Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Spatial Coordinates as a Cell Language: A Multi-Sentence Framework for Imaging Mass Cytometry Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Spatial2Sentence: adding neighbor cells to prompts improves LLM cell typing on IMC data.

desk verdict A genuinely new prompting idea for spatial single-cell LLMs, but the headline numbers don't match the tables and the clinical status result likely leaks donor identity. read the letter →

arxiv 2506.01918 v1 pith:EHAVULVV submitted 2025-06-02 cs.CL

Chi-Jane Chen , Yuhang Chen , Sukwon Yun , Natalie Stanley , Tianlong Chen This is my paper
classification cs.CL
keywords imagingmasscytometrysingle-celllanguagemodelspatialcontextmulti-sentencepromptingcontrastivelearningcell-typeclassificationclinicalstatuspredictionproteinexpression
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

This paper claims that large language models can analyze imaging mass cytometry (IMC) data more accurately when each cell is represented as a multi-sentence prompt rather than a single sentence of its own protein levels. The proposed framework, Spatial2Sentence, appends to a cell's protein-expression sentence the sentences of its most expression-similar and spatially nearest cells, and contrasts them with the most dissimilar and farthest cells. On preprocessed diabetes and brain tumor IMC datasets, this approach improves cell-type classification by 5.98% and clinical status prediction by 4.18% over existing single-cell LLMs, with the gains attributed to letting the model read cell-cell relationships in both expression and space. The paper further shows through ablations that the multi-sentence context, the positive/negative pair structure, and the ranking by expression similarity each contribute to the improvement.

What carries the argument

The load-bearing mechanism is the contrastive multi-sentence prompt built from two ranked neighbor lists. An expression similarity matrix $G$ (cosine similarity between cells' protein profiles) and a spatial distance matrix $D$ (Euclidean distance between cells' coordinates) are computed once per dataset; for each cell, cells are ranked by both measures. The positive prompt concatenates the target cell's protein sentence with the sentences of the top-K expression-similar and top-K spatially nearest cells; the negative prompt substitutes the top-K dissimilar and farthest cells. These prompts are fed to Llama-3.2-1B fine-tuned with multi-task learning, and the ablation results indicate that removing either pair type, replacing the ranking with random selection, or dropping the multi-sentence context degrades accuracy.

What would settle it

Re-run the clinical-status evaluation with all cells from each donor assigned exclusively to either the training or test set. If Spatial2Sentence's status-prediction advantage over the baselines disappears under this split, the reported 4.18% gain likely reflects donor-specific memorization rather than generalizable disease-status prediction.

Watch

Extended reading notes

Core claim

Spatial2Sentence's central claim is that spatial and cell-cell interaction information can be encoded directly into natural-language prompts, allowing an LLM to outperform dedicated single-cell models without architectural changes. Starting from Cell2Sentence's representation of a cell as a ranked list of its protein names, Spatial2Sentence constructs two pairwise matrices — cosine similarity of expression profiles and Euclidean distance of spatial coordinates — and uses them to rank every other cell. For each target cell, the model is prompted with positive pairs (top-K expression-similar cells and top-K spatially nearest cells) and negative pairs (top-K dissimilar and farthest cells), then fine-tuned to predict cell type and clinical status. Reported results on the authors' preprocessed diabetes and brain tumor IMC datasets place the method ahead of scGPT, Geneformer, Cell2Sentence, and related baselines on most cell-type classification tasks, and the authors attribute the gains to the contrastive multi-sentence context.

Load-bearing premise

The clinical-status results assume that the 90/10 data split separates donors, so that no donor contributes cells to both training and testing; the paper does not state that this separation was performed.

Editorial extensions

If this is right

  • If the claim holds, any single-cell LLM that currently treats cells independently can incorporate spatial context simply by changing its input prompt, without redesigning the model architecture.
  • Clinical status prediction from IMC tissue sections improves enough on these datasets to make patient-level classification a feasible downstream task for cell-level language models.
  • The negative-pair ablation shows that the top 1-3 most dissimilar cells provide the strongest contrastive signal, implying that prompt design has a measurable sweet spot rather than 'more contrast is always better'.
  • The gains persist across several LLM backbones (GPT-2, Llama, Gemma), suggesting the multi-sentence prompting strategy will transfer to larger and newer models as they become available.

Reading between the lines

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

  • The prompt essentially performs retrieval-augmented inference over cells: the model reads about neighbors rather than learning a spatial embedding, so the same recipe may apply to other graph-structured biological data expressible as text.
  • The reported clinical-status advantage should be tested under a donor-stratified split; if cells from the same donor appear in both training and test sets, part of the status accuracy could come from memorizing donor-specific protein signatures rather than learning disease biology.
  • Since the ablation shows that removing expression-similar sentences hurts more than removing spatial-neighbor sentences, a minimal version of the method might keep only expression-similar neighbors and still retain most of the gain, which would matter for datasets lacking coordinates.
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

4 major / 5 minor

Summary. The paper introduces Spatial2Sentence, a framework that converts imaging mass cytometry (IMC) protein expression profiles into rank-ordered textual sentences and augments each cell's sentence with positive and negative neighbor sentences chosen by cosine expression similarity and Euclidean spatial distance. This multi-sentence prompt is fed to a fine-tuned LLM (Llama-3.2-1B by default) with single- or multi-task learning for cell-type classification and donor-level clinical status prediction. Experiments on preprocessed Diabetes and brain-tumor IMC datasets report gains over single-cell LLM baselines, and ablations examine the contributions of positive/negative pairs, expression vs. spatial context, hyperparameter K, similarity metrics, and backbone LLMs.

Significance. The idea of encoding spatial interactions as contrastive text is a practical contribution, and the paper ships code, extensive ablations, and interpretability analyses. The method is not circular: labels are not used to construct prompts, and positive/negative pairs come only from expression and geometry. If the leakage concerns are resolved, the framework could provide a lightweight way to inject spatial context into LLM-based cell annotation. However, the status-prediction half of the central claim currently rests on an underspecified split, and the headline improvements in the abstract do not match any pairwise comparison in the tables; the empirical claims therefore need verification before the stated state-of-the-art conclusion can be accepted.

major comments (4)
  1. [4.2] Section 4.2 states only that 'the data set is divided into 90% for training and 10% for validation, with a separate test set reserved,' with no mention of donor-stratified splitting. Because the clinical status labels are donor-level (33 non-diabetic vs. 34 long-term T1D donors for Diabetes; 37 vs. 37 for brain tumor), a random cell-level split places cells from the same donor in both training and test, enabling the model to memorize donor-specific expression patterns. Under this setup, the reported status accuracies (e.g., 74.02 in Table 1 multi-task Diabetes) and the abstract's 4.18% improvement would not demonstrate generalization to unseen donors. Please provide a donor-stratified split, report donor-level status accuracy, or clearly identify the existing design if it is already donor-stratified.
  2. [3.2, 4.2] Section 3.2 constructs the expression similarity matrix G and spatial distance matrix D over all N cells before the data split is described in Section 4.2. The paper does not state that positive and negative neighbor sets are computed separately within the training, validation, and test splits. If the matrices are built on the full dataset, a test cell can appear as context inside training prompts, and a test query's prompt can include training cells, both of which leak information. Please specify the order of operations (matrix construction, neighbor selection, and split), and if necessary rerun with split-aware neighbor search.
  3. [Abstract, Section 1, Tables 1 and 3] The abstract and Section 1 claim improvements of 5.98% in cell-type classification and 4.18% in clinical status prediction on the Diabetes dataset, but these margins do not correspond to any pairwise comparison in Tables 1 or 3. In Table 1's multi-task Diabetes Type column, the margin over LangCell is 0.81 percentage points and over scGPT is 7.48 percentage points; in the multi-task Status column, Spatial2Sentence's 74.02 is below C2S w/ Spatial's 74.11. In Table 3 (Llama-3.2-1B), the gains over C2S are 4.44 percentage points for Type and 1.47 percentage points for Status. Please specify the exact baselines and single/multi-task settings used for the headline numbers, or revise the claims.
  4. [4.2, Tables 1-6] All tables report only mean accuracy over three seeds, with no standard deviations or significance tests. Several of the claimed improvements are small (e.g., 0.81 percentage points over LangCell in Table 1, and 1.47 percentage points over C2S in Table 3), so it is unclear whether they are statistically reliable. Please report per-seed results or confidence intervals, especially for the headline gains.
minor comments (5)
  1. [Tables 4 and 5] Table 4 lists 'Cosine Distance' under spatial distance methods and Table 5 lists 'Euclidean Distance' under expression similarity metrics; clarify what these variants compute, since the usual cosine distance for coordinates and Euclidean distance for expression vectors are not standard choices.
  2. [Figure 3] Figure 3 lacks axis labels and a legend; specify which accuracy (cell type vs. status, single vs. multi-task, dataset) is shown as a function of K.
  3. [4.1, Section 1] Section 4.1 contains grammatical errors, e.g., 'For those patients in the advanced stages of Type 1 diabetes are typically marked by...' and Section 1 'which containing protein expression matrices'; please edit for clarity.
  4. [4.2] Section 4.2 does not specify the size or composition of the reserved test set; state how many cells/samples are held out in each dataset.
  5. [4.2] The evaluation details for baselines are minimal; for example, it is unclear whether scGPT and Geneformer are fine-tuned on the same train split and with the same spatial preprocessing as Spatial2Sentence, which is needed to make the comparisons in Table 1 interpretable.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the method constructs prompts and contrastive pairs from expression and spatial inputs, not from target labels; the main validity concern is donor-level split leakage, which is a correctness issue, not circularity.

full rationale

The paper's derivation chain is self-contained and non-circular. The input representation is obtained by rank-ordering each cell's protein expression into a sentence (Eq. 1-2), following C2S. The positive and negative pairs are selected from a cosine similarity matrix (Eq. 3-4) and a Euclidean distance matrix (Eq. 5-6), both computed entirely from the input expression matrix and spatial coordinates. These pairs are formatted into prompts, and the model is fine-tuned to predict cell type or clinical status from the prompt text. Nowhere are the target labels used to construct the prompts or the similarity/distance matrices, so the predicted quantities are not defined in terms of the model outputs nor vice versa. The self-citations in the paper (e.g., Stanley et al. 2020 for cell-type annotation and Yun et al. 2024 for spatial analysis background) are not load-bearing for the central derivation; they concern preprocessing tools and related work. The strongest concern raised by a skeptical reading is that the clinical-status split in Section 4.2 is not explicitly donor-stratified, which could inflate status accuracy through donor-level leakage. That is a serious experimental-validity issue, but it is not a circularity of the kind defined by the analysis rules: it does not make the claimed result equivalent to its inputs by construction. The method is also evaluated against external baselines and includes ablations, further supporting that the core claim has independent empirical content. Overall, the paper shows no significant circularity; at most it has minor non-load-bearing self-citations, so a low score is appropriate.

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

The central claim relies on the C2S rank-based sentence representation, the choice of similarity/distance metrics, and the assumption that adding neighbor sentences improves prediction. No invented entities are introduced. The free parameters are metric choices and the prompt size K, all selected on the test data.

free parameters (3)
  • K (number of similar/distant cells in prompt) = 1-3 (top 1-3 used in best config)
    Hyperparameter controlling how many neighbor sentences are concatenated; selected via sensitivity analysis (Figure 3, Table 6).
  • Expression similarity metric = cosine similarity
    Chosen over Pearson correlation and Euclidean distance based on Table 5 results on the diabetes and brain tumor test sets.
  • Spatial distance metric = Euclidean distance
    Chosen over L1 norm and cosine distance based on Table 4 results.
assumptions (3)
  • domain assumption Rank-ordered protein names (C2S representation) preserve sufficient cell identity information for classification.
    The framework converts each cell's expression to a sentence by ranking proteins (Section 3.1), relying on C2S's hypothesis that rank reflects cell properties.
  • domain assumption Cells sharing expression similarity and spatial proximity are relevant context for predicting the current cell's type and clinical status.
    Central to positive/negative pair construction in Section 3.3; no biological validation is provided for these relationships on these datasets.
  • domain assumption The random 90/10 split of cells separates donors so that clinical status labels do not leak between train and test.
    The paper states the split without specifying donor stratification (Section 4.2); if cells from the same donor appear in both sets, donor-level status can be memorized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial Coordinates as a Cell Language: A Multi-Sentence Framework for Imaging Mass Cytometry Analysis." pith.science (2026). https://pith.science/paper/EHAVULVV

@misc{pith2026250601918,
  author       = {Pith},
  title        = {Pith review of: Spatial Coordinates as a Cell Language: A Multi-Sentence Framework for Imaging Mass Cytometry Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHAVULVV}},
  note         = {Machine review of arXiv:2506.01918}
}
read the original abstract

Image mass cytometry (IMC) enables high-dimensional spatial profiling by combining mass cytometry's analytical power with spatial distributions of cell phenotypes. Recent studies leverage large language models (LLMs) to extract cell states by translating gene or protein expression into biological context. However, existing single-cell LLMs face two major challenges: (1) Integration of spatial information: they struggle to generalize spatial coordinates and effectively encode spatial context as text, and (2) Treating each cell independently: they overlook cell-cell interactions, limiting their ability to capture biological relationships. To address these limitations, we propose Spatial2Sentence, a novel framework that integrates single-cell expression and spatial information into natural language using a multi-sentence approach. Spatial2Sentence constructs expression similarity and distance matrices, pairing spatially adjacent and expressionally similar cells as positive pairs while using distant and dissimilar cells as negatives. These multi-sentence representations enable LLMs to learn cellular interactions in both expression and spatial contexts. Equipped with multi-task learning, Spatial2Sentence outperforms existing single-cell LLMs on preprocessed IMC datasets, improving cell-type classification by 5.98% and clinical status prediction by 4.18% on the diabetes dataset while enhancing interpretability. The source code can be found here: https://github.com/UNITES-Lab/Spatial2Sentence.

Figures

Figures reproduced from arXiv: 2506.01918 by the authors.

Figure 1
Figure 1. Overall framework of Spatial2Sentence. (a) Given IMC data, we integrate both the protein expression matrix and the spatial coordinate information simultaneously. The protein expression matrix (with cells as rows and proteins as columns) is transposed (i.e., V T ) and multiplied to compute cosine similarity. Similarly, the spatial coordinate matrix (with columns representing the X and Y coordinates) is processed to o… view at source ↗
Figure 2
Figure 2. Given a (a) multi-sample IMC dataset (b), we used Cellpose to detect cell centers within superpixels (c) and extracted cells from IMC images for cell-type annotation. (d) We then applied the SLIC algorithm to segment images into superpixel region, (e) generating a cell × protein feature matrix for analysis [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Accuracy performance of Spatial2Sentence across different values of the hyperparameter K. mance metrics is classification accuracy. Experiment Details. We use the Llama-3.2- 1B (Dubey et al., 2024) model for our experiments by default, fine-tuning it with the following train￾ing parameters: the batch size is set to 8 per device, and we apply a learning rate of 2e-4 with a cosine learning rate scheduler. The model is… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Cell-type distribution in brain tumor dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [1]

    Reinders, and Ahmed Mahfouz

    Tamim Abdelaal, Vincent van Unen, Thomas Höllt, Frits Koning, Marcel J.T. Reinders, and Ahmed Mahfouz. 2018. Predicting cell types in single cell mass cytometry data. bioRxiv, 19(6):759--769

    work page 2018

  2. [2]

    Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine S \"u sstrunk. 2010. Slic superpixels

    work page 2010

  3. [3]

    Sean C Bendall, Garry P Nolan, Mario Roederer, and Pratip K Chattopadhyay. 2012. A deep profiler's guide to cytometry. Trends in immunology, 33(7):323--332

    work page 2012

  4. [4]

    Petter Brodin, Darragh Duffy, and Lluis Quintana-Murci. 2019. A call for blood—in human immunology. Immunity, 50(6):1335--1336

    work page 2019

  5. [5]

    Yiqun Chen and James Zou. 2024. Genept: a simple but effective foundation model for genes and cells built from chatgpt. bioRxiv, pages 2023--10

    2024

  6. [6]

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. 2024. scgpt: toward building a foundation model for single-cell multi-omics using generative ai. Nature Methods, pages 1--11

    work page 2024

  7. [7]

    Nicolas Damond, Stefanie Engler, Vito RT Zanotelli, Denis Schapiro, Clive H Wasserfall, Irina Kusmartseva, Harry S Nick, Fabrizio Thorel, Pedro L Herrera, Mark A Atkinson, et al. 2019. A map of human type 1 diabetes progression by imaging mass cytometry. Cell metabolism, 29(3):755--768

    work page 2019

  8. [8]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv 2024

Show all 38 references
  1. [9]

    Chen Fang, Yidong Wang, Yunze Song, Qingqing Long, Wang Lu, Linghui Chen, Guihai Feng, Yuanchun Zhou, and Xin Li. 2024. How do large language models understand genes and cells. ACM Transactions on Intelligent Systems and Technology

    2024

  2. [10]

    Charlotte Giesen, Hao AO Wang, Denis Schapiro, Nevena Zivanovic, Andrea Jacobs, Bodo Hattendorf, Peter J Sch \"u ffler, Daniel Grolimund, Joachim M Buhmann, Simone Brandt, et al. 2014. Highly multiplexed imaging of tumor tissues with subcellular resolution by mass cytometry. N...

    2014

  3. [11]

    Felix J Hartmann and Sean C Bendall. 2020. Immune monitoring using mass cytometry and related high-dimensional imaging approaches. Nature Reviews Rheumatology, 16(2):87--99

    2020

  4. [12]

    Karthik A Jagadeesh, Kushal K Dey, Daniel T Montoro, Rahul Mohan, Steven Gazal, Jesse M Engreitz, Ramnik J Xavier, Alkes L Price, and Aviv Regev. 2022. Identifying disease-critical cell types and cellular processes by integrating single-cell rna-sequencing and human genetics. ...

    2022

  5. [13]

    Vijayakumar R Kakade, Marlene Weiss, and Lloyd G Cantley. 2021. Using imaging mass cytometry to define cell identities and interactions in human tissues. Frontiers in Physiology, 12:817181

    2021

  6. [14]

    Elham Karimi, Miranda W Yu, Sarah M Maritan, Lucas JM Perus, Morteza Rezanejad, Mark Sorin, Matthew Dankner, Parvaneh Fallah, Samuel Dor \'e , Dongmei Zuo, et al. 2023. Single-cell spatial immune landscapes of primary and metastatic brain tumours. Nature, 614(7948):555--563

    2023

  7. [15]

    Leeat Keren, Marc Bosse, Diana Marquez, Roshan Angoshtari, Samir Jain, Sushama Varma, Soo-Ryum Yang, Allison Kurian, David Van Valen, Robert West, et al. 2018. A structured tumor-immune microenvironment in triple negative breast cancer revealed by multiplexed ion beam imaging....

    2018

  8. [16]

    Wei Lan, Guohang He, Mingyang Liu, Qingfeng Chen, Junyue Cao, and Wei Peng. 2024. Transformer-based single-cell language model: A survey. Big Data Mining and Analytics, 7(4):1169--1186

    2024

  9. [17]

    Hao-Chih Lee, Roman Kosoy, Christine E Becker, Joel T Dudley, and Brian A Kidd. 2017. Automated cell type discovery and classification through knowledge transfer. Bioinformatics, 33(11):1689--1695

    2017

  10. [18]

    Daniel Levine, Syed Asad Rizvi, Sacha L \'e vy, Nazreen Pallikkavaliyaveetil, David Zhang, Xingyu Chen, Sina Ghadermarzi, Ruiming Wu, Zihe Zheng, Ivan Vrkic, et al. 2023. Cell2sentence: teaching large language models the language of biology. BioRxiv, pages 2023--09

    2023

  11. [19]

    Tianyu Liu, Tianqi Chen, Wangjie Zheng, Xiao Luo, and Hongyu Zhao. 2023. scelmo: Embeddings from language models are good learners for single-cell data analysis. bioRxiv, pages 2023--12

    2023

  12. [20]

    Vivien Marx. 2021. Method of the year: spatially resolved transcriptomics. Nature methods, 18(1):9--14

    2021

  13. [21]

    Vladan Milosevic. 2023. Different approaches to imaging mass cytometry data analysis. Bioinformatics Advances, 3(1):vbad046

    2023

  14. [22]

    Nitya Nair, Henrik E Mei, Shih-Yu Chen, Matthew Hale, Garry P Nolan, Holden T Maecker, Mark Genovese, C Garrison Fathman, and Chan C Whiting. 2015. Mass cytometry as a platform for the discovery of cellular biomarkers to guide effective rheumatic disease therapy. Arthritis res...

    2015

  15. [23]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  16. [24]

    Anjali Rao, Dalia Barkley, Gustavo S Fran c a, and Itai Yanai. 2021. Exploring tissue architecture using spatial transcriptomics. Nature, 596(7871):211--220

    2021

  17. [25]

    Amy Reece, Bingzhao Xia, Zhongliang Jiang, Benjamin Noren, Ralph McBride, and John Oakey. 2016. Microfluidic techniques for high throughput single cell analysis. Current opinion in biotechnology, 40:90--96

    2016

  18. [26]

    Luca D Schreurs, Alexander F Vom Stein, Stephanie T J \"u nger, Marco Timmer, Ka-Won Noh, Reinhard Buettner, Hamid Kashkar, Volker Neuschmelting, Roland Goldbrunner, and Phuong-Hien Nguyen. 2025. The immune landscape in brain metastasis. Neuro-Oncology, 27(1):50--62

    2025

  19. [27]

    Aya M Shaaban, Nancy M Salem, and Lamees N Mahmoud. 2024. Cutting-edge approaches to cell segmentation in imaging mass cytometry: A detailed review. In 2024 6th Novel Intelligent and Leading Emerging Sciences Conference (NILES), pages 469--474. IEEE

    2024

  20. [28]

    Natalie Stanley, Ina A Stelzer, Amy S Tsai, Ramin Fallahzadeh, Edward Ganio, Martin Becker, Thanaphong Phongpreecha, Huda Nassar, Sajjad Ghaemi, Ivana Maric, et al. 2020. Vopo leverages cellular heterogeneity for predictive modeling of single-cell data. Nature communications, ...

    2020

  21. [29]

    Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. 2021. Cellpose: a generalist algorithm for cellular segmentation. Nature methods, 18(1):100--106

    2021

  22. [30]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L \'e onard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram \'e , et al. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118

    2024 arXiv

  23. [31]

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. 2023. Transfer learning enables predictions in network biology. Nature, 618(7965):616--624

    2023

  24. [32]

    Luyi Tian, Fei Chen, and Evan Z Macosko. 2023. The expanding vistas of spatial transcriptomics. Nature Biotechnology, 41(6):773--782

    2023

  25. [33]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  26. [34]

    Logan A Walsh and Daniela F Quail. 2023. Decoding the tumor microenvironment with spatial technologies. Nature Immunology, 24(12):1982--1993

    2023

  27. [35]

    Sukwon Yun, Jie Peng, Alexandro E Trevino, Chanyoung Park, and Tianlong Chen. 2024. Mew: Multiplexed immunofluorescence image analysis through an efficient multiplex network. In European Conference on Computer Vision, pages 127--144. Springer

    2024

  28. [36]

    Suyuan Zhao, Jiahuan Zhang, Yushuai Wu, Yizhen Luo, and Zaiqing Nie. 2024. Langcell: Language-cell pre-training for cell identity understanding. arXiv preprint arXiv:2405.06708

    2024 arXiv

  29. [37]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  30. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.