of the accompanying paper.

---

1. Segmentation Dataset

Purpose

The segmentation subset is used to train and evaluate instance segmentation models for accurate delineation of swarming colony boundaries, which subsequently serve as inputs to the prediction model.

Data Statistics

Split	Number of Samples	Augmentation
Training	1815	Yes
Validation	78	No
Test	78	No
Total	1971	—

• All 1971 samples originate from experimentally acquired images.
• Data augmentation is applied only to the training set.
• Validation and test sets remain unaltered to ensure unbiased evaluation.

Annotations

• Each image is annotated with instance-level segmentation masks.
• Annotations are curated to ensure:
  • Closed and valid instance contours
  • Consistent boundary definition across samples
• Augmentations preserve annotation fidelity through synchronized geometric transformations.

---

2. Prediction Dataset

Purpose

The prediction subset is designed for temporal modeling of swarming morphogenesis, enabling the prediction of future colony shapes from past observations.

Data Organization

• Each prediction sample is a grouped sequence of temporally consecutive images.
• Sequences represent the evolution of a single colony over time.
• Prediction is performed at the sequence (group) level, not the individual frame level.

Data Statistics

Split	Number of Sequences	Augmentation
Training	260	Yes
Test	16	No
Total	276	—

Temporal Resolution

• All sequences are sampled at a fixed temporal interval of 1 minute.
• Temporal downsampling, when used, is performed uniformly across sequences.

Augmentation Strategy

• Sequence-level augmentation is applied only to training sequences.
• All frames within a sequence share identical transformation parameters to preserve temporal coherence.
• Augmentations are restricted to transformations that do not alter growth dynamics, such as:
  • In-plane rotation
  • Translation
  • Horizontal and vertical flipping

---

Preprocessing and Standardization

• Original image resolution: 1250 × 1250
• During model training and evaluation, images or masks may be resized to match network input requirements.
• Binary masks used for prediction are derived from segmentation outputs and remain consistent within each experimental setup.

---

Recommended Usage

• Segmentation dataset
  Train and evaluate instance segmentation models for accurate colony boundary extraction.
  Segmentation data are organized under SwarmEvo/segmentation, where all annotations are provided in YOLO-format mask files corresponding to each image.

• Prediction dataset
  Use segmented masks as inputs for temporal models to predict future colony morphology under autoregressive settings.
  Prediction data are located under SwarmEvo/prediction. The subdirectory SwarmEvo/prediction/raw_imgs contains the original image sequences for all colonies, while the corresponding segmentation masks are provided in YOLO format, enabling direct reuse in downstream temporal modeling pipelines.

To avoid data leakage, training, validation, and test splits are strictly separated at the sequence level for temporal prediction, while segmentation samples may originate from the same colony but are temporally well separated, ensuring that no contiguous or dynamically correlated observations are shared across splits.

---

Citation

If you use the SwarmEvo in your research, please cite the accompanying paper:

From shape to fate: making bacterial swarming expansion predictable

@article{duan2026shapetofate,
  title     = {From shape to fate: making bacterial swarming expansion predictable},
  author    = {Duan, Shengyou and Wang, Zhaoyang and Xiong, Kaiyi and Zhu, Jin and Gu, Pengxi and Chen, Weijie and Xin, Hongyi and Qu, Zijie},
  journal   = {arXiv preprint arXiv:2602.01056},
  year      = {2026},
  url       = {https://arxiv.org/abs/2602.01056}
}

---

License

The SwarmEvo dataset is released for academic research use only.

---

Contact

For questions regarding the dataset, please contact the corresponding authors listed in the paper.