henhao
Sign in
Systems BiologybioRxiv
Preprint

TxCyto: A machine learning framework for estimating cytokine activity from whole transcriptome

Rahul Kumar, Shan LiiD, Peng JiangiD, Sridhar S. HannenhalliiD

Abstract

Cytokines are critical mediators of intercellular communication, and a comprehensive characterization of their activity is essential for understanding health and disease. Existing tools to infer cytokine activity rely on experimental measurements. However, such measurements are available only for a small minority (43) of cytokines, and moreover, cytokine activity and response are highly context-specific, making a comprehensive experimental profiling across tissues, disease states, and biological contexts impractical. To address this gap, we developed TxCyto - a deep learning-based framework that infers the activity of cytokines, and more broadly of the tumor secretome, directly from the whole transcriptome profile of a sample. Trained on pan-cancer TCGA tumor transcriptomes, TxCyto was extensively validated in multiple independent datasets, including cytokine perturbation experiments. Across multiple cancer immunotherapy cohorts, TxCyto identified cytokines whose predicted activity was associated with therapeutic response. Furthermore, in spatial transcriptomic data for Liver cancer, TxCyto discovered spatial niches associated with response to immunotherapy. Overall, we develop a machine learning tool -TxCyto, for predicting the activity of 645 cytokines and tumor secretome from readily available whole transcriptomes. The TxCyto framework is generally applicable to other classes of regulatory molecules and TxCyto code base, and the tools are provided at https://github.com/Rahulncbs/TxCyto.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (29 Sept 2026) · CC0 1.0

A fully connected neural network trained on 9491 TCGA bulk tumor transcriptomes estimates activity of 645 cytokines and secreted proteins, checked against perturbation and protein data.

Existing cytokine activity tools such as CytoSig learn from perturbation experiments that cover only 43 cytokines in few immune contexts, and profiling every cytokine experimentally across tissues is impractical. TxCyto avoids perturbation data by learning from bulk tumor transcriptomes how global expression relates to each cytokine's mRNA level, which extends coverage to 645 secreted proteins.

Key results

  1. ShownThe bulk-transcriptome FCNN predicted held-out cytokine mRNA better than 11 deconvolved cell type-specific models and beat linear ridge and tree-based XGBoost baselines. It partly generalized to tumor types left out of training.

    Median 5-fold CV Pearson r 0.78 for bulk FCNN vs 0.38 for ridge regression and about 0.70 for XGBoost; 482 of 645 cytokines exceed CV 0.6 (442 for the malignant-cell deconvolved model); leave-one-tissue-out gave median accuracy >0.4 in 19 of 33 tumor types. · TxCyto accurately estimates cytokine activity…; Model construction and performance evaluation

  2. SuggestedTraining on bulk TCGA transcriptomes beat training on scRNA-seq pseudobulk in cross-modality transfer and in detecting cytokine induction. For some cytokines, predictions matched measured protein better than the observed mRNA did.

    Induction data: positive delta in 134 of 202 cytokine-experiment conditions (about 66%) for the TCGA-trained model vs 106 (about 52%) for the scRNA-seq pseudobulk-trained model (1062 samples). Ovarian cohort (32 samples, 23 cytokines): predicted level beat observed mRNA in protein correlation for 9 cytokines. · Tissue-level bulk transcriptomes capture…; Model construction and performance evaluation

  3. ShownPredicted activity fell after cytokine inhibitor treatment in most cases, slightly more often than with CytoSig, though the two agreed on direction in only 11 of 20 cases. It rose after in vitro stimulation in most PBMC cell types and correlated positively with canonical downstream pathway scores.

    Inhibitor cohorts: negative delta in 16 of 20 cases (80%) vs 15 of 20 for CytoSig. PBMC stimulation (12 donors): 46% of 765 tests (45 cytokines x 17 cell types) significant at Wilcoxon p <= 0.05, above 50% of cytokines in B cells, CD8 memory T cells and dendritic cells; 15 cytokine-pathway pairs mostly positively correlated. · TxCyto correctly predicts lower cytokine…; TxCyto predicts cytokine-stimulation across…; Cytokine-specific downstream pathways positively…; Independent validation of the model

  4. SuggestedBefore treatment, predicted activity of an IFN-gamma and T-cell-recruiting chemokine program was associated with response to checkpoint blockade across cancer types. Cytokines linked to non-response differed by tumor type (angiogenesis, matrix remodeling, developmental and neuroendocrine programs).

    19 studies, 5 cancer types; one-sided Mann-Whitney U, Benjamini-Hochberg FDR<0.25, same direction in three or more cohorts (renal ccRCC, lung NSCLC, melanoma) or two or more (head and neck HNSCC, breast). Responder-linked: CXCL9, CXCL10, CXCL11, FLT3LG, IL-15, EBI3; non-responder: VEGFA, ANGPT2 (ccRCC, melanoma), COL11A1 (ccRCC, breast). · Identification of cytokines associated with…; Application

  5. SuggestedSpot-level secretome activity profiles from Visium data separated responder from non-responder tumor regions with moderate accuracy, in hepatocellular carcinoma (immunotherapy) and high-grade serous ovarian cancer (chemotherapy). Both cohorts were very small.

    Logistic regression on 645 predicted activities, leave-one-slide-out. HCC (7 patients; 4 responders, 3 non-responders to neoadjuvant cabozantinib plus nivolumab): spot AUC 0.75, sample AUC 0.66. Ovarian (12 patients; 6 vs 6 to neoadjuvant chemotherapy): spot AUC 0.77, sample AUC 0.83. · Spatially resolved tumor-secretome activity…; Application

Methods, data and reproducibility

A fully connected neural network (FCNN) takes 6075 genes shared across cohorts (secretome genes excluded) and predicts, independently for each of 645 secreted proteins, its mRNA level as a proxy for activity. It is trained on pan-cancer TCGA bulk RNA-seq and scored by 5-fold cross-validation (CV) Pearson correlation.

  • TCGA pan-cancer bulk RNA-seq (Broad GDAC Firehose), 33 tumor types, plus CODEFACS-deconvolved expression for 11 cell types · human (TCGA) · n = 9491 samples · 10.5281/ZENODO.4469784 (deconvolved data)
  • Pseudobulked scRNA-seq tumor cohort with malignant and immune populations, used as an alternative training set · n = 1062 samples · 10.5281/ZENODO.15554080
  • Cytokine induction transcriptomic experiments used to compare TCGA-trained and pseudobulk-trained models · n = 202 cytokine-experiment conditions, 18 cytokines · GEO IDs in Supplementary Table 2
  • Ovarian cancer cohort with matched transcriptomes and cytokine protein quantification · n = 32 samples, 23 cytokines
  • Cytokine inhibitor studies (lupus, psoriasis, atopic dermatitis, rheumatoid arthritis) with skin biopsy or whole blood RNA-seq · human · n = 20 inhibitor-treatment cases · GEO IDs in Supplementary Table 3
  • PBMC scRNA-seq after in vitro stimulation with 90 cytokines, pseudobulked per immune cell type · human · n = 12 donors; 45 cytokines x 17 cell types analyzed
  • Pre-treatment transcriptomes from immunotherapy cohorts (breast, melanoma, HNSCC, NSCLC, ccRCC) · human (clinical cohorts) · n = 19 studies · GEO IDs in Supplementary Table 6
  • 10x Visium spatial transcriptomics: HCC (neoadjuvant cabozantinib plus nivolumab) and high-grade serous ovarian cancer (neoadjuvant chemotherapy) · human (patient samples) · n = 7 HCC and 12 ovarian patient samples
Methods & tools
TxCyto (fully connected neural network), CODEFACS (sample-specific deconvolution), Ridge regression (baseline), XGBoost (baseline), CytoSig (comparator), ssGSEA / GSEA pathway scoring, Logistic regression (spatial response classifier), Mann-Whitney U test with Benjamini-Hochberg correction, Wilcoxon test, 10x Genomics Visium
Code
https://github.com/Rahulncbs/TxCyto
Data & accessions
https://gdac.broadinstitute.org/https://zenodo.org/records/7074291
Reproducibility gaps
Several cohorts (ovarian protein, PBMC stimulation, lung receptor, Visium) are cited without accessions in the text; GEO IDs are deferred to supplementary tables; no software versions or hardware are given.
Limitations (6)
  • Cytokine mRNA abundance stands in for activity because protein measurements are scarce; the authors note that even protein level only approximates activity, which also depends on post-translational modification and functional interactions. · stated by the authors
  • The authors acknowledge that the HCC spatial cohort has very few donors (4 responders, 3 non-responders), so the sample-level AUC of 0.66, obtained by averaging spot probabilities per slide, rests on minimal data. · stated by the authors
  • Because the model is trained on bulk tissue, changes within a single cell compartment may not fully reflect the global cytokine shift, a caveat the authors raise when applying it to pseudobulked PBMC cell types. · stated by the authors
  • The cytokines associated with immunotherapy response are offered as candidate biomarkers and targets that the authors say still need experimental validation. · stated by the authors
  • Accuracy is benchmarked mainly as reconstruction of each cytokine's own mRNA. Margins over CytoSig on inhibitor data (16 vs 15 of 20 cases) and over scRNA-trained models (10 vs 9 of 18 cytokines) are small and come without formal significance tests. · noted in this summary
  • The Discussion frames the ovarian spatial classifier as predicting immunotherapy response, yet that cohort was treated with neoadjuvant chemotherapy; the immunotherapy cytokine screen also relies on a permissive FDR<0.25 cutoff. · noted in this summary

Who should read:Read this if you want to estimate the activity of hundreds of cytokines or secreted proteins from bulk, pseudobulk, or spatial transcriptomes, or are looking for cytokine biomarkers of checkpoint immunotherapy response.

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
THISJavaid '23Mukherjee '25Kus '25Bagherlou '26Yamagishi '22Wang '23
61% semantic matchBioinformatics
Azka Javaid, Hildreth Robert Frost·bioRxiv·Oct 17, 2023· 5 citations