henhaobio
Sign in

PLoS Computational BiologyAnimal Behavior & CognitionPeer-reviewed journal article

An improved neural network model enables worm tracking in challenging conditions and increases signal-to-noise ratio in phenotypic screens

Weheliye H. Weheliye, Javier RodrígueziD, Luigi FerianiiD, Avelino JaveriD, Virginie UhlmanniD, André EX BrowniD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

High-resolution posture tracking of C. elegans has applications in genetics, neuroscience, and drug screening. While classic methods can reliably track isolated worms on uniform backgrounds, they fail when worms overlap, coil, or move in complex environments. Model-based tracking and deep learning approaches have addressed these issues to an extent, but there is still significant room for improvement in tracking crawling worms. Here we train a version of the DeepTangle algorithm developed for swimming worms using a combination of data derived from Tierpsy tracker and hand-annotated data for more difficult cases. DeepTangleCrawl (DTC) outperforms existing methods, reducing failure rates and producing more continuous, gap-free worm trajectories that are less likely to be interrupted by collisions between worms or self-intersecting postures (coils). We show that DTC enables the analysis of previously inaccessible behaviours and increases the signal-to-noise ratio in phenotypic screens, even for data that was specifically collected to be compatible with legacy trackers including low worm density and thin bacterial lawns. DTC broadens the applicability of high-throughput worm imaging to more complex behaviours that involve worm-worm interactions and more naturalistic environments including thicker bacterial lawns.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (4 Feb 2025) · CC BY 4.0

DeepTangleCrawl, DeepTangle retrained for crawling C. elegans, skeletonises more coiled and overlapping worms than Tierpsy, Omnipose or a PAF tracker, with longer tracks and larger screen effect sizes than Tierpsy.

Classic skeletonisation trackers such as Tierpsy lose crawling worms when they coil, overlap or sit on cluttered backgrounds, and the earlier DeepTangle model was built for swimming worms, whose postures and overlaps are simpler. Retraining it on real crawling-worm clips (Tierpsy-derived, synthetically overlapped and hand-annotated) carries its use of neighbouring frames over to these harder cases, recovering behaviour that legacy tracking leaves as gaps.

Key results

  1. ShownOn held-out data, DeepTangleCrawl (DTC) fails to output a skeleton less often than Tierpsy, Omnipose or a part affinity field (PAF) landmark tracker, and its error is smaller than the disagreement between human annotators.

    Median root mean squared deviation (RMSD) to annotated or Tierpsy skeletons is 2.2 pixels (about half a body width) on 2000 held-out test skeletons. All models have a lower modal RMSD than the >4 pixel RMSD among three annotators. Omnipose and PAF have a lower modal RMSD than DTC when they do predict. DTC has the fewest no-prediction cases (counts only in Figure 2B). · Pose estimation accuracy; Training data; RMSD calculation and failed predictions; Figure legends

  2. ShownDTC produces longer continuous trajectories and more skeletons per frame than Tierpsy at both 3 and 15 worms per well. In harder conditions (dense worms, eggs, collisions, coiling) it keeps skeletonising worms where Tierpsy fails, and it keeps worm identities through collisions.

    Trajectory duration (log scale) and skeletons per frame both rise from Tierpsy to DTC at 3 and at 15 worms per well (48 and 240 worms per video). DTC's skeleton counts come closer to the nominal worm number. The overlap, collision and coil examples are qualitative (Supplementary Movies 1 to 3). · DTC improves tracking in challenging conditions; Figure legends

  3. ShownWhen a published insecticide and nematicide screen (recorded with Tierpsy in mind) is reanalysed, DTC and Tierpsy give broadly similar behavioural features. The features that disagree tend to be ones that are noisy anyway.

    Median midbody speed is highly correlated between the two trackers across 290 randomly sampled wells. The modal Pearson correlation across the Tierpsy256 features is 0.71. Many weakly correlated features also have low F-statistics across drug classes. · Improved signal to noise ratio in phenotypic…; Phenotypic screen analysis; Figure legends

  4. SuggestedFeatures from DTC give larger drug-versus-DMSO effect sizes than features from Tierpsy. In a focused example, curvature under the coiling-inducing spiroindoline SY1713 is higher and less variable with DTC.

    For SY1713, DTC's curvature estimate is higher at every dose and its distribution is narrower. Across all features and drugs, the mean absolute Hedges' d versus DMSO is correlated between the two trackers but larger for DTC (points below the diagonal in Figure 4D). The text reports no formal test of the difference and uses one screen. · Improved signal to noise ratio in phenotypic…; Phenotypic screen analysis; Figure legends

  5. ShownInference time for the neural models does not grow with the number of worms, while Tierpsy's does. Running DTC on every third frame and filling the gaps with a spline cuts time at a small cost in accuracy.

    Per-frame times are compared at 3 and 15 worms per well (Figure 2D). DTC and PAF run on GPU, Tierpsy on CPU only, and Omnipose on GPU plus CPU skeletonisation. DTC on 1:3 subsampled frames plus a 3D smoothing spline lowers total time with slightly worse RMSD, but the gain is modest at 15 worms per well. The text gives no absolute times. · Pose estimation accuracy; Fitting 3D splines; Figure legends

Methods, data and reproducibility

DeepTangle, a YOLO-style network that predicts each skeleton as 72 principal-component shape coefficients from 11-frame clips, was retrained in JAX on Tierpsy-derived crawling-worm clips, synthetic overlaps and hand-annotated hard cases, after SVD background subtraction. It was compared with Tierpsy, Omnipose and a PAF landmark tracker, then used to reanalyse a published drug screen.

  • Tierpsy-tracked clips of isolated, non-coiling crawling worms (11 frames each, sampled from 2, 4 and 8 s sequences); pairs were stacked each epoch to simulate overlaps · C. elegans · n = 50,000 clips (25,000 effective after stacking)
  • Manually annotated clips of hard cases (coils, complex overlaps, non-uniform backgrounds), made with a custom annotation GUI · C. elegans · n = 1000 clips
  • Held-out test skeletons drawn from the same experiments as the training data · C. elegans · n = 2000 skeletons · 10.5281/zenodo.14517883
  • Previously published video screen of worms treated with diverse insecticides and nematicides at several doses with DMSO controls, reanalysed with both Tierpsy and DTC · C. elegans · n = 290 wells sampled for speed comparison; full screen size not stated
  • Source videos from megapixel camera arrays (12.4 µm/pixel, 25 frames/s; 16 wells of a 96-well plate per camera) from disease-model, drug, insecticide and nematicide experiments · C. elegans · n = not stated (tracking and timing comparisons use 3 or 15 worms per well)
Methods & tools
DeepTangle, DeepTangleCrawl (DTC), Tierpsy tracker, Tierpsy256 features, Omnipose, Part Affinity Fields (PAF), JAX, PyTorch, ResNet, UNET, YOLO, Adam optimizer
Assay conditions & compute
DTC and Omnipose were each trained on an NVIDIA RTX A6000 (48 GB). At inference, DTC and PAF run on GPU, Tierpsy on CPU only, and Omnipose on GPU plus CPU skeletonisation. Training durations and absolute inference times are not given in the text.
Reproducibility gaps
No link to DTC weights or training code is printed. Zenodo holds the 2000 test skeletons and, per the Omnipose section, training data, a model and inference code. Release of the full DTC training set and software versions is not stated.
Limitations (6)
  • DTC still fails on tightly coiled worms that barely move and on worms in prolonged overlaps or parallel contact, because without motion it cannot use neighbouring frames. Every tested model fails on complex overlaps of several worms. · stated by the authors
  • Where Omnipose and PAF do output a skeleton, they are more accurate than the current DTC (lower modal RMSD). · stated by the authors
  • Training data come from one imaging setup and lab, and the authors expect that much more training data would be needed for the model to generalise to recordings from other trackers. · stated by the authors
  • Long-lived close contacts such as mating remain hard for DTC, and the authors do not expect large dense aggregates to be trackable from brightfield imaging alone. · stated by the authors
  • The authors retrained Omnipose and PAF on their own data, and PAF is an in-house landmark tracker resembling SLEAP rather than SLEAP itself, so the authors designed and ran the whole comparison. · noted in this summary
  • Test skeletons come from the same experiments as the training clips, and the effect-size gain rests on one previously published screen with no formal test of the difference. · noted in this summary

Read this if you run high-throughput C. elegans posture tracking or behavioural drug screens and lose data to coiling, overlapping worms or thicker bacterial lawns.

Citations
12
FWCI
4.15
Field percentile
top 4%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBanerjee '22Saha '26Kiel '18Kuze '25Hebert '20Safaeian '26
76% semantic matchAnimal Behavior & Cognition
S. C. Banerjee, K. A. Khan, R. Sharma·bioRxiv·Aug 18, 2022