CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis

Cedric Caruzzo; Jong Chul Ye

REVIEW 4 major objections 5 minor 24 references

CellPainTR claims that a Transformer trained on Cell Painting feature vectors, conditioned on source-specific context tokens, learns representations that correct batch effects on datasets it has never seen, and reports out-of-distribution r

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

CellPainTR applies Hyena-based Transformers with source context tokens to Cell Painting profiles, achieving better batch integration and OOD generalization than ComBat and Harmony.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful new architecture for Cell Painting batch correction, but the OOD claim rests on a hand-picked proxy token and zero-padding that are not yet controlled for. the 4 major comments →

arxiv 2509.06986 v1 pith:VRYWF2W3 submitted 2025-09-02 cs.CV cs.AI

CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis

Cedric Caruzzo , Jong Chul Ye This is my paper

classification cs.CV cs.AI

keywords Cell Paintingbatch effect correctionout-of-distribution generalizationrepresentation learningTransformerHyena operatorcontrastive learningmorphological profiling

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper is trying to establish that batch-effect correction for Cell Painting data can be a reusable learned model rather than a one-time statistical fix. It introduces CellPainTR, a Transformer that reads thousands of morphological features, conditions on a learnable token identifying the experimental source, and is trained in three stages: masked feature reconstruction, within-source contrastive learning, then cross-source contrastive learning. On the JUMP benchmark, this yields a better balance of batch integration and biological signal preservation than ComBat, Harmony, and Sphering. The decisive claim is out-of-distribution: applied without fine-tuning to the unseen Bray dataset, with only 275 of 4,765 features available, CellPainTR outperforms ComBat and Harmony that were re-fit directly on that dataset. If true, a lab could analyze new data through a pre-trained model and compare against a stable atlas instead of re-processing the combined dataset from scratch.

Core claim

The central discovery is that a Transformer trained on Cell Painting profiles, with a source-context token prepended to the feature sequence, learns a representation in which biological identity is separated from technical provenance well enough to transfer to a dataset from a different lab and a different CellProfiler version. The architecture uses linear adaptation of continuous feature values, feature-identity embeddings instead of positional encodings, and a learnable source token chosen from a codebook. The three-stage training—masked feature prediction, intra-source supervised contrastive learning, then inter-source supervised contrastive learning—first teaches general morphology, then

What carries the argument

The central mechanism is the source context token: a learnable embedding, one per training source, prepended to the feature sequence and optimized during inter-source contrastive training, which conditions the model on data provenance and lets it subtract source-specific batch effects at inference time. It is carried by Bidirectional Hyena layers, which give near-linear complexity on long feature sequences, and trained by the three-step curriculum: Channel-Wise Masked Morphology (CWMM) reconstruction, intra-source supervised contrastive learning, and inter-source supervised contrastive learning.

Load-bearing premise

The load-bearing premise is that a source token selected by metadata similarity from training data approximates the batch effect of the unseen dataset; if metadata similarity does not imply batch-effect similarity, or if compound labels are not comparable across datasets, the OOD gains would shrink.

What would settle it

Take a held-out JUMP source and compare OOD-quality scores using (a) the metadata-similar source token, (b) a randomly chosen token, and (c) an adversarially mismatched token; if random or mismatched tokens give the same score, the source token is not carrying the batch-effect signal. Alternatively, evaluate on a second unseen dataset whose metadata is similar to a training source but whose actual batch shift is known to differ; a performance drop would show the generalization depends on metadata similarity rather than learned batch-effect structure.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

If CellPainTR is correct, batch correction no longer requires re-analyzing the combined data: a pre-trained model can be applied to a new dataset directly.
A cumulative morphological atlas becomes feasible, where new experiments are embedded into a fixed reference space and compared against prior data without retraining.
The proxy-token strategy—selecting a training source by metadata similarity—becomes a practical way to adapt to unseen sources without fine-tuning.
The feature-space design keeps biological interpretability while still gaining the benefits of deep representation learning.
The same three-stage curriculum could be transferred to other high-dimensional biological feature sets beyond Cell Painting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the OOD evidence is consistent with the stronger claim that the learned representation approximates a canonical morphology space; testing on a second unseen dataset with genuinely different biology would separate that stronger claim from simple batch robustness.
Editorial inference: the proxy-token assumption suggests a testable design rule—if metadata-similarity scores predict OOD performance across held-out sources, then automatic source-token selection could replace manual metadata matching.
Editorial inference: because the model zero-pads missing features, it may tolerate feature-set drift across CellProfiler versions, which would make it useful for retrospectively integrating legacy datasets.
Editorial inference: the source token may implicitly encode experimental protocol rather than lab identity, in which case matching by protocol metadata rather than lab metadata would improve transfer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A useful new architecture for Cell Painting batch correction, but the OOD claim rests on a hand-picked proxy token and zero-padding that are not yet controlled for.

read the letter

Short version: CellPainTR is a legitimately new combination—Hyena layers, feature context embeddings, source codebook tokens, and a three-stage masked/contrastive curriculum—and it beats ComBat/Harmony on the held-out JUMP set. The engineering is competent and the ablation story is clear. But the headline claim of OOD generalization to Bray is softer than the abstract suggests. The OOD numbers depend on selecting Source 10 as a proxy token by metadata similarity, with no check that metadata similarity tracks batch-effect similarity, and on padding 275 observed features into 4,765 positions. That zero-padding creates a distinctive input pattern the model may have learned to treat as 'unseen domain.' The paper flags the proxy limitation in Section 5.2, which is honest, but the abstract and conclusion still call the OOD result 'unprecedented,' which overstates the evidence.

What works: the in-distribution aggregate is the best overall (0.60 vs 0.56), and the OOD numbers, even if confounded, suggest the representation transfers better than simple re-fitting. The curriculum design—from masked reconstruction to intra-source then inter-source contrastive—is a sensible way to navigate the batch-vs-biology trade-off. Operating on engineered features with interpretability as a stated goal is defensible.

Weaknesses, in order. First, no error bars or seeds, so a 0.60 vs 0.56 gain could be noise. Second, only three classical baselines appear; no deep representation baselines like a standard MLP or plain transformer, so we cannot tell how much of the gain comes from the architecture versus the curriculum. Third, the OOD bridge—proxy token plus zero-padding—is where the 'foundational model' claim hangs, and it is the least controlled part of the study. The biological mAP on Bray is 0.07 (Table 2), near floor, which makes the biological-preservation claim on OOD hard to sell. The compound-overlap question is unaddressed; if the same compounds appear in both datasets, the OOD framing weakens further.

Who is this for: people building representation models for image-based profiling and anyone doing batch-effect correction in Cell Painting. It is a useful proof-of-concept, not the foundational breakthrough it claims. It deserves a serious referee, but the referee should demand more baselines, error bars, an overlap check, and a cleaner OOD setup—either a learned adaptation mechanism or a validation that the proxy token is actually doing what is claimed.

Recommendation: send to peer review, but expect major revision. The core idea is worth engaging with.

Referee Report

4 major / 5 minor

Summary. CellPainTR is a Transformer-style encoder for Cell Painting morphological feature profiles. It uses a linear adaptor, per-feature context embeddings, a learnable source context token drawn from a codebook, and Bidirectional Hyena layers. Training proceeds in three stages: masked feature reconstruction (CWMM), intra-source supervised contrastive learning, and inter-source supervised contrastive learning. On a held-out JUMP test set, the final model reports the best aggregate score (0.60) versus ComBat (0.54), Harmony (0.56), and Sphering (0.51). For OOD generalization, the model is applied to the unseen Bray et al. dataset using a proxy source token (Source 10) selected by metadata similarity and zero-padding from 275 common features to 4,765; it reports Overall 0.40 versus 0.26 for ComBat/Harmony. The paper claims this demonstrates effective OOD generalization and positions CellPainTR as a foundational model for image-based profiling.

Significance. If the OOD result were clean, this would be a valuable step toward reusable models for image-based profiling: the architecture is plausible, the use of Hyena for long feature vectors is sensible, and the authors provide code and detailed training hyperparameters. However, the OOD evidence is not yet clean. The proxy source token injects training-derived information into the 'unseen' test, and zero-padding may be exploited as a domain marker. The in-distribution gain is modest (0.60 vs. 0.56) and no variance or seed-level results are reported. The significance is therefore contingent on additional controls that separate the contributions of the source token, the zero-padding, and the learned representation.

major comments (4)

[§4.3, Table 2; §5.2] The central OOD claim rests on using a pre-trained Source 10 token selected by 'experimental metadata similarity' to the Bray et al. dataset. No evidence is provided that metadata similarity implies batch-effect similarity, so the comparison is not a clean OOD evaluation: the model receives training-derived information (the token) while ComBat and Harmony are re-fit but do not receive an equivalent prior. The paper itself flags this limitation in §5.2, but the abstract and conclusion still claim 'effective OOD generalization' and 'unprecedented' performance. Please provide ablations: (i) evaluate with all K source tokens, (ii) with a random token, and (iii) with a token inferred from a held-out portion of Bray (if possible), and report whether the ranking over baselines is stable. Without this, the OOD claim is unsupported.
[§4.3, Appendix B.2] The 275-feature Bray input is zero-padded to 4,765 positions and processed with learnable per-feature embeddings and biases. This creates a highly distinctive input pattern; the model may learn to recognize zero-padded positions as an 'unseen domain' marker rather than learning a generalizable batch-invariance mechanism. No control isolates this effect. Please train or evaluate with random feature dropout on JUMP (e.g., zero out the same fraction of features), or retrain the model on the 275 common feature positions only, and show that the OOD gains persist. Without such a control, the OOD improvement could be attributed to padding artifacts.
[§4.2, Tables 1–2] All scores are reported as point estimates with no error bars, seeds, or significance tests. The in-distribution Overall gain (0.60 vs. 0.56) is modest; the OOD biological metrics are very low in absolute terms (mAP no rep 0.07, Leiden ARI 0.03). The claim that CellPainTR 'drastically outperformed' baselines requires mean/std over at least three seeds and a statement of whether the differences are statistically significant. This is also necessary to assess whether the OOD Overall gap (0.40 vs. 0.26) is robust or within noise.
[§2, §4.2, §5.2] The comparison is limited to three classical baselines (ComBat, Harmony, Sphering). No deep-learning-based batch correction or recent Cell Painting representation learning method is included. The limitation is acknowledged in §5.2, but the 'state-of-the-art' and 'foundational model' conclusions require at least one strong learned baseline, such as a supervised contrastive model without source tokens or a VAE/scVI-style correction, to disentangle the contribution of the source-token mechanism and to support the broader claim of generalizable representation learning.

minor comments (5)

[§4.1 / References] The target dataset is referred to as 'Bray et al. (2017)' in the text, but the reference is Bray et al. (2016). Please correct the inconsistency.
[Eq. (1)] The ComBat model is written in a simplified form that is not fully defined (e.g., the dimensions of x_j and the meaning of α_i/β_i are ambiguous). Please provide the standard ComBat equation or a precise reference.
[Eq. (8)] The symbol S_k is used both for the embedding lookup and the appended token; the dimension and the indexing are not clearly distinguished. Please make the notation consistent.
[§4.3] The criterion for 'most similar experimental metadata' is never defined. Please state the metadata fields and the similarity measure used to select Source 10.
[Appendix C.2] The definitions of 'mAP control' and 'mAP no rep' are not fully specified. Please clarify the query set, the negative set construction, and how replicates are defined in the OOD setting.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central OOD claim rests on several assumptions: metadata similarity predicts batch effect similarity, zero-padding is a lossless feature-mismatch handling, InChIKey labels transfer across datasets, and JUMP/Bray compound sets do not overlap. None are justified in the paper.

free parameters (3)

Proxy source token index for OOD test = Source 10 (chosen by metadata similarity)
The OOD evaluation selects an existing training source token to represent the unseen Bray dataset; this choice is hand-selected per dataset and could materially affect results.
Temperature tau in supervised contrastive loss = 0.1
Temperature is a hyperparameter chosen without sensitivity analysis; contrastive embedding geometry depends on it.
Masking probability range = [0.05, 0.4]
The masking schedule in CWMM is chosen by hand and not ablated.

axioms (5)

domain assumption Cell Painting features, after MAD normalization, contain biological signal that is separable from batch effects.
The entire method assumes the feature space carries recoverable biological information independent of technical variation, stated in Sections 1 and 4.1.
domain assumption InChIKey serves as a valid biological label for contrastive learning across JUMP and Bray datasets.
Supervised stages (Steps 2 and 3) use InChIKey as the positive-pair identity; this assumes labels are consistent across datasets and perturbations, Section 3.2.
ad hoc to paper A source token selected by experimental metadata similarity adequately represents the batch effect of an unseen dataset.
The OOD protocol in Section 4.3 uses Source 10 as a proxy token for Bray data, assuming metadata similarity implies batch-effect similarity.
ad hoc to paper Zero-padding absent features preserves model validity for feature mismatch.
When only 275 of 4,765 features overlap, the model pads with zeros; this assumes the linear adaptor's bias and learned feature context can safely ignore missing features, Appendix B.2.
domain assumption JUMP held-out test set and Bray dataset have no compound overlap that would inflate OOD performance.
The paper does not check whether compounds in the Bray dataset appear in the JUMP training set; if they do, the contrastive pretraining could memorize them, Section 4.3.

reviewed 2026-08-05 · how reviews work

0 comments

Cite this review

Pith. "Pith review of CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis." pith.science (2026). https://pith.science/paper/VRYWF2W3

@misc{pith2026250906986,
  author       = {Pith},
  title        = {Pith review of: CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VRYWF2W3}},
  note         = {Machine review of arXiv:2509.06986}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Large-scale biological discovery requires integrating massive, heterogeneous datasets like those from the JUMP Cell Painting consortium, but technical batch effects and a lack of generalizable models remain critical roadblocks. To address this, we introduce CellPainTR, a Transformer-based architecture designed to learn foundational representations of cellular morphology that are robust to batch effects. Unlike traditional methods that require retraining on new data, CellPainTR's design, featuring source-specific context tokens, allows for effective out-of-distribution (OOD) generalization to entirely unseen datasets without fine-tuning. We validate CellPainTR on the large-scale JUMP dataset, where it outperforms established methods like ComBat and Harmony in both batch integration and biological signal preservation. Critically, we demonstrate its robustness through a challenging OOD task on the unseen Bray et al. dataset, where it maintains high performance despite significant domain and feature shifts. Our work represents a significant step towards creating truly foundational models for image-based profiling, enabling more reliable and scalable cross-study biological analysis.

Figures

Figures reproduced from arXiv: 2509.06986 by Cedric Caruzzo, Jong Chul Ye.

**Figure 1.** Figure 1: Conceptual Overview of the CellPainTR Framework and Training Curriculum. The standard Cell Painting workflow consists of (a) data preparation through cell seeding, perturbation, staining, and high-content screening (HCS), followed by (b) data generation, where raw cell images are processed by CellProfiler to produce high-dimensional feature vectors. Our novel contribution, shown in (c), is the CellPainTR m… view at source ↗

**Figure 2.** Figure 2: Detailed Architecture of the CellPainTR Model. The model processes single-cell features through a multi-stage pipeline. (a) The core of the model is a Transformer encoder composed of a stack of Bidirectional Hyena Layers. To prepare the input, the raw feature vector is processed by two parallel modules: (b) a Linear Adaptor which embeds the continuous feature values, and (c) a Feature Embedding which provi… view at source ↗

**Figure 3.** Figure 3: Overview of the CellPainTR Training Curriculum. The model is trained in a multi-stage process with evolving objectives. (a) First, input features are grouped by their imaging channel and the cellular compartment, and a subset of tokens are masked for the reconstruction task (Channel-Wise Masked Morphology). (b) A batch of these profiles is then processed. The different colors (e.g., pink, purple) represent… view at source ↗

**Figure 4.** Figure 4: UMAP visualizations of in-distribution results. Comparing (from left to right) Baseline (uncorrected) CellPainTR step 2 and CellPainTR step 3. The top row uses MoA coloring to show biological signal, while the bottom row uses source coloring to show batch effects. The final CellPainTR model (right) achieves cohesive batch integration while preserving clear compoundspecific patterns, demonstrating its effe… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

[1]

Michael Ando, Cory Y

D. Michael Ando, Cory Y. McLean, and Marc Berndl. Improving phenotypic measurements in high-content imaging screens. bioRxiv, 2017

work page 2017
[2]

Ewald, Robert van Dijk, Anne E

John Arevalo, Ellen Su, Jessica D. Ewald, Robert van Dijk, Anne E. Carpenter, and Shantanu Singh. Evaluating batch correction methods for image-based cell profiling. Nature Communications, 15: 0 50613, 2024

work page 2024
[3]

Decoding phenotypic screening: A comparative analysis of image representations

Adriana Borowa et al. Decoding phenotypic screening: A comparative analysis of image representations. Computational and Structural Biotechnology Journal, 23: 0 1181--1188, 2024

work page 2024
[4]

M.-A. Bray, S. Singh, H. Han, C. T. Davis, B. Borgeson, C. Hartland, and A. E. Carpenter. Cell painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nature Protocols, 11: 0 1757--1774, 2016

work page 2016
[5]

J. C. Caicedo, S. Cooper, F. Heigwer, S. Warchal, P. Qiu, C. Molnar, and A. E. Carpenter. Data-analysis strategies for image-based cell profiling. Nature Methods, 14: 0 849--863, 2017

work page 2017
[6]

Carpenter, Srinivas Niranj Chandrasekaran, Hugo Ceulemans, and Justin D

Anne E. Carpenter, Srinivas Niranj Chandrasekaran, Hugo Ceulemans, and Justin D. Boyd. Image-based profiling for drug discovery: due for a machine-learning upgrade? Nature Reviews Drug Discovery, 20: 0 145--159, 2020

work page 2020
[7]

Safiye Celik, Jan-Christian Huetter, Sandra Melo, Nathan Lazar, Rahul Mohan, Conor Tillinghast, Tommaso Biancalani, Marta Fay, Berton Earnshaw, and Imran S. Haque. Biological cartography: Building and benchmarking representations of life. In NeurIPS 2022 Workshop on Learning Meaningful Representations of Life, 2022

work page 2022
[8]

Michael Ando, John Arevalo, Melissa Bennion, Nicolas Boisseau, Adriana Borowa, Justin D

Srinivas Niranj Chandrasekaran, Jeanelle Ackerman, Eric Alix, D. Michael Ando, John Arevalo, Melissa Bennion, Nicolas Boisseau, Adriana Borowa, Justin D. Boyd, Laurent Brino, Patrick J. Byrne, Hugo Ceulemans, Carolyn Ch'ng, Beth A. Cimini, Djork-Arne Clevert, Nicole Deflaux, John G. Doench, Thierry Dorval, Regis Doyonnas, Vincenza Dragone, Ola Engkvist, P...

work page 2023
[9]

A simple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020

Pith/arXiv arXiv 2002
[10]

Bert: Pre-training of deep bidirectional transformers for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

Pith/arXiv arXiv 2018
[11]

S. M. Gustafsdottir, V. Ljosa, K. L. Sokolnicki, J. A. Wilson, D. Walpita, M. M. Kemp, and A. E. Carpenter. Multiplex cytological profiling assay to measure diverse cellular states. PLoS ONE, 8: 0 e80999, 2013

work page 2013
[12]

Evan Johnson, Cheng Li, and Ariel Rabinovic

W. Evan Johnson, Cheng Li, and Ariel Rabinovic. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics, 8: 0 118--127, 2007

work page 2007
[13]

Fast, sensitive and accurate integration of single-cell data with harmony

Ilya Korsunsky, Nathan Millard, Jean Fan, Kamil Slowikowski, Fan Zhang, Kevin Wei, Yulia Baglaenko, Michael Brenner, Po-Ru Loh, and Soumya Raychaudhuri. Fast, sensitive and accurate integration of single-cell data with harmony. Nature Methods, 16: 0 1289--1296, 2019

work page 2019
[14]

Kraus, K

O. Kraus, K. Kenyon-Dean, S. Saberian, M. Fallah, P. McLean, J. Leung, V. Sharma, A. Khan, J. Balakrishnan, S. Celik, and D. Beaini. Masked autoencoders for microscopy are scalable learners of cellular biology. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11757--11768, 2024

work page 2024
[15]

Ljosa, K

V. Ljosa, K. L. Sokolnicki, and A. E. Carpenter. Annotated high-throughput microscopy image sets for validation. Nature Methods, 9: 0 637, 2013

work page 2013
[16]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

2019
[17]

schyena: Foundation model for full-length single-cell rna-seq analysis in brain

Gyutaek Oh, Baekgyu Choi, Inkyung Jung, and Jong Chul Ye. schyena: Foundation model for full-length single-cell rna-seq analysis in brain. bioRxiv, 2023

work page 2023
[18]

Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré

Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré. Hyena hierarchy: Towards larger convolutional language models. arXiv preprint arXiv:2302.10866, 2023

Pith/arXiv arXiv 2023
[19]

Rousseeuw

Peter J. Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20: 0 53--65, 1987

work page 1987
[20]

Carpenter, Ola Spjuth, and Andreas Bender

Srijit Seal, Jordi Carreras-Puigvert, Shantanu Singh, Anne E. Carpenter, Ola Spjuth, and Andreas Bender. From pixels to phenotypes: Integrating image-based profiling with cell health data as biomorph features improves interpretability. Molecular Biology of the Cell, 35: 0 mr2, 1--13, 2024

work page 2024
[21]

Singh, M.-A

S. Singh, M.-A. Bray, T. R. Jones, and A. E. Carpenter. Pipeline for illumination correction of images for high-throughput microscopy. Nature Protocols, 12: 0 1709--1725, 2017

work page 2017
[22]

Traag, Ludo Waltman, and Nees Jan van Eck

Vincent A. Traag, Ludo Waltman, and Nees Jan van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific reports, 9 0 (1): 0 5233, 2019

work page 2019
[23]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30: 0 5998--6008, 2017

work page 2017
[24]

Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance

Nguyen Xuan Vinh, Julien Epps, and James Bailey. Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance. Journal of Machine Learning Research, 11: 0 2837--2854, 2010

work page 2010

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.

[1] [1]

Michael Ando, Cory Y

D. Michael Ando, Cory Y. McLean, and Marc Berndl. Improving phenotypic measurements in high-content imaging screens. bioRxiv, 2017

work page 2017

[2] [2]

Ewald, Robert van Dijk, Anne E

John Arevalo, Ellen Su, Jessica D. Ewald, Robert van Dijk, Anne E. Carpenter, and Shantanu Singh. Evaluating batch correction methods for image-based cell profiling. Nature Communications, 15: 0 50613, 2024

work page 2024

[3] [3]

Decoding phenotypic screening: A comparative analysis of image representations

Adriana Borowa et al. Decoding phenotypic screening: A comparative analysis of image representations. Computational and Structural Biotechnology Journal, 23: 0 1181--1188, 2024

work page 2024

[4] [4]

M.-A. Bray, S. Singh, H. Han, C. T. Davis, B. Borgeson, C. Hartland, and A. E. Carpenter. Cell painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nature Protocols, 11: 0 1757--1774, 2016

work page 2016

[5] [5]

J. C. Caicedo, S. Cooper, F. Heigwer, S. Warchal, P. Qiu, C. Molnar, and A. E. Carpenter. Data-analysis strategies for image-based cell profiling. Nature Methods, 14: 0 849--863, 2017

work page 2017

[6] [6]

Carpenter, Srinivas Niranj Chandrasekaran, Hugo Ceulemans, and Justin D

Anne E. Carpenter, Srinivas Niranj Chandrasekaran, Hugo Ceulemans, and Justin D. Boyd. Image-based profiling for drug discovery: due for a machine-learning upgrade? Nature Reviews Drug Discovery, 20: 0 145--159, 2020

work page 2020

[7] [7]

Safiye Celik, Jan-Christian Huetter, Sandra Melo, Nathan Lazar, Rahul Mohan, Conor Tillinghast, Tommaso Biancalani, Marta Fay, Berton Earnshaw, and Imran S. Haque. Biological cartography: Building and benchmarking representations of life. In NeurIPS 2022 Workshop on Learning Meaningful Representations of Life, 2022

work page 2022

[8] [8]

Michael Ando, John Arevalo, Melissa Bennion, Nicolas Boisseau, Adriana Borowa, Justin D

Srinivas Niranj Chandrasekaran, Jeanelle Ackerman, Eric Alix, D. Michael Ando, John Arevalo, Melissa Bennion, Nicolas Boisseau, Adriana Borowa, Justin D. Boyd, Laurent Brino, Patrick J. Byrne, Hugo Ceulemans, Carolyn Ch'ng, Beth A. Cimini, Djork-Arne Clevert, Nicole Deflaux, John G. Doench, Thierry Dorval, Regis Doyonnas, Vincenza Dragone, Ola Engkvist, P...

work page 2023

[9] [9]

A simple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020

Pith/arXiv arXiv 2002

[10] [10]

Bert: Pre-training of deep bidirectional transformers for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

Pith/arXiv arXiv 2018

[11] [11]

S. M. Gustafsdottir, V. Ljosa, K. L. Sokolnicki, J. A. Wilson, D. Walpita, M. M. Kemp, and A. E. Carpenter. Multiplex cytological profiling assay to measure diverse cellular states. PLoS ONE, 8: 0 e80999, 2013

work page 2013

[12] [12]

Evan Johnson, Cheng Li, and Ariel Rabinovic

W. Evan Johnson, Cheng Li, and Ariel Rabinovic. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics, 8: 0 118--127, 2007

work page 2007

[13] [13]

Fast, sensitive and accurate integration of single-cell data with harmony

Ilya Korsunsky, Nathan Millard, Jean Fan, Kamil Slowikowski, Fan Zhang, Kevin Wei, Yulia Baglaenko, Michael Brenner, Po-Ru Loh, and Soumya Raychaudhuri. Fast, sensitive and accurate integration of single-cell data with harmony. Nature Methods, 16: 0 1289--1296, 2019

work page 2019

[14] [14]

Kraus, K

O. Kraus, K. Kenyon-Dean, S. Saberian, M. Fallah, P. McLean, J. Leung, V. Sharma, A. Khan, J. Balakrishnan, S. Celik, and D. Beaini. Masked autoencoders for microscopy are scalable learners of cellular biology. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11757--11768, 2024

work page 2024

[15] [15]

Ljosa, K

V. Ljosa, K. L. Sokolnicki, and A. E. Carpenter. Annotated high-throughput microscopy image sets for validation. Nature Methods, 9: 0 637, 2013

work page 2013

[16] [16]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

2019

[17] [17]

schyena: Foundation model for full-length single-cell rna-seq analysis in brain

Gyutaek Oh, Baekgyu Choi, Inkyung Jung, and Jong Chul Ye. schyena: Foundation model for full-length single-cell rna-seq analysis in brain. bioRxiv, 2023

work page 2023

[18] [18]

Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré

Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré. Hyena hierarchy: Towards larger convolutional language models. arXiv preprint arXiv:2302.10866, 2023

Pith/arXiv arXiv 2023

[19] [19]

Rousseeuw

Peter J. Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20: 0 53--65, 1987

work page 1987

[20] [20]

Carpenter, Ola Spjuth, and Andreas Bender

Srijit Seal, Jordi Carreras-Puigvert, Shantanu Singh, Anne E. Carpenter, Ola Spjuth, and Andreas Bender. From pixels to phenotypes: Integrating image-based profiling with cell health data as biomorph features improves interpretability. Molecular Biology of the Cell, 35: 0 mr2, 1--13, 2024

work page 2024

[21] [21]

Singh, M.-A

S. Singh, M.-A. Bray, T. R. Jones, and A. E. Carpenter. Pipeline for illumination correction of images for high-throughput microscopy. Nature Protocols, 12: 0 1709--1725, 2017

work page 2017

[22] [22]

Traag, Ludo Waltman, and Nees Jan van Eck

Vincent A. Traag, Ludo Waltman, and Nees Jan van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific reports, 9 0 (1): 0 5233, 2019

work page 2019

[23] [23]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30: 0 5998--6008, 2017

work page 2017

[24] [24]

Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance

Nguyen Xuan Vinh, Julien Epps, and James Bailey. Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance. Journal of Machine Learning Research, 11: 0 2837--2854, 2010

work page 2010

CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it