bio.rodeo
HomeCompetitorsLeaderboardOrganizations
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

© 2026 bio.rodeo. All rights reserved.
Imaging

BPD

Chan Zuckerberg Initiative

Fifth-place solution from the CZII CryoET Kaggle competition; an ensemble of four lightweight 3D U-Nets for protein particle localization in cryo-ET tomograms.

Released: 2025

Overview

BPD — Biological Particle Detector — is a deep learning model for localizing protein complexes in cryo-electron tomography (cryo-ET) tomograms, developed as a competitive solution to the CZII CryoET Object Identification Kaggle challenge organized by the Chan Zuckerberg Imaging Institute (CZII) from November 2024 to February 2025. BPD placed fifth among 1,135 participants and is now hosted on the CZ Virtual Cells Platform (v1.0) as a publicly accessible particle picking tool integrated with the copick cryo-ET dataset ecosystem.

The CZII Kaggle competition challenged participants to automate the detection of six protein complexes — apoferritin, beta-amylase, beta-galactosidase, 80S ribosomes, thyroglobulin, and virus-like particles — within experimental and simulated cryo-ET tomographic volumes. BPD approached this as a 3D segmentation problem using a compact U-Net architecture trained with multiple random seeds and combined through ensemble averaging, a strategy that improves prediction robustness without requiring architectural novelty. The result is a lightweight, easy-to-train model that achieves competitive accuracy with substantially fewer parameters than deeper architectures, making it practical for laboratories with limited GPU compute.

By making fifth-place competition solutions available alongside first-place models on the Virtual Cells Platform, CZII provides users with a range of trade-offs between performance and computational cost, broadening accessibility to the cryo-ET particle picking community.

Key Features

  • Lightweight 3D U-Net: Uses a compact architecture with only two downsampling and two upsampling levels, processing features through 28, 32, and 36 channels per resolution, making it significantly lighter than deeper ensembles while retaining multi-scale context.
  • Trilinear interpolation upsampling: Employs trilinear interpolation for upsampling in the decoder rather than transposed convolutions, reducing artifacts and parameter count in the upsampling path.
  • Four-seed ensemble: Trains four identical 3D U-Net models with different random seeds and averages their predictions, providing variance reduction that improves recall and precision beyond any single model without requiring architectural diversity.
  • Heatmap-based localization: Predicts a continuous 3D probability density for each particle class rather than hard coordinate proposals, with local maxima extraction used to recover particle center positions at inference time.
  • Multi-class simultaneous detection: A single forward pass produces heatmaps for all six target molecular species, making inference efficient on large tomographic volumes.
  • Copick compatibility: Integrated with the copick storage-agnostic API, enabling direct application to datasets registered in the CZ cryo-ET Data Portal or any local copick-format project.

Technical Details

The BPD architecture is a 3D U-Net with a deliberately shallow depth of two encoder-decoder levels. The encoder performs two rounds of 3D convolution followed by stride-2 downsampling, progressing through feature channel counts of 28, 32, and 36. Two dual-convolution blocks at each resolution extract local features before downsampling. The decoder mirrors this structure with two upsampling stages using trilinear interpolation (bilinear analog in 3D) followed by 3D convolution, with skip connections from the corresponding encoder levels. The final output head produces a multi-channel heatmap volume with one channel per target particle class.

Training uses the experimental and simulated tomograms from the CZII phantom dataset: 6 tomograms from the phantom dataset (with 1 held out for validation) and 27 simulated runs from the competition dataset. Each training example is a 3D patch cropped from a full tomogram volume, with data augmentation including random flips, rotations, and noise injection to improve generalization across the variable signal-to-noise conditions present in experimental cryo-ET data. Four models are trained from scratch using the same architecture and data but different random seeds, and their heatmap outputs are averaged before coordinate extraction. The ensemble strategy targets variance reduction: individual models trained on small datasets exhibit high variance in their predictions, and averaging across seeds reduces this without requiring additional data or larger architectures.

Applications

BPD is well suited for cryo-ET practitioners who need an accessible particle picking baseline that can be trained and run on modest GPU hardware. Its shallow architecture means that training converges rapidly and can be completed on a single GPU in hours rather than days, making it practical for iterative experimental workflows where picking models may need to be retrained as new tomograms are collected. The four-seed ensemble strategy is straightforwardly reproducible: users can train BPD on a new dataset simply by running the training script four times with different seeds and averaging outputs. The copick integration means that picked coordinates feed directly into the CZ cryo-ET Data Portal ecosystem for collaborative visualization and quality control. BPD is particularly appropriate for studies involving the six protein complexes in the CZII phantom benchmark, where pre-trained weights can be applied directly without fine-tuning.

Impact

BPD's fifth-place performance in a 1,135-participant competition demonstrates that lightweight ensemble approaches can achieve competitive particle picking accuracy without architectural complexity, providing a useful counterpoint to deeper and more parameter-heavy first-place solutions such as TopCUP. Its inclusion on the CZ Virtual Cells Platform alongside higher-ranked models gives users an explicit choice between computational overhead and accuracy, recognizing that many cryo-ET laboratories operate on resource-constrained computing environments. As of early 2026, BPD has not been the subject of a standalone peer-reviewed publication; its performance was documented in the context of the CZII competition. Users applying BPD to targets outside the six phantom complexes should expect to perform task-specific fine-tuning to achieve reliable results.

Tags

particle pickingsegmentationCNNtransfer learningcryo-ETstructural biology

Resources

Official Website