Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Color Flow Imaging Microscopy Improves Identification of Stress Sources of Protein Aggregates in Biopharmaceuticals

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Color Flow Imaging Microscopy improves stress-source classification of protein aggregates, with mixed-color training reaching 97.1% accuracy.

desk verdict A well-run empirical comparison of color vs monochrome FIM images with a new dataset, but the color advantage may be a buffer/background shortcut, and validation-set reuse weakens the quantitative claims. read the letter →

arxiv 2501.15492 v1 pith:GT6B4LGL submitted 2025-01-26 cs.CV cs.AI

Michaela Cohrs , Shiwoo Koak , Yejin Lee , Yu Jin Sung , Wesley De Neve , Hristo L. Svilenov , Utku Ozbulak This is my paper · ORCID
classification cs.CVcs.AI
keywords flowimagingmicroscopyproteinaggregationsubvisibleparticlesstresssourceclassificationcolorversusgrayscaledeeplearningbiopharmaceuticalqualitycontrolself-supervised
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Protein-based drugs can form subvisible particles when exposed to stresses such as heat or mechanical shaking, and knowing which stress caused a particle matters for quality control. This paper asks whether color Flow Imaging Microscopy (FIM), an imaging technique that photographs particles in flowing liquid, carries information that black-and-white images lose. Using 16,000 subvisible particle images from eight commercial monoclonal antibodies, the authors train ResNet-50 and ViT-B/16 models under several supervised and self-supervised pretraining schemes and find that models trained on RGB images consistently outperform those trained on grayscale or single-channel images. Adding color as a random training augmentation pushes the best model to 97.1% overall accuracy. The paper concludes that color FIM is worth using for stress-source classification of protein aggregates.

What carries the argument

The mechanism under test is the color-mode representation of FIM images. The authors convert RGB images to grayscale with the standard ITU-R 601-2 luminance formula $L = 0.299R + 0.587G + 0.114B$, which gives a monochrome baseline with identical particle morphology and distribution. They also train on individual red, green, and blue channels. The mixed-color strategy uses color mode as a form of data augmentation by randomly converting each training image to one of the color modes, forcing the model to learn features shared across color presentations. The stress-source task is binary (heat versus mechanical), and the dataset includes two antibodies stressed both ways and two antibodies withheld from training to test generalization.

What would settle it

Capture matched samples under controlled monochromatic illumination with a true monochrome sensor and with a color sensor, or train classifiers on RGB images whose color channels are randomly permuted across images; if the RGB advantage disappears when channel identity is decoupled from stress type, the color signal is an artifact of the imaging system rather than a property of the aggregates.

Watch

Extended reading notes

Core claim

The central claim is that deep-learning classifiers for stress-source identification of subvisible protein particles perform better with color FIM images than with monochrome conversions of the same particles. Across more than 800 training runs covering five color modes and seven model/pretraining combinations, RGB input is the best or tied-best in every row, with the largest RGB advantage reaching about 1.5 percentage points over the second-best mode—roughly a third fewer misclassified particles. The paper also reports that the green channel carries more discriminative signal than red or blue, that self-supervised pretraining matches or exceeds supervised pretraining, and that a mixed-color augmentation scheme, which randomly converts training images to red, green, blue, or grayscale, reaches 97.1% overall accuracy while improving per-antibody true positive rates.

Load-bearing premise

The load-bearing premise is that the color differences in the FIM images come from the protein aggregates themselves, not from lighting, focus, or flow-cell optics that happen to differ between heat- and mechanically-stressed samples; if color is an imaging artifact, the RGB advantage would not reflect a real benefit of color FIM.

Editorial extensions

If this is right

  • Color FIM offers a concrete accuracy gain for stress-source classification, so manufacturers using only monochrome instruments can expect measurably fewer misclassified particles by upgrading to color imaging.
  • The green channel's higher discriminative value suggests that wavelength-specific information about protein aggregates is present and may be exploitable beyond simple RGB images.
  • Self-supervised pretraining is a viable alternative to supervised pretraining for this domain, which matters because labeled subvisible particle images are scarce.
  • Mixed-color augmentation yields the best overall accuracy (97.1%) and improves per-antibody consistency, making it a practical training recipe for similar quality-control tasks.
  • Models generalize to antibodies never seen during training, supporting the use of such classifiers for new drug products without recollecting stress data for every molecule.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the green-channel advantage raises a testable hypothesis that aggregate optical properties near green wavelengths correlate with stress type; a band-pass illumination or spectroscopy experiment could check this independently of FIM.
  • Editorial inference: mixed-color augmentation may act as domain randomization that helps transfer models across FIM devices with different color calibration, but this cross-instrument transfer is not tested in the paper.
  • Editorial inference: because the grayscale baseline is derived from RGB rather than captured by a true monochrome sensor, the paper leaves open whether a dedicated monochrome camera would show the same performance gap; a hardware-level comparison would settle that.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper curates a new dataset of 16,000 Flow Imaging Microscopy (FIM) images of subvisible protein particles from eight commercial monoclonal antibodies subjected to heat or mechanical stress, and trains ResNet-50 and ViT-B/16 models under supervised and self-supervised pretraining to classify the stress source. Using RGB images, individual red/green/blue channels, a luminance-based grayscale conversion (Eq. 1), and a mixed-color augmentation scheme, the authors report that RGB-trained models consistently outperform grayscale-trained models, with the best RGB models at about 97.0% accuracy and a mixed-color model reaching 97.1%.

Significance. If the central claim is validated, this would be the first systematic demonstration that color FIM provides a practically meaningful accuracy gain over monochrome FIM for stress-source classification in biopharmaceutical quality control. The paper's strengths include a new dataset spanning eight commercial antibodies, a large set of 800+ training runs across two architectures and multiple pretraining methods, and a grayscale baseline derived from the same RGB images, which controls for particle morphology. The mixed-color augmentation idea is also interesting. However, the current evaluation protocol and the absence of controls for color artifacts leave the central claim not yet established; the reported advantage could stem from buffer/background color cues or from validation-set model selection rather than from intrinsic color information carried by the aggregates.

major comments (3)
  1. [§2.5, §3, Table 2] The same validation set is used for early stopping, hyperparameter selection over the 24 training runs, and the final accuracy reporting, with no independent test set and no repeated-seed error bars. The 'unseen antibody' generalization claim is also weakened because mAb5 and mAb8 appear in the validation set that drives model selection; they are unseen only in the sense of not being in the training set. The reported 1–1.5% RGB advantage may therefore reflect selection noise. Please add a genuinely held-out test set (or nested cross-validation) and report confidence intervals or multiple-seed statistics.
  2. [§2.1, §2.2, §2.5, Table 2] The central RGB-versus-grayscale comparison is not controlled for where the color signal lives. Heat-stress and mechanical-stress samples are prepared with different buffers and workflows (Section 2.1), and images are padded with the median color of the image (Section 2.2); because most pixels are background, the median color essentially tracks the buffer or background color. The random-resized-crop augmentation does not guarantee removal of background pixels, and the grayscale baseline inherits luminance background cues while losing hue. Without a control such as a true monochrome sensor, a flat-field/white-balance calibration, or a particle-masked neutral-background comparison, the RGB advantage and the green-channel result cannot be attributed to color information carried by the aggregates rather than by the imaging/protocol environment. Please add such a control or explicitly restrict the claim to the current acquisition setup.
  3. [§3, Table 3] The mixed-color training result (97.1%) is presented as a further improvement, but the same confound applies: if the color signal exploited by the RGB model is a buffer/background artifact, then mixing color modes as augmentation only augments that artifact. Moreover, the comparison between the mixed-color model and the single-color models is made on the same validation set used for early stopping and grid-search selection, so the improvement is not statistically grounded. A test-set evaluation with confidence intervals is needed before the mixed-color advantage can be considered established.
minor comments (5)
  1. [§2.5] The heading 'T raining' should be 'Training'.
  2. [§2.1] The text states '10 µM acetate buffer at pH 5'; this is likely a typo for '10 mM acetate buffer', since a 10 micromolar buffer would be neither practical nor pharmaceutically relevant. Please correct and verify.
  3. [§3] In the mixed-color training paragraph, 'color as a an augmentation' should read 'color as an augmentation'.
  4. [Figure 1] The two panels use different axis limits, which makes the aspect-ratio distributions difficult to compare visually; please use aligned axes for both panels.
  5. [§3, Tables 2 and 3] The selected hyperparameters for the reported best-performing models are not listed; for reproducibility, please provide the final learning rate, weight decay, momentum, and pretraining checkpoint for each row of Tables 2 and 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the color-versus-monochrome comparison is a controlled empirical evaluation with no fitted constant disguised as a prediction.

full rationale

The paper's central claim is an empirical comparison of deep learning models trained on RGB, individual-channel, and grayscale versions of the same Flow Imaging Microscopy images. No derivation step reduces to its own input: the grayscale version is obtained deterministically from the RGB images via the standard ITU-R 601-2 LUMA formula (Eq. 1), so the comparison holds morphology and particle distribution fixed and the observed accuracy differences are measured on a held-out validation set rather than being fitted. Mixed-color training is described as a data-augmentation scheme and its 97.1% accuracy is again reported on validation data; no target result is used to select the reported numbers in a way that would force the conclusion. The only self-citation (reference [22]) is a survey of self-supervised learning cited for context in the methodology section; it is not load-bearing for the color-comparison claim. The paper does not invoke a uniqueness theorem, smuggle in an ansatz via citation, or rename a known result. The skeptical concern that color differences might originate from buffer, padding, or optics rather than from the aggregates is a question about external validity and experimental control, not circular reasoning: the comparison is internally consistent, so this does not affect the circularity score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This is an empirical study; it introduces no new physical entities or fitted constants. The claims rest on domain assumptions about the visual distinguishability of stress types and the faithfulness of the grayscale conversion, which are listed.

assumptions (3)
  • domain assumption Heat and mechanical stress produce visually distinguishable subvisible particle morphologies that generalize across the eight mAbs tested.
    The classification task assumes that stress type is visually identifiable from FIM images, but the paper does not independently verify that the visual differences are due to stress type rather than antibody identity or imaging conditions.
  • domain assumption The ITU-R 601-2 LUMA grayscale conversion (Eq. 1) produces images equivalent to those from a true monochrome FIM instrument.
    Real monochrome FIM sensors may have different spectral sensitivity, so the comparison of RGB against this conversion may not represent a comparison against an actual monochrome device.
  • domain assumption Color channel values in FIM images reflect particle properties and not imaging-system artifacts.
    No control experiment is provided to rule out color artifacts correlated with stress type, so the RGB advantage could be an instrument effect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Color Flow Imaging Microscopy Improves Identification of Stress Sources of Protein Aggregates in Biopharmaceuticals." pith.science (2026). https://pith.science/paper/GT6B4LGL

@misc{pith2026250115492,
  author       = {Pith},
  title        = {Pith review of: Color Flow Imaging Microscopy Improves Identification of Stress Sources of Protein Aggregates in Biopharmaceuticals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GT6B4LGL}},
  note         = {Machine review of arXiv:2501.15492}
}
read the original abstract

Protein-based therapeutics play a pivotal role in modern medicine targeting various diseases. Despite their therapeutic importance, these products can aggregate and form subvisible particles (SvPs), which can compromise their efficacy and trigger immunological responses, emphasizing the critical need for robust monitoring techniques. Flow Imaging Microscopy (FIM) has been a significant advancement in detecting SvPs, evolving from monochrome to more recently incorporating color imaging. Complementing SvP images obtained via FIM, deep learning techniques have recently been employed successfully for stress source identification of monochrome SvPs. In this study, we explore the potential of color FIM to enhance the characterization of stress sources in SvPs. To achieve this, we curate a new dataset comprising 16,000 SvPs from eight commercial monoclonal antibodies subjected to heat and mechanical stress. Using both supervised and self-supervised convolutional neural networks, as well as vision transformers in large-scale experiments, we demonstrate that deep learning with color FIM images consistently outperforms monochrome images, thus highlighting the potential of color FIM in stress source classification compared to its monochrome counterparts.

Figures

Figures reproduced from arXiv: 2501.15492 by the authors.

Figure 1
Figure 1. Distribution of the height and width of the SvP images used in this study [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example set of subvisible protein aggregate images is shown in (a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages

  1. [1]

    Journal of Pharmaceutical Sciences 89(5), 646–651 (2000)

    Arakawa, T., Kita, Y.: Protection of bovine serum albumin from aggregation by tween 80. Journal of Pharmaceutical Sciences 89(5), 646–651 (2000)

    work page 2000

  2. [2]

    arXiv preprint arXiv:2105.04906 (2021)

    Bardes, A., Ponce, J., LeCun, Y.: VICReg: Variance-invariance-covariance regu- larization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021)

    arXiv 2021

  3. [3]

    Advances in Neural Information Processing Systems33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems33, 1877–1901 (2020)

    2020

  4. [4]

    Advances in Neural Information Processing Systems 33, 9912–9924 (2020)

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems 33, 9912–9924 (2020)

    work page 2020

  5. [5]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Caron, M., Touvron, H., Misra, I., J´ egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650–9660 (2021)

    2021

  6. [6]

    In: International Conference on Machine Learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597–1607. PMLR (2020)

    2020

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15750–15758 (2021) 10 Cohrs et al

    work page 2021

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 9640–9649 (2021)

    2021

Show all 32 references
  1. [9]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    He, K., Chen, X., Xie, S., Li, Y., Doll´ ar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 16000–16009 (2022)

    2022

  3. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729–9738 (2020)

    2020

  4. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016)

    2016

  5. [13]

    Kanan, C., Cottrell, G.W.: Color-to-grayscale: does the method matter in image recognition? PloS one 7(1), e29740 (2012)

    2012

  6. [14]

    In: 2022 2nd International Conference on Artificial Intelligence (ICAI)

    Khan, A., AlBarri, S., Manzoor, M.A.: Contrastive self-supervised learning: a sur- vey on different architectures. In: 2022 2nd International Conference on Artificial Intelligence (ICAI). pp. 1–6. IEEE (2022)

    2022

  7. [15]

    Journal of pharmaceutical sciences 97(10), 4347–4366 (2008)

    Kiese, S., Papppenberger, A., Friess, W., Mahler, H.C.: Shaken, not stirred: me- chanical stress testing of an igg1 antibody. Journal of pharmaceutical sciences 97(10), 4347–4366 (2008)

    2008

  8. [16]

    In: MAbs

    Lai, P.K., Gallegos, A., Mody, N., Sathish, H.A., Trout, B.L.: Machine learning prediction of antibody aggregation and viscosity for high concentration formula- tion development of protein therapeutics. In: MAbs. vol. 14, p. 2026208. Taylor & Francis (2022)

    2022

  9. [17]

    Proceedings of the IEEE (1998)

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE (1998)

    1998

  10. [18]

    Journal of Pharmaceutical Sciences 108(7), 2397–2405 (2019)

    Matter, A., Koulov, A., Singh, S., Mahler, H.C., Reinisch, H., Langer, C., Zu- col, B., Mathaes, R.: Variance between different light obscuration and flow imag- ing microscopy instruments and the impact of instrument calibration. Journal of Pharmaceutical Sciences 108(7), 2397...

    2019

  11. [19]

    Journal of pharmaceutical sciences 105(2), 417–430 (2016)

    Moussa, E.M., Panchal, J.P., Moorthy, B.S., Blum, J.S., Joubert, M.K., Narhi, L.O., Topp, E.M.: Immunogenicity of therapeutic protein aggregates. Journal of pharmaceutical sciences 105(2), 417–430 (2016)

    2016

  12. [20]

    Nature reviews

    Mullard, A.: Fda approves 100th monoclonal antibody product. Nature reviews. Drug discovery 20(7), 491–495 (2021)

    2021

  13. [21]

    Journal of pharmaceutical sciences 104(6), 1899–1908 (2015)

    Narhi, L.O., Corvari, V., Ripple, D.C., Afonina, N., Cecchini, I., Defelippis, M.R., Garidel, P., Herre, A., Koulov, A.V., Lubiniecki, T., et al.: Subvisible (2-100 µm) particle analysis during biotherapeutic drug product development: part 1, con- siderations and strategy. Jou...

    2015

  14. [22]

    arXiv preprint arXiv:2305.13689 (2023)

    Ozbulak, U., Lee, H.J., Boga, B., Anzaku, E.T., Park, H., Van Messem, A., De Neve, W., Vankerschaver, J.: Know your self-supervised learning: A survey on image-based generative and discriminative training. arXiv preprint arXiv:2305.13689 (2023)

    2023 arXiv

  15. [23]

    Pharmaceutical Research 40(3), 721–733 (2023) Color Flow Imaging Microscopy Improves Stress-type Detection of SvPs 11

    Poozesh, S., Cannav` o, F., Manikwar, P.: Sensitivity and uncertainty analysis of micro-flow imaging for sub-visible particle measurements using artificial neural network. Pharmaceutical Research 40(3), 721–733 (2023) Color Flow Imaging Microscopy Improves Stress-type Detectio...

    2023

  16. [24]

    Journal of pharmaceutical sciences 101(10), 3568–3579 (2012)

    Ripple, D.C., Dimitrova, M.N.: Protein particles: What we know and what we do not know. Journal of pharmaceutical sciences 101(10), 3568–3579 (2012)

    2012

  17. [25]

    International Journal of Computer Vision 115(3), 211–252 (2015)

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211–252 (2015)

    2015

  18. [26]

    The AAPS journal 12(3), 455–464 (2010)

    Sharma, D.K., King, D., Oma, P., Merchant, C.: Micro-flow imaging: flow mi- croscopy applied to sub-visible particulate analysis in protein formulations. The AAPS journal 12(3), 455–464 (2010)

    2010

  19. [27]

    Journal of pharma- ceutical sciences 111(10), 2730–2744 (2022)

    Thite, N.G., Ghazvini, S., Wallace, N., Feldman, N., Calderon, C.P., Randolph, T.W.: Machine learning analysis provides insight into mechanisms of protein par- ticle formation inside containers during mechanical agitation. Journal of pharma- ceutical sciences 111(10), 2730–2744 (2022)

    2022

  20. [28]

    International Journal of Pharmaceutics 631, 122490 (2023)

    Thorlaksen, C., Schultz, H.S., Gammelgaard, S.K., Jiskoot, W., Hatzakis, N.S., Nielsen, F.S., Solberg, H., Foder` a, V., Bartholdy, C., Groenning, M.: In vitro and in vivo immunogenicity assessment of protein aggregate characteristics. International Journal of Pharmaceutics 63...

    2023

  21. [29]

    Advances in neural information pro- cessing systems 30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems 30 (2017)

    2017

  22. [30]

    Nature Biotechnology 40(12), 1722–1760 (2022)

    Walsh, G., Walsh, E.: Biopharmaceutical benchmarks 2022. Nature Biotechnology 40(12), 1722–1760 (2022)

    2022

  23. [31]

    Interna- tional journal of pharmaceutics 289(1-2), 1–30 (2005)

    Wang, W.: Protein aggregation and its inhibition in biopharmaceutics. Interna- tional journal of pharmaceutics 289(1-2), 1–30 (2005)

    2005

  24. [32]

    Journal of pharmaceutical sciences 101(3), 914–935 (2012)

    Z¨ olls, S., Tantipolphan, R., Wiggenhorn, M., Winter, G., Jiskoot, W., Friess, W., Hawe, A.: Particles in therapeutic protein formulations, part 1: Overview of ana- lytical methods. Journal of pharmaceutical sciences 101(3), 914–935 (2012)

    2012

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.