Combining BulkFormer and TabPFN to predict post- transplant function from kidney biopsies during machine perfusion or cold storage

Combining BulkFormer and TabPFN to predict post- transplant function from kidney biopsies during machine perfusion or cold storage | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Combining BulkFormer and TabPFN to predict post- transplant function from kidney biopsies during machine perfusion or cold storage Samuel J Tingle, Georgios Kourounis, Sofia Kazerouni, Harry VM Spiers, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9242336/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Generating predictions from transcriptomic data poses a unique challenge due to the high number of genes, and often small sample size. BulkFormer and TabPFN have emerged as leading transformer-based foundation models for bulk transcriptomic and tabular data respectively. We explore an artificial intelligence pipeline using BulkFormer-TabPFN v2.5 which generates zero-shot predictions from raw RNA-Seq count data without retraining. This was tested on three cohorts of biopsies taken from donated human kidneys. BulkFormer-TabPFN was able to predict delayed kidney function using RNA-Seq counts from kidneys undergoing ex-situ normothermic machine perfusion (NMP; c-statistic=0.82, 95% CI=0.67–0.97). Predictive discrimination was optimised under the following conditions: BulkFormer-TabPFN versus TabPFN alone, maximum absolute aggregation of BulkFormer gene-level embeddings, biopsies taken during ex-situ NMP versus cold storage. BulkFormer-TabPFN predictions were modified by cytokine filter treatment during ex-situ NMP, suggesting they could be a dynamic surrogate endpoint for novel therapeutics, which is intrinsically linked to post-transplant outcome. This demonstrates for the first time synergistic benefits of these foundation models, to generate zero-shot predictions without model retraining. The provided code provides a blueprint to replicate generating predictions from RNA-Seq count data, which could be applied in a wide range of biomedical contexts within and beyond organ transplantation. Biological sciences/Biological techniques Health sciences/Biomarkers Biological sciences/Computational biology and bioinformatics Health sciences/Nephrology BulkFormer TabPFN foundation model transcriptomics Bulk RNA-Seq kidney transplantation Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Lack of certainty around how a donated organ will function after transplantation frequently leads to organ non-use, contributing to a global shortage of transplants. 1 A technique termed normothermic machine perfusion (NMP), where a warm blood-based solution is circulated through an isolated organ to maintain it outside of the body, provides an opportunity to assess organ health. 2–4 However, more accurate measures of kidney viability during NMP are urgently needed. 4,5 Previous studies have identified transcriptomic signatures during kidney NMP that are associated with poor graft outcomes post-transplant, highlighting the potential for using transcriptomic analysis in this setting. 6,7 However, using transcriptomic data for prediction is challenging for several reasons. Firstly, data is generally available for relatively few samples, compared to the large clinical cohorts underpinning most prediction models. 8 Secondly, there is a very large number of features, with information on tens of thousands of gene transcripts per sample. Standard statistical prediction modelling has to reduce this large number of genes to select a set of promising genes to use for modelling; a process where overfitting and overoptimism are virtually guaranteed, especially in small datasets. 9–11 In addition, this manual process is inherently variable and analyst-dependent, introducing potential bias in feature selection and reducing reproducibility. Combining BulkFormer and TabPFN v2.5 may allow us to overcome these issues. 12–14 These are both state of the art foundation models which use a transformer-based architecture; the same architecture type underpinning large language models such as ChatGPT. The models take in data as a ‘prompt’, and output distilled information or predictions, without requiring any model retraining; this is termed “zero-shot” predictions. TabPFN has been at the heart of a recent revolution in the analysis of tabular data (data organised in rows and columns). 15 It is a state of the art model, outperforming established machine learning approaches for tabular data, such as XGBoost and random forests. 13 TabPFN is especially strong when sample size is low, making it an attractive option for use with current transcriptomic datasets. However, performance has been reported to degrade with high-dimensional data, especially when the number of features exceeds 2,000. 13,16 Standard approaches to dimensionality reduction often lose important data, or rely on relationships between predictors and outcome thereby potentially amplifying noise leading to overfitting. BulkFormer is a foundation model trained on over 500,000 bulk transcriptomic samples. 14 It can process transcriptomic data on 20,010 genes and distil this information into a compact set of embeddings, each of which reflects a key biological programme. This distilled data can then be used for a range of tasks, including disease annotation, patient prognosis prediction and drug response prediction. 14 Despite these conceptual advantages to deal with the specific challenges posed by transcriptomics data sets, there are no published articles describing the combination of BulkFormer and TabPFN v2.5. Our aim was therefore to evaluate the performance of a novel combination BulkFormer-TabPFN pipeline to make outcome predictions based on raw transcriptomic count data. Methods This is an exploratory study to assess the ability of a BulkFormer-TabPFN pipeline to generate clinical outcome predictions from small sample-size transcriptomic datasets. All data and code used in this study is available in our accompanying GitHub repository. 17 Ethical approval was not required as all raw data (RNA-Seq raw counts and metadata) was retrieved from previous publications, either via the gene expression omnibus (GEO) or direct contact with the study authors. This raw count data, along with its metadata, is also available in our GitHub repository. 17 A summary of the computing pipeline is given in Figure 1A. All computing steps shown were performed zero-shot, without any model retraining. Our primary clinical outcome was delayed graft function (DGF), often defined as the need for dialysis within the first week post-transplant. This measure of early post-transplant kidney function has been shown to correlate with long-term kidney survival, and is frequently used as the primary outcome in both preclinical and clinical research. 5,6,18,19 As dialysis in the first day post-transplant often reflects hyperkalaemia, rather than truly reflecting function of the transplanted kidney, recipients receiving dialysis only during the first 24 hours post-transplant (with no further dialysis in the first week), were classified as not having DGF. Therefore, our definition of DGF was requirement for dialysis in the first week post-transplant, outside of the first 24 hours. This approach is consistent with prior work demonstrating that transplants where dialysis is only performed in the first 24 hours have transcriptomic profiles comparable to kidneys where no post-transplant dialysis is performed. 6,7 Summary of study cohorts and methodology A summary of the three study cohorts is given in Figure 1B and Table 1. All three cohorts include bulk RNA-Seq raw count data from renal cortex tissue biopsies of human kidneys donated for transplantation. For all cohorts, tissue was homogenised, and RNA isolated with the pure link RNA mini kit (Ambion). Library preparation was performed using the TruSeq Stranded total RNA library prep kit (Illumina), and sequenced on a Hiseq sequencer (Illumina). Alignment to the human genome (Hg38) was performed with HISAT2, and featurecounts was used to generate the raw count data. 6,18 Further details on the study cohorts, and on precise methodologies for generating the raw count data, can be found in the original publications. 6,18 The NMP trial cohort consists of biopsies performed at 1 hour of NMP as part of the United Kingdom NMP randomised trial. 5 The data for N=34 kidneys was kindly provided by Ferdinand et al, and included the 33 kidneys included in their manuscript, plus an additional case. All biopsies had linked DGF data, in-line with our definition above. 6 The Cytokine Filter cohort included paired human kidney experiments (N=10 kidneys). These were performed on human kidneys retrieved for transplant but then deemed unsuitable and offered for research. One kidney per pair received 4 hours of NMP plus cytokine filter treatment, and the other kidney was an NMP-only control. None of these kidneys were transplanted. Raw count data was taken from GEO (accession GSE121447), with additional metadata provided in the associated manuscript. 6 This cohort allowed us to assess whether DGF predictions were modifiable by therapeutic intervention during NMP, with the paired design being a particular strength. 6 The Quality in Organ Donation (QUOD) retrieval biopsy cohort contains biopsies which were taken at the time of kidney organ retrieval (following cold flush of the organs), prior to static cold storage and transport. 18 Raw count data was taken from GEO (accession GSE215108), with metadata taken from manuscript supplementary data. 18 Of the available biopsies, 263 had sufficient metadata to allow labelling of DGF in line with our definition above. The NMP and QUOD retrieval biopsy cohorts were analysed and evaluated entirely independently, with no cross-cohort prediction between NMP and QUOD retrieval biopsy cohorts. Processing of raw count data The raw count data for all three cohorts was indexed using gene symbols, rather than the Ensembl IDs required for BulkFormer. These gene symbols were mapped to Ensembl gene IDs using Ensembl BioMart (using biomaRt; hsapiens_gene_ensembl ). 20–22 Symbols not mapped by BioMart were resolved via the HGNC REST API to identify approved symbols from aliases/previous symbols, and these approved symbols were then re-mapped to Ensembl gene IDs via biomaRt. 22,23 The mapping file was created on 11 th Feb 2026 using the most recent Ensembl (release 115) and HGNC versions, and is available on our GitHub. 17,20 Where a single gene symbol mapped to multiple Ensembl IDs, we prioritised Ensembl IDs that are used in the BulkFormer pipeline (either as one of the input genes or a gene with length information for their transcript per million, TPM, calculation). 24,25 Counts for any duplicate Ensembl ID were summed. Any gene symbols which could not be mapped to Ensembl ID used by BulkFormer were initially retained for TPM calculations. Raw count data was then converted to TPM and log-transformed to log e (TPM); this was performed using code provided in the BulkFormer GitHub, to ensure consistency with the data used for previous BulkFormer training. 24 Generating BulkFormer Embeddings We used the BulkFormer (version 2) 147M model to generate embeddings. 14,24,25 The BulkFormer GitHub repository was cloned on 10 th January 2026, and all associated files (including model weights and gene annotation files) were downloaded on the same day from Zenodo. 25 The code for generating embeddings was taken directly from the “bulkformer_extract_feature.ipynb” Jupyter notebook. The key area where the BulkFormer code was modified was the method for aggregating gene-level to sample level embeddings. BulkFormer generates latent space embeddings (n=643 for the model used here) per gene, per sample. For prediction tasks, gene-level embeddings need to be aggregated to sample-level embeddings, such that each sample is now represented by n=643 unique embeddings. The available options for aggregation were: maximum (used in BulkFormer manuscript) 14 , mean and median. We have added an option for maximum absolute value, so that large negative embeddings are also captured. As sensitivity analyses we also added options for using the 99.9 th percentile rather than the maximum (attempting to make the maximum aggregation approaches robust to potential single gene outliers). These new options are available on our GitHub. 17 TabPFN v2.5 TabPFN v2.5 was used with model weights downloaded from Hugging Face (a repository for model weights) on 19th February 2026. 26 We utilized the default model (tabpfn-v2.5-classifier-v2.5_default.ckpt) for all feature sets and additionally evaluated the XL model variant (tabpfn-v2.5-classifier-v2.5_large-features-XL.ckpt) for log(TPM) data. The models operate by generating predictions conditioned on contextual examples from the input data, rather than undergoing traditional parameter retraining. The input to these models was either: (1) BulkFormer sample-level embeddings (n=643 features) generated as described above, and (2) log(TPM) values containing either all genes or only the 20,010 genes used for BulkFormer inference; this was done to assess the impact of omitting the BulkFormer step. We evaluated model performance using leave-one-out cross-validation (LOO-CV) within the NMP trial and QUOD retrieval biopsy cohorts separately. Within each cohort, outcome-labelled data were provided to the pipeline as context, with an outcome value for a single kidney (occurrence of DGF) masked and predicted by the model. This process of masking a single kidney’s outcome was repeated for all kidneys within the cohort, allowing comparison of predicted and observed outcomes to assess predictive performance. This was performed in the NMP trial and QUOD cohorts completely separately (neither cohort used any information from the other cohort in any way). We also generated the predicted probability of DGF for each sample in the cytokine filter cohort, using the NMP trial cohort as context. This was performed to assess whether DGF predictions were dynamic, and whether they could be influenced by novel therapies. We refer to these as “pseudoprobabilities” because these biopsies were taken at different timepoints of NMP compared with the NMP trial cohort. Therefore, they do not represent cohort-calibrated estimated true probabilities. Higher values indicate a greater predicted probability of DGF compared with other cases in the dataset, but the numerical values themselves should not be assumed to represent absolute probabilities. Statistical analysis Statistical analysis and reporting were conducted in line with recent best-practice guidance for reporting binary prediction model performance in medical AI research. 27 Model discrimination was assessed using the c-statistic (area under the receiver operating characteristic curve, AUC), which was prespecified as the primary performance metric. Overall predictive performance was evaluated using the Brier score. 27 Confidence intervals (CI) were calculated using a bootstrap approach for all metrics, except for the c-statistic, where the DeLong method implemented in the pROC R library was used. 28 For visualisation, ROC curves were plotted to illustrate discrimination, and violin plots were used to display the distribution of performance estimates across feature representations and aggregation strategies. The code used for performance evaluation and visualisation was adapted from the best-practice guidance paper’s accompanying GitHub repository. 29 To assess the difference in pseudoprobabilities of DGF in the paired kidney experiments (Cytokine Filter cohort) we used a two-tailed paired t-test. Results The novel computing pipeline is displayed in Fig. 1 A, and a summary of study cohorts is given in Table 1 and Fig. 1 B. Kidney cohort sample sizes were: 34 for the NMP trial samples (9/34 with DGF), 10 for the Cytokine Filter preclinical cohort (5 pairs) and 263 for the QUOD retrieval biopsy cohort (73/263 with DGF). Full details on every tested model is given in Supplementary Table 1, and includes all predictive model metrics recommended by Calster et al. 27 Combining BulkFormer and TabPFN allows DGF prediction We initially analysed the NMP trial cohort. Our BulkFormer-TabPFN pipeline allowed RNA-Seq raw count data from biopsies taken at 1 hour of normothermic machine perfusion to be used to predict early post-transplant function, as measured by delayed graft function (Fig. 2 ). Assessing various options for aggregation of BulkFormer gene-level to sample-level embeddings revealed that maximum absolute aggregation resulted in the best model performance (Fig. 2 A&B). Maximum absolute and maximum aggregation produced models with C-statistics whose 95% confidence intervals excluded 0.5, indicating discrimination significantly better than chance, whereas mean and median aggregation did not. As a sensitivity analysis we also aggregated embeddings using 99.9th percentile, to see whether maximum embeddings were overly influenced by outliers; these models showed inferior predictive performance (Supplementary Table 1). We next assessed predictive performance of a pipeline which did not use BulkFormer and instead applied TabPFN directly to the transcriptomic data (Fig. 1 A). When the BulkFormer step was omitted, predictive performance was lost, with TabPFN-only pipelines generating predictions that were worse than chance (Fig. 2 A&B). This was true for both the default TabPFN v2.5 models, and TabPFN models specifically trained for high feature count data (the v2.5 XL models). Figure 2 C displays receiver-operating characteristic curves for the top three performing model pipelines, based on c-statistic. The best overall model was the BulkFormer-TabPFN model which used maximum absolute embedding aggregation (c-statistic = 0.82, 95% CI = 0.67–0.97; Brier score = 0.15, 95% CI = 0.08–0.22). The distribution of predicted probabilities of DGF, stratified by actual DGF status, is displayed in Fig. 2 D. Predicted DGF probability is dynamic and modulated by a novel therapy To assess whether the predicted DGF probability from this pipeline is sensitive to therapeutic modulation of the kidney, we used a cohort of biopsies from a preclinical study of a novel therapy. 6 This consisted of paired kidneys (i.e. both kidneys from the same donor) undergoing normothermic machine perfusion with or without cytokine filter treatment. Using embedding and outcome data from the NMP trial cohort as the context, we generated zero-shot predicted probability of DGF for each of the Cytokine Filter kidney biopsies (displayed in Fig. 3 ). As the machine perfusion timepoints differed between groups (1 hour in NMP trial cohort; 2 and 4 hours in cytokine filter cohort), we have labelled this as a pseudoprobability. At the end of cytokine filter treatment (4 hours), there was a significant reduction in predicted pseudoprobability of DGF in the Cytokine Filter group (mean difference in DGF pseudoprobability = -21.3%, 95% CI = -38.7% to -4.0%, paired t-test p = 0.027) compared to the control group. This occurred despite no difference between groups at the 2-hour timepoint (mean difference in DGF pseudoprobability = + 4.7%, 95% CI = -24.8% to 34.2%, p = 0.683). Figure 3 displays these results. Biopsy during NMP allows better predictions than biopsies during cold storage We next assessed whether the same pipeline (Fig. 1 A) could be used to predict DGF, in an independent cohort of biopsies taken from kidneys flushed with cold preservation solution (thus at 4 o C temperature) at the time of organ retrieval (QUOD cohort), as opposed to kidney biopsies taken during normothermic machine perfusion (Fig. 1 B). This analysis was conducted entirely within the QUOD cohort, with no use of NMP trial samples at any stage, including as ‘context’. Full model performance metrics in the retrieval biopsy cohort are provided in Supplementary Fig. 1 and Supplementary Table 1. Mirroring the results in the NMP cohort, the BulkFormer-TabPFN model incorporating maximum absolute embedding aggregation ranked highest by c-statistic (Supplementary Fig. 1A). Within the QUOD retrieval biopsy cohort, only BulkFormer-TabPFN using absolute and maximum aggregation produced models with C-statistics whose 95% confidence intervals excluded 0.5. Consistent with the NMP findings, models omitting the BulkFormer step demonstrated no evidence of predictive discrimination. With an identical model pipeline, outcome definition, and an identical transcriptomic pipeline performed in the same laboratory, 6,18 predictive performance was significantly higher when using NMP biopsies versus the QUOD kidney biopsies (Fig. 4 ). For example, with the maximum absolute aggregation approach, discrimination was strong in the NMP cohort (c-statistic = 0.82, 95% CI = 0.67–0.97) but substantially lower in the QUOD cohort (c-statistic = 0.59, 95% CI = 0.52–0.67). Discussion We show for the first time the ability of an AI pipeline combining BulkFormer Version 2 and TabPFN v2.5 to generate outcome predictions from raw transcriptomic data. In the setting of the small datasets examined here, the addition of BulkFormer for context-aware distillation of gene data into embeddings significantly improved predictive performance, compared with pipelines using TabPFN alone. Predictive performance was sensitive to the method used to aggregate gene-level into sample-level embeddings; we found that maximum absolute aggregation of BulkFormer embeddings was optimal, a method which we believe has not been described previously. 14 , 24 All code is freely available on GitHub and includes tools enabling readers to apply the pipeline to their own raw count data. 17 Due to the small sample size, and single centre collection/processing of samples, this work is exploratory and is clearly not a clinical prediction model which is ready for deployment. Nevertheless, there are important messages and opportunities for transplantation. Firstly, this adds to the body of evidence for improved organ assessment using NMP; 2–4 with our pipeline, transcriptomic signatures during NMP led to superior predictions compared with transcriptomic signatures at retrieval. Importantly, the NMP trial and QUOD retrieval biopsy cohorts were analysed completely independently, with no sharing of samples, features, or contextual information at any stage. QUOD cohort predictions were not informed by the NMP cohort, and the observed differences in performance are due to the underlying biological signal themselves. Secondly, we have provided evidence that these predictions are responsive to the delivery of a novel NMP therapy (Cytokine filters). 6 With larger cohorts of outcome-linked biopsy data, this could be a potential surrogate endpoint in preclinical or early-phase trials, which could potentially reduce the need for larger, more expensive clinical trials in the future. As a continuous variable it offers increased power and lower sample sizes in comparison to binary endpoints, and it is intrinsically tied to post-transplant outcome. This is particularly important given the substantial volume of preclinical NMP therapeutic studies, where robust, translatable efficacy readouts are currently limited. 30 , 31 Although larger cohorts of outcome-linked data are required, the code we provide 17 acts as a blueprint for the generation of DGF predicted probabilities from kidney biopsy raw count data. Transcriptomic assessment of kidney biopsies has entered select clinical practice for post-transplant biopsies, most notably with the Molecular Microscope Diagnostics System (MMDx). 32 Although sequencing costs have reduced substantially (< $ 150 for bulk RNA-Seq), turn-around time remains a hurdle in the time-pressured setting of organ assessment. Advances in next-generation sequencing with nanopore have already been reported to provide sample-to-answer times of just 6 hours. 33 As RNA-Seq workflows continue to improve, and technologies for extending organ preservation time evolve, 34 transcriptomic assessment could one day guide organ accepting decisions. 34 – 37 This manuscript provides a blueprint for how AI foundation models can be combined to make rapid predictions from raw count data, without manual bioinformatic analysis. A key advantage of our pipeline is that it performs predictions ‘zero-shot’, meaning that no model retraining is performed, and none of the model weights are changed. This is possible as both BulkFormer and TabPFN are foundation models. They use transformer-based architectures, and train on massive amounts of data. 12 , 14 After this training, they can then take new labelled data as a ‘prompt’ and provide predictions with a single forward pass through the network, akin to how a large language model can take text as a prompt and provide a response (without retraining the model weights). This contrasts with standard approaches to generate predictions from RNA-Seq data. Standard statistical methods generally rely on feature selection (selection of ‘significant’ genes) prior to subsequent modelling. This process is inherently unstable, and in small datasets overfitting and overoptimism are virtually guaranteed. 9 – 11 These issues are magnified if variable/feature selection is performed using association with the outcome being predicted. 38 , 39 Even options for dimensionality reduction such as principal components can distort inference by inflating type I error, whilst losing biologically relevant signal. 40 , 41 Established machine-learning methods, including random forests and gradient boosting, remain vulnerable to the combined challenges of high feature dimensionality, noise, and small sample size, often resulting in unstable models which are overfitted to noise. 42 , 43 For these reasons, we have deliberately avoided comparison of our zero-shot approach, to any methods which require retraining, as they would not provide a meaningful benchmark. TabPFN is trained to approximate Bayesian inference, and has demonstrated state-of-the-art performance for tabular data predictions across a wide range of disciplines, outperforming bespoke models using established machine learning approaches. 13 In addition to improved predictive performance, TabPFN offers practical advantages: prediction is rapid as model retraining is entirely avoided, implementation is often simpler as hyperparameter tuning is not required, and performance is particular strong in the setting of low sample-size. 12 One limitation of TabPFN is its significantly decreased performance with data that exceeds 2,000 dimensions. 13 , 16 This is clearly a challenge for transcriptomic data, where the number of genes exceeds 10,000–20,000. This is where we turned to BulkFormer version 2, 25 which utilises a graph neural-network to incorporate known gene-gene interactions, and a performer-based architecture which was trained on over 500,000 bulk RNA-Seq samples. 14 It takes RNA-Seq data on 20,010 genes, and distils this into a compressed set of embeddings which represent biological signal, whilst minimising information loss. In the original BulkFormer manuscript these embeddings were fed into a random forest model to generate predictions. 14 BulkFormer outperformed existing single-cell trained models on a range of tasks including patient prognosis prediction. 14 The benefits of these two models are complementary for transcriptomic data. TabPFN performs well across multiple disciplines, especially in the setting of small sample sizes, however struggles with high feature count data. BulkFormer allows the number of features to be reduced, whilst preserving biological signal. Our findings support these conceptual benefits. TabPFN would also allow very simple integration of donor characteristics into the same model, as additional features, which may further improve predictions. Despite these advantages, and growing use of both models individually, to our knowledge only a single pre-print article has described the combination of BulkFormer and TabPFN. 44 This found no benefit of integrating BulkFormer with TabPFN. There are several potential reasons for the difference in conclusions with this study. Firstly, they aggregated gene-level embeddings using mean aggregation only; an approach which also failed to improve predictions in our manuscript. They also used older versions of both BulkFormer and TabPFN, and used datasets with over 10,000 samples (which is not where TabPFN shows the most benefit). 13 , 25 , 44 This study falls outside of the standard paradigm of internal and external validation of clinical prediction models, as it uses zero-shot models. 45 As there is no model retraining, many of the standard approaches for internal validation studies, such as optimism calculation/correction, do not apply. However, this does not represent true external validation, as the study cohort used as the input is still used for generating predictions. 45 , 46 The most recent TRIPOD + AI guidelines explicitly excluded foundational models, and the TRIPOD-LLM guidelines do not cover tabular models. In the future new guidelines will be required to cover the application of foundational tabular models for clinical prediction. 45 The main limitations relate to the retrospective nature of the study, and reliance on available sample numbers and metadata. Overall, there was a small available sample size of biopsy data linked to post-transplant outcome, with minimal metadata on other important clinical factors (such as donor age). As a result it is unclear how much this transcriptomic data adds to standard clinical factors. However, our BulkFormer-TabPFN pipeline achieved superior predictive discrimination (c-statistic) than DGF predictions models using clinical data, and predictions were modifiable by treatment (in kidney pairs with identical clinical variables). 47 Whilst combining BulkFormer-TabPFN improved performance compared with TabPFN alone, combining two models amplifies the “black-box” nature of foundation models, making it challenging to attribute predictions to specific genes. Another limitation is that all biopsies from all cohorts were processed by the same laboratory and subjected to an identical RNA-Seq generation pipeline, which may limit generalisability and transferability to external cohorts. In addition, NMP biopsies linked to outcome data were only available at the 1-hour timepoint. As transcriptomic profiles evolve over the course of NMP, 6 the applicability of models derived from early biopsies to samples obtained after prolonged perfusion remains uncertain and warrants further investigation. The benefit of our approach is that as larger outcome-linked RNA-Seq datasets become available, they can be incorporated into our pipeline, allowing iterative evaluation of generalisability, consistency, transportability and predictive performance. In conclusion, transcriptomic data poses a unique challenge due to the high number of genes, relative to the often small sample size. Standard methods for prediction rely on feature selection, which leads to overfitting and optimism about model performance. Here we combine two state of the art artificial intelligence models, BulkFormer and TabPFN. These transformer-based foundation models can generate predictions “zero-shot”, without retraining on new data. This manuscript is the first to demonstrate the synergistic benefits of combining these models in terms of predictive performance. Our freely available code offers a blueprint to apply this pipeline to generate predictions from raw RNA-Seq, which could be applied in a wide range of contexts beyond deceased donor kidney transplantation. Declarations Acknowledgements Thank you to Professor Menna Clatworthy, Benjy Tan and John Ferdinand for their help in accessing the raw data used in this study. We would also like to thank Vasilis Kosmoliaptsis for the help of his laboratory team in acquiring raw data, and for providing comments on the manuscript contents. Funding declaration SJT was funded for this work via a Medical Research Council Clinical Research Training Fellowship (MRC/Y000676/1), which was part-funded by Kidney Research UK. This study is funded by the National Institute for Health and Care Research (NIHR) Blood and Transplant Research Unit in Organ Donation and Transplantation (NIHR203332), a partnership between the National Health Service (NHS) Blood and Transplant, the University of Cambridge, and Newcastle University. The views expressed are those of the author(s) and not necessarily those of the NIHR, NHS Blood and Transplant, or the Department of Health and Social Care. ChatGPT v5.2 was used for copy editing only, after the first draft was finalised. The authors remain accountable for all aspects of the work. Author contributions Study concept: SJT Study design: SJT, GK Curation of existing data sources: SJT Building BulkFormer-TabPFN pipeline: SJT, GK Data analysis: SJT, GK Interpretation of results: All authors Initial drafting of manuscript: SJT, GK Critical review and editing of manuscript: All authors Review of final manuscript version: All authors Data availability statement All data and code are freely available open access via our GitHub repository for the project https://github.com/TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics Competing Interests Statement The authors declare no competing interests. References Olawade, D. B., Marinze, S., Qureshi, N., Weerasinghe, K. & Teke, J. Transforming organ donation and transplantation: Strategies for increasing donor participation and system efficiency. Eur. J. Intern. Med. 133 , 14–24. 10.1016/j.ejim.2024.11.010 (2025). Jaynes, C. L., Goggins, W. C., Holzner, M. L., Garonzik-Wang, J. & Leuvenink, H. G. D. No Kidney Left Behind: Rescuing Unused Donor Kidneys for Transplant at the First Centralized Assessment and Repair Center. Transpl. Int. 38 , 15424. 10.3389/ti.2025.15424 (2025). Hosgood, S. A., Barlow, A. D., Hunter, J. P. & Nicholson, M. L. Ex vivo normothermic perfusion for quality assessment of marginal donor kidney transplants. Br. J. Surg. 102 (11), 1433–1440. 10.1002/bjs.9894 (2015). Tingle, S. J. et al. Contrast-Enhanced Ultrasound to Assess Kidney Quality During Ex Situ Normothermic Machine Perfusion. Transpl. Int. 38 , 14268. 10.3389/ti.2025.14268 (2025). Hosgood, S. A. et al. Normothermic machine perfusion versus static cold storage in donation after circulatory death kidney transplantation: a randomized controlled trial. Nat. Med. 29 (6), 1511–1519. 10.1038/s41591-023-02376-7 (2023). Ferdinand, J. R. et al. Cytokine absorption during human kidney perfusion reduces delayed graft function-associated inflammatory gene signature. Am. J. Transpl. 21 (6), 2188–2199. 10.1111/ajt.16371 (2021). Spiers, H. V. M. et al. Transcriptomic profiling during normothermic machine perfusion of human kidneys reveals a pro-inflammatory cellular landscape and gene expression signature associated with severe ischemia-reperfusion injury and delayed graft function. Front. Immunol. 16 10.3389/fimmu.2025.1679251 (2025). Rao, P. S. et al. A comprehensive risk quantification score for deceased donor kidneys: the kidney donor risk index. Transplantation 88 (2), 231–236. 10.1097/TP.0b013e3181ac620b (2009). Clarke, R. et al. The properties of high-dimensional data spaces: implications for exploring gene and protein expression data. Nat. Rev. Cancer . 8 (1), 37–49. 10.1038/nrc2294 (2008). Simon, R., Radmacher, M. D., Dobbin, K. & McShane, L. M. Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J. Natl. Cancer Inst. 95 (1), 14–18. 10.1093/jnci/95.1.14 (2003). Ambroise, C. & McLachlan, G. J. Selection bias in gene extraction on the basis of microarray gene-expression data. Proceedings of the National Academy of Sciences . ;99(10):6562–6566. (2002). 10.1073/pnas.102102699 Hollmann, N. et al. Accurate predictions on small data with a tabular foundation model. Nature 637 (8045), 319–326. 10.1038/s41586-024-08328-6 (2025). Grinsztajn, L. et al. TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models. arXiv . Preprint posted online February 5, 2026:arXiv:2511.08667. 10.48550/arXiv.2511.08667 Kang, B., Fan, R., Yi, M., Cui, C. & Cui, Q. A large-scale foundation model for bulk transcriptomes. bioRxiv . Preprint posted online June 17, 2025:2025.06.11.659222. 10.1101/2025.06.11.659222 Zhang, Q., Tan, Y. S., Tian, Q. & Li, P. TabPFN: One Model to Rule Them All? arXiv . Preprint posted online November 27, 2025:arXiv:2505.20003. 10.48550/arXiv.2505.20003 Ye, H. J., Liu, S. Y. & Chao, W. L. A Closer Look at TabPFN v2: Understanding Its Strengths and Extending Its Capabilities. arXiv . Preprint posted online June 11, 2025:arXiv:2502.17361. 10.48550/arXiv.2502.17361 Tingle, S. J. & Kourounis, G. TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics: Data and code repository for project. GitHub. Accessed February 27, (2026). https://github.com/TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics Zhang, R. et al. Assessment of biological organ age using molecular pathology in pre-transplant kidney biopsies. Kidney Int. 106 (2), 302–316. 10.1016/j.kint.2024.03.028 (2024). Venkataraman, K., Irish, G. L., Collins, M. G. & Clayton, P. A. The Association Between Early Graft Function, Donor Type and Long-Term Kidney Transplant Outcomes. Transpl. Int. 38 , 14197. 10.3389/ti.2025.14197 (2025). Dyer, S. C. et al. Ensembl 2025. Nucleic Acids Res. 53 (D1), D948–D957. 10.1093/nar/gkae1071 (2025). Durinck, S. et al. BioMart and Bioconductor: a powerful link between biological databases and microarray data analysis. Bioinformatics 21 (16), 3439–3440. 10.1093/bioinformatics/bti525 (2005). Durinck, S., Spellman, P. T., Birney, E. & Huber, W. Mapping identifiers for the integration of genomic datasets with the R/Bioconductor package biomaRt. Nat. Protoc. 4 (8), 1184–1191. 10.1038/nprot.2009.97 (2009). Braschi, B. et al. Genenames.org: the HGNC and VGNC resources in 2019. Nucleic Acids Res. 47 (D1), D786–D792. 10.1093/nar/gky930 (2019). Kang, B. KangBoming/BulkFormer. Published online February 7, 2026. Accessed January 10, (2026). https://github.com/KangBoming/BulkFormer Kang, B., Fan, R., Cui, C. & Cui, Q. A large-scale foundation model for human bulk transcriptomes (version 2). Published online Dec. 11 10.5281/zenodo.15744294 (2025). Prior-Labs/tabpfn_2_5 · Hugging Face. November 11. Accessed February 27, 2026. (2025). https://huggingface.co/Prior-Labs/tabpfn_2_5 Calster, B. V. et al. Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance. Lancet Digit. Health . 7 (12). 10.1016/j.landig.2025.100916 (2025). Robin, X. et al. pROC: an open-source package for R and S + to analyze and compare ROC curves. BMC Bioinform. 12 , 77 (2011). benvancalster. benvancalster/PerfMeasuresOverview. Published online February 5. Accessed February 27, 2026. (2026). https://github.com/benvancalster/PerfMeasuresOverview Franchin, B. et al. Normothermic Kidney Perfusion: Current Status and Future Perspectives. Kidney Int. Rep. 10 (9), 2943–2952. 10.1016/j.ekir.2025.06.041 (2025). Rabelink, T. J. et al. Opportunities and challenges with the implementation of normothermic machine perfusion in kidney transplantation. Nat. Commun. 16 (1), 6883. 10.1038/s41467-025-60410-3 (2025). Halloran, P. F., Madill-Thomsen, K. S. & Reeve, J. The Molecular Phenotype of Kidney Transplants: Insights From the MMDx Project. Transplantation 108 (1), 45–71. 10.1097/TP.0000000000004624 (2024). Gu, W. et al. Rapid pathogen detection by metagenomic next-generation sequencing of infected body fluids. Nat. Med. 27 (1), 115–124. 10.1038/s41591-020-1105-z (2021). Dumbill, R. et al. Prolonged normothermic perfusion of the kidney prior to transplantation: a historically controlled, phase 1 cohort study. Nat. Commun. 16 (1), 4584. 10.1038/s41467-025-59829-5 (2025). Saju, A. F., Mukundan, A., Divyashree, M. S., Hariharapura, R. C. & Mahadev Rao, A. RNA diagnostics and therapeutics: a comprehensive review. RNA Biol. 22 (1), 1–11. 10.1080/15476286.2024.2449277 (2025). Chu, V. T. et al. Implementation and outcomes of a rapid response genomic hospital epidemiology programme at an academic medical centre over 7 years. Lancet Microbe . 7 (2), 101277. 10.1016/j.lanmic.2025.101277 (2026). Shuja, A. Advances in RNA-Seq Data Analysis: Towards More Reliable Clinical Translation: Advances in RNA-Seq Data Analysis: Towards Reliable Clinical Translation. Dev. MEDICO-LIFE-SCIENCES ; 2 (8):1–2. doi: 10.69750/dmls.02.08.0143 (2025). Collins, G. S. et al. Evaluation of clinical prediction models (part 1): from development to external validation. BMJ 384 , e074819. 10.1136/bmj-2023-074819 (2024). Harrell, F. E. xRegression Modeling Strategies: With Applications, to Linear Models, Logistic and Ordinal Regression, and Survival Analysis, 2nd ed. Heidelberg: Springer. Biometrics . ;72(3):1006–1007. (2016). 10.1111/biom.12569 Jolliffe, I. T. & Cadima, J. Principal component analysis: a review and recent developments. Philos Trans A Math Phys Eng Sci . ;374(2065):20150202. (2016). 10.1098/rsta.2015.0202 Nygaard, V., Rødland, E. A. & Hovig, E. Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses. Biostatistics 17 (1), 29–39. 10.1093/biostatistics/kxv027 (2016). Feldner-Busztin, D. et al. Dealing with dimensionality: the application of machine learning to multi-omics data. Bioinformatics 39 (2), btad021. 10.1093/bioinformatics/btad021 (2023). Kumar, S., Agarwal, A. & Chatterjee, S. Feature learning augmented with sampling and heuristics (FLASH) improves model performance and biomarker identification. NPJ Syst. Biol. Appl. 11 , 137. 10.1038/s41540-025-00614-x (2025). Zhou, S., Agarwal, V., Gopinath, A. & Kassis, T. The Limitations of TabPFN for High-Dimensional RNA-seq Analysis. bioRxiv . Preprint posted online September 17, 2025:2025.08.15.670537. 10.1101/2025.08.15.670537 Collins, G. S. et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385 , e078378. 10.1136/bmj-2023-078378 (2024). Gallifant, J. et al. The TRIPOD-LLM reporting guideline for studies using large language models. Nat. Med. 31 (1), 60–69. 10.1038/s41591-024-03425-5 (2025). Irish, W. D., Ilsley, J. N., Schnitzler, M. A., Feng, S. & Brennan, D. C. A Risk Prediction Model for Delayed Graft Function in the Current Era of Deceased Donor Renal Transplantation. Am. J. Transplant. 10 (10), 2279–2286. 10.1111/j.1600-6143.2010.03179.x (2010). Table Table 1 – study cohort details, and our aims for each cohort. Normothermic Machine Perfusion (NMP) is a way of keeping an isolated kidney alive outside of the body. Delayed graft function (DGF) was our primary measure of how well the kidney was functioning after transplant. Cohort Details Source Aim NMP Trial samples N=34 kidneys biopsied during the United Kingdom NMP randomised trial. 5 All kidneys were subsequently transplanted, with data on early post-transplant function available for all. 9/34 recipients experienced DGF. Data provided upon request from the authors of Ferdinand 2021 6 Assess ability of our pipeline to predict DGF using biopsies taken during NMP Cytokine Filter samples N=10 kidneys (5 pairs). Preclinical normothermic machine perfusion with or without a cytokine filter treatment. The aim of the cytokine filter treatment is to improve post-transplant kidney function. Ferdinand 2021 6 and associated GEO accession GSE121447 Assess whether DGF predictions are modifiable by therapies delivered during NMP QUOD retrieval biopsy samples Kidneys biopsied at the donor hospital (before transport of the kidney on ice). These biopsies were collected and stored in the Quality in Organ Donation biobank. N=263 kidneys with sufficient data on DGF. 73/263 recipients experienced DGF. Zhang 2024 18 and associated GEO accession GSE215108 Assess ability to predict DGF with retrieval biopsies (in kidneys not receiving NMP). Additional Declarations No competing interests reported. Supplementary Files Supplementarydata.pdf Supplementary figure caption Supplementary Figure 1 – Predictive performance for early post-transplant kidney function (defined as Delayed Graft Function; DGF) in the QUOD retrieval biopsy cohort. The discrimination (measured using c-statistic) and overall performance (Brier score) for each of the tested models are shown in panel A and B, and are ranked best to worst. Models using the full BulkFormer-TabPFN pipeline are shown in green, and are labelled with which BulkFormer method was used for aggregation of gene-level embeddings (e.g. Max absolute). Models using only TabPFN directly on the transcriptomic data (without BulkFormer) are shown in orange, and labelled with which genes were used, and whether the TabPFN default or high feature count (XL model) were used. C) Receiver-operating characteristic curves for the top three performing models by c-statistic. D) Distribution of predicted probabilities stratified by subsequent delayed graft function status. Supplementary table caption Supplementary table 1 – All model performance metrics for all models in the NMP and QUOD cohorts. For models using the BulkFormer-TabPFN pipeline (BF-TabPFN) the labels reflect the BulkFormer method for aggregating gene-level embeddings. For models using the just TabPFN on the raw transcript per million data (without BulkFormer), the label represents which genes were used, and whether the TabPFN default or high feature count (XL model) were used. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 15 May, 2026 Reviewers invited by journal 14 May, 2026 Editor invited by journal 03 Apr, 2026 Editor assigned by journal 28 Mar, 2026 Submission checks completed at journal 28 Mar, 2026 First submitted to journal 27 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9242336","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":617761070,"identity":"011759e1-b05e-4764-b101-6e2f1bb558ee","order_by":0,"name":"Samuel J Tingle","email":"data:image/png;base64,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","orcid":"","institution":"Newcastle University","correspondingAuthor":true,"prefix":"","firstName":"Samuel","middleName":"J","lastName":"Tingle","suffix":""},{"id":617761071,"identity":"3e2c9a54-3af5-4fd8-b767-a7e438666796","order_by":1,"name":"Georgios Kourounis","email":"","orcid":"","institution":"Newcastle University","correspondingAuthor":false,"prefix":"","firstName":"Georgios","middleName":"","lastName":"Kourounis","suffix":""},{"id":617761072,"identity":"a1cbaf17-6ac9-4590-bf38-60ad697206d2","order_by":2,"name":"Sofia Kazerouni","email":"","orcid":"","institution":"Newcastle University","correspondingAuthor":false,"prefix":"","firstName":"Sofia","middleName":"","lastName":"Kazerouni","suffix":""},{"id":617761073,"identity":"10c62550-fe77-4de3-aab2-366b314486ce","order_by":3,"name":"Harry VM Spiers","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Harry","middleName":"VM","lastName":"Spiers","suffix":""},{"id":617761074,"identity":"7cd55b2e-3c00-4fd4-9beb-c4ad8a4a475f","order_by":4,"name":"Miguel Larraz","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Miguel","middleName":"","lastName":"Larraz","suffix":""},{"id":617761075,"identity":"e8686dd1-8468-4083-a112-2ce2244eb117","order_by":5,"name":"Maithili Mehta","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Maithili","middleName":"","lastName":"Mehta","suffix":""},{"id":617761076,"identity":"ace5b193-01ac-4e59-ac86-835991018b39","order_by":6,"name":"Serena MacMillan","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Serena","middleName":"","lastName":"MacMillan","suffix":""},{"id":617761077,"identity":"0a5777a0-6c6c-43fc-97cd-686b63c78b6d","order_by":7,"name":"Sarah A Hosgood","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"A","lastName":"Hosgood","suffix":""},{"id":617761078,"identity":"8d012a49-abc9-4b51-b494-2d9ba090420b","order_by":8,"name":"Michael L Nicholson","email":"","orcid":"","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"L","lastName":"Nicholson","suffix":""},{"id":617761079,"identity":"8a30a920-f0c2-41e6-8fd1-e1cc404a4f3d","order_by":9,"name":"Neil S Sheerin","email":"","orcid":"","institution":"Newcastle University","correspondingAuthor":false,"prefix":"","firstName":"Neil","middleName":"S","lastName":"Sheerin","suffix":""},{"id":617761080,"identity":"63340d2c-2cc6-4c52-b1f3-e6768fdf858f","order_by":10,"name":"Colin H Wilson","email":"","orcid":"","institution":"Newcastle University","correspondingAuthor":false,"prefix":"","firstName":"Colin","middleName":"H","lastName":"Wilson","suffix":""}],"badges":[],"createdAt":"2026-03-27 08:42:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9242336/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9242336/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106235498,"identity":"d37272b6-e31c-481e-a3fd-4dc2f676b2c2","added_by":"auto","created_at":"2026-04-06 13:42:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":416140,"visible":true,"origin":"","legend":"\u003cp\u003eStudy overview. A) Conceptual overview of computational pipeline to generate predictions on post-transplant outcome from bulk RNA-Seq raw count data. B) Overview of the three study cohorts, showing timing of kidney biopsy, relative to other key pretransplant events.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/a1e345ad7290122064b154fb.png"},{"id":106235501,"identity":"d0767954-ebce-43be-ac2d-958c96505db1","added_by":"auto","created_at":"2026-04-06 13:42:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":148164,"visible":true,"origin":"","legend":"\u003cp\u003ePredictive performance for early post-transplant kidney function (defined as Delayed Graft Function; DGF) in the NMP Trial cohort. The discrimination (measured using c-statistic) and overall performance (Brier score) for each of the tested models are shown in panel A and B, and are ranked best to worst. Models using the full BulkFormer-TabPFN pipeline are shown in green, and are labelled with which BulkFormer method was used for aggregation of gene-level embeddings (e.g. Max absolute). Models using only TabPFN directly on the transcriptomic data (without BulkFormer) are shown in orange, and labelled with which genes were used, and whether the TabPFN default or high feature count (XL model) were used. Error bars represent 95% confidence intervals. C) Receiver-operating characteristic curves for the top three performing models by c-statistic. D) Distribution of predicted probabilities stratified by subsequent delayed graft function status.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/7af7d3367bd23e176f3fb2f2.png"},{"id":106235499,"identity":"db4232fd-4a1a-41c7-8f60-ceb896c24fa4","added_by":"auto","created_at":"2026-04-06 13:42:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":129029,"visible":true,"origin":"","legend":"\u003cp\u003epredicted pseudoprobability of delayed graft function (DGF) in the Cytokine Filter cohort. Each dot is a single kidney, with lines connecting pairs of kidneys. The predictions were generated based on the NMP Trial samples, using the BulkFormer TabPFN pipeline with the maximum absolute aggregation method. Panels represent cortex biopsy samples taken at 2 hours and 4 hours of normothermic machine perfusion (A and B respectively). DGF = delayed graft function.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/2635d90f1fce1c9647295eae.png"},{"id":106235497,"identity":"02210039-6644-4bd1-83f0-c8a595ade51d","added_by":"auto","created_at":"2026-04-06 13:42:20","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":206471,"visible":true,"origin":"","legend":"\u003cp\u003eComparing model pipeline predictive performance in the normothermic machine perfusion cohort versus the cold storage cohort (QUOD retrieval biopsy). The discrimination (measured using c-statistic) and overall performance (Brier score) for BulkFormer-TabPFN models are shown in panel A and B, labelled with the BulkFormer gene-level embedding aggregation method. Predictive performance in the NMP and QUOD cohort are represented by red and blue bars respectively. Error bars represent 95% confidence intervals. C) Receiver-operating characteristic curves for BulkFormer-TabPFN models, stratified by study cohort. D) Distribution of predicted probabilities stratified by subsequent delayed graft function status, with facets for aggregation method and study cohort (red = NMP, blue = QUOD).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/29c04c0a8c091d6ba209585a.png"},{"id":106402974,"identity":"76ffc0f1-41af-4c87-95aa-f0798c60b570","added_by":"auto","created_at":"2026-04-08 09:13:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1380697,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/24b58508-e9d0-4964-a673-00fc86b5322d.pdf"},{"id":106235496,"identity":"2989214e-366c-489c-a57e-c2447d2aa1f6","added_by":"auto","created_at":"2026-04-06 13:42:20","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":329733,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary figure caption\u003c/p\u003e\n\u003cp\u003eSupplementary Figure 1 – Predictive performance for early post-transplant kidney function (defined as Delayed Graft Function; DGF) in the QUOD retrieval biopsy cohort. The discrimination (measured using c-statistic) and overall performance (Brier score) for each of the tested models are shown in panel A and B, and are ranked best to worst. Models using the full BulkFormer-TabPFN pipeline are shown in green, and are labelled with which BulkFormer method was used for aggregation of gene-level embeddings (e.g. Max absolute). Models using only TabPFN directly on the transcriptomic data (without BulkFormer) are shown in orange, and labelled with which genes were used, and whether the TabPFN default or high feature count (XL model) were used. C) Receiver-operating characteristic curves for the top three performing models by c-statistic. D) Distribution of predicted probabilities stratified by subsequent delayed graft function status.\u003c/p\u003e\n\u003cp\u003eSupplementary table caption\u003c/p\u003e\n\u003cp\u003eSupplementary table 1 – All model performance metrics for all models in the NMP and QUOD cohorts. For models using the BulkFormer-TabPFN pipeline (BF-TabPFN) the labels reflect the BulkFormer method for aggregating gene-level embeddings. For models using the just TabPFN on the raw transcript per million data (without BulkFormer), the label represents which genes were used, and whether the TabPFN default or high feature count (XL model) were used.\u003c/p\u003e","description":"","filename":"Supplementarydata.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9242336/v1/2bbf28a8355e9c7881f14644.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Combining BulkFormer and TabPFN to predict post- transplant function from kidney biopsies during machine perfusion or cold storage","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLack of certainty around how a donated organ will function after transplantation frequently leads to organ non-use, contributing to a global shortage of transplants.\u003csup\u003e1\u003c/sup\u003e A technique termed normothermic machine perfusion (NMP), where a warm blood-based solution is circulated through an isolated organ to maintain it outside of the body, provides an opportunity to assess organ health.\u003csup\u003e2–4\u003c/sup\u003e However, more accurate measures of kidney viability during NMP are urgently needed.\u003csup\u003e4,5\u003c/sup\u003e Previous studies have identified transcriptomic signatures during kidney NMP that are associated with poor graft outcomes post-transplant, highlighting the potential for using transcriptomic analysis in this setting.\u003csup\u003e6,7\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHowever, using transcriptomic data for prediction is challenging for several reasons. Firstly, data is generally available for relatively few samples, compared to the large clinical cohorts underpinning most prediction models.\u003csup\u003e8\u003c/sup\u003e Secondly, there is a very large number of features, with information on tens of thousands of gene transcripts per sample. Standard statistical prediction modelling has to reduce this large number of genes to select a set of promising genes to use for modelling; a process where overfitting and overoptimism are virtually guaranteed, especially in small datasets.\u003csup\u003e9–11\u003c/sup\u003e In addition, this manual process is inherently variable and analyst-dependent, introducing potential bias in feature selection and reducing reproducibility.\u003c/p\u003e\n\u003cp\u003eCombining BulkFormer and TabPFN v2.5 may allow us to overcome these issues.\u003csup\u003e12–14\u003c/sup\u003e These are both state of the art foundation models which use a transformer-based architecture; the same architecture type underpinning large language models such as ChatGPT. The models take in data as a ‘prompt’, and output distilled information or predictions, without requiring any model retraining; this is termed “zero-shot” predictions. TabPFN has been at the heart of a recent revolution in the analysis of tabular data (data organised in rows and columns).\u003csup\u003e15\u003c/sup\u003e It is a state of the art model, outperforming established machine learning approaches for tabular data, such as XGBoost and random forests.\u003csup\u003e13\u003c/sup\u003e TabPFN is especially strong when sample size is low, making it an attractive option for use with current transcriptomic datasets. However, performance has been reported to degrade with high-dimensional data, especially when the number of features exceeds 2,000.\u003csup\u003e13,16\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eStandard approaches to dimensionality reduction often lose important data, or rely on relationships between predictors and outcome thereby potentially amplifying noise leading to overfitting. BulkFormer is a foundation model trained on over 500,000 bulk transcriptomic samples.\u003csup\u003e14\u003c/sup\u003e It can process transcriptomic data on 20,010 genes and distil this information into a compact set of embeddings, each of which reflects a key biological programme. This distilled data can then be used for a range of tasks, including disease annotation, patient prognosis prediction and drug response prediction.\u003csup\u003e14\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eDespite these conceptual advantages to deal with the specific challenges posed by transcriptomics data sets, there are no published articles describing the combination of BulkFormer and TabPFN v2.5. Our aim was therefore to evaluate the performance of a novel combination BulkFormer-TabPFN pipeline to make outcome predictions based on raw transcriptomic count data.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis is an exploratory study to assess the ability of a BulkFormer-TabPFN pipeline to generate clinical outcome predictions from small sample-size transcriptomic datasets. All data and code used in this study is available in our accompanying GitHub repository.\u003csup\u003e17\u003c/sup\u003e Ethical approval was not required as all raw data (RNA-Seq raw counts and metadata) was retrieved from previous publications, either via the gene expression omnibus (GEO) or direct contact with the study authors. This raw count data, along with its metadata, is also available in our GitHub repository.\u003csup\u003e17\u003c/sup\u003e A summary of the computing pipeline is given in Figure 1A. All computing steps shown were performed zero-shot, without any model retraining.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOur primary clinical outcome was delayed graft function (DGF), often defined as the need for dialysis within the first week post-transplant. This measure of early post-transplant kidney function has been shown to correlate with long-term kidney survival, and is frequently used as the primary outcome in both preclinical and clinical research.\u003csup\u003e5,6,18,19\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs dialysis in the first day post-transplant often reflects hyperkalaemia, rather than truly reflecting function of the transplanted kidney, recipients receiving dialysis only during the first 24 hours post-transplant (with no further dialysis in the first week), were classified as not having DGF.\u0026nbsp;Therefore, our definition of DGF was requirement for dialysis in the first week post-transplant, outside of the first 24 hours. This approach is consistent with prior work demonstrating that transplants where dialysis is only performed in the first 24 hours have transcriptomic profiles comparable to kidneys where no post-transplant dialysis is performed.\u003csup\u003e6,7\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eSummary of study cohorts and methodology\u003c/h2\u003e\n\u003cp\u003eA summary of the three study cohorts is given in Figure 1B and Table 1. All three cohorts include bulk RNA-Seq raw count data from renal cortex tissue biopsies of human kidneys donated for transplantation. For all cohorts, tissue was homogenised, and RNA isolated with the pure link RNA mini kit (Ambion). Library preparation was performed using the TruSeq Stranded total RNA library prep kit (Illumina), and sequenced on a Hiseq sequencer (Illumina). Alignment to the human genome (Hg38) was performed with HISAT2, and featurecounts was used to generate the raw count data.\u003csup\u003e6,18\u003c/sup\u003e Further details on the study cohorts, and on precise methodologies for generating the raw count data, can be found in the original publications.\u003csup\u003e6,18\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eThe NMP trial cohort consists of biopsies performed at 1 hour of NMP as part of the United Kingdom NMP randomised trial.\u003csup\u003e5\u003c/sup\u003e The data for N=34 kidneys was kindly provided by Ferdinand et al, and included the 33 kidneys included in their manuscript, plus an additional case. All biopsies had linked DGF data, in-line with our definition above.\u003csup\u003e6\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe Cytokine Filter cohort included paired human kidney experiments (N=10 kidneys). These were performed on human kidneys retrieved for transplant but then deemed unsuitable and offered for research. One kidney per pair received 4 hours of NMP plus cytokine filter treatment, and the other kidney was an NMP-only control. None of these kidneys were transplanted. Raw count data was taken from GEO (accession GSE121447), with additional metadata provided in the associated manuscript.\u003csup\u003e6\u003c/sup\u003e This cohort allowed us to assess whether DGF predictions were modifiable by therapeutic intervention during NMP, with the paired design being a particular strength.\u003csup\u003e6\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eThe Quality in Organ Donation (QUOD) retrieval biopsy cohort contains biopsies which were taken at the time of kidney organ retrieval (following cold flush of the organs), prior to static cold storage and transport.\u003csup\u003e18\u003c/sup\u003e Raw count data was taken from GEO (accession GSE215108), with metadata taken from manuscript supplementary data.\u003csup\u003e18\u003c/sup\u003e Of the available biopsies, 263 had sufficient metadata to allow labelling of DGF in line with our definition above.\u003c/p\u003e\n\u003cp\u003eThe NMP and QUOD retrieval biopsy cohorts were analysed and evaluated entirely independently, with no cross-cohort prediction between NMP and QUOD retrieval biopsy cohorts.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eProcessing of raw count data\u003c/h2\u003e\n\u003cp\u003eThe raw count data for all three cohorts was indexed using gene symbols, rather than the Ensembl IDs required for BulkFormer. These gene symbols were mapped to Ensembl gene IDs using Ensembl BioMart (using biomaRt; \u003cem\u003ehsapiens_gene_ensembl\u003c/em\u003e).\u003csup\u003e20–22\u003c/sup\u003e Symbols not mapped by BioMart were resolved via the HGNC REST API to identify approved symbols from aliases/previous symbols, and these approved\u0026nbsp;symbols were then re-mapped to Ensembl gene IDs via biomaRt.\u003csup\u003e22,23\u003c/sup\u003e The mapping file was created on 11\u003csup\u003eth\u003c/sup\u003e Feb 2026 using the most recent Ensembl (release 115) and HGNC versions, and is available on our GitHub.\u003csup\u003e17,20\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eWhere a single gene symbol mapped to multiple Ensembl IDs, we prioritised Ensembl IDs that are used in the BulkFormer pipeline (either as one of the input genes or a gene with length information for their transcript per million, TPM, calculation).\u003csup\u003e24,25\u003c/sup\u003e Counts for any duplicate Ensembl ID were summed. Any gene symbols which could not be mapped to Ensembl ID used by BulkFormer were initially retained for TPM calculations. Raw count data was then converted to TPM and log-transformed to log\u003csub\u003ee\u003c/sub\u003e(TPM); this was performed using code provided in the BulkFormer GitHub, to ensure consistency with the data used for previous BulkFormer training.\u003csup\u003e24\u003c/sup\u003e\u003c/p\u003e\n\u003ch2\u003eGenerating BulkFormer Embeddings\u003c/h2\u003e\n\u003cp\u003eWe used the BulkFormer (version 2) 147M model to generate embeddings.\u003csup\u003e14,24,25\u003c/sup\u003e The BulkFormer GitHub repository was cloned on 10\u003csup\u003eth\u003c/sup\u003e January 2026, and all associated files (including model weights and gene annotation files) were downloaded on the same day from Zenodo.\u003csup\u003e25\u003c/sup\u003e The code for generating embeddings was taken directly from the “bulkformer_extract_feature.ipynb” Jupyter notebook.\u003c/p\u003e\n\u003cp\u003eThe key area where the BulkFormer code was modified was the method for aggregating gene-level to sample level embeddings. BulkFormer generates latent space embeddings (n=643 for the model used here) per gene, per sample. For prediction tasks, gene-level embeddings need to be aggregated to sample-level embeddings, such that each sample is now represented by n=643 unique embeddings. The available options for aggregation were: maximum (used in BulkFormer manuscript)\u003csup\u003e14\u003c/sup\u003e, mean and median. We have added an option for maximum absolute value, so that large negative embeddings are also captured. As sensitivity analyses we also added options for using the 99.9\u003csup\u003eth\u003c/sup\u003e percentile rather than the maximum (attempting to make the maximum aggregation approaches robust to potential single gene outliers). These new options are available on our GitHub.\u003csup\u003e17\u003c/sup\u003e\u003c/p\u003e\n\u003ch2\u003eTabPFN v2.5\u003c/h2\u003e\n\u003cp\u003eTabPFN v2.5 was used with model weights downloaded from Hugging Face (a repository for model weights) on 19th February 2026.\u003csup\u003e26\u003c/sup\u003e We utilized the default model (tabpfn-v2.5-classifier-v2.5_default.ckpt) for all feature sets and additionally evaluated the XL model variant (tabpfn-v2.5-classifier-v2.5_large-features-XL.ckpt) for log(TPM) data. The models operate by generating predictions conditioned on contextual examples from the input data, rather than undergoing traditional parameter retraining.\u003c/p\u003e\n\u003cp\u003eThe input to these models was either: (1) BulkFormer sample-level embeddings (n=643 features) generated as described above, and (2) log(TPM) values containing either all genes or only the 20,010 genes used for BulkFormer inference; this was done to assess the impact of omitting the BulkFormer step.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe evaluated model performance using leave-one-out cross-validation (LOO-CV) within the NMP trial and QUOD retrieval biopsy cohorts separately. Within each cohort, outcome-labelled data were provided to the pipeline as context, with an outcome value for a single kidney (occurrence of DGF) masked and predicted by the model. This process of masking a single kidney’s outcome was repeated for all kidneys within the cohort, allowing comparison of predicted and observed outcomes to assess predictive performance. This was performed in the NMP trial and QUOD cohorts completely separately (neither cohort used any information from the other cohort in any way).\u003c/p\u003e\n\u003cp\u003eWe also generated the predicted probability of DGF for each sample in the cytokine filter cohort, using the NMP trial cohort as context. This was performed to assess whether DGF predictions were dynamic, and whether they could be influenced by novel therapies. We refer to these as “pseudoprobabilities” because these biopsies were taken at different timepoints of NMP compared with the NMP trial cohort. Therefore, they do not represent cohort-calibrated estimated true probabilities. Higher values indicate a greater predicted probability of DGF compared with other cases in the dataset, but the numerical values themselves should not be assumed to represent absolute probabilities.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eStatistical analysis\u003c/h2\u003e\n\u003cp\u003eStatistical analysis and reporting were conducted in line with recent best-practice guidance for reporting binary prediction model performance in medical AI research.\u003csup\u003e27\u003c/sup\u003e Model discrimination was assessed using the c-statistic (area under the receiver operating characteristic curve, AUC), which was prespecified as the primary performance metric. Overall predictive performance was evaluated using the Brier score.\u003csup\u003e27\u003c/sup\u003e Confidence intervals (CI) were calculated using a bootstrap approach for all metrics, except for the c-statistic, where the DeLong method implemented in the pROC R library was used.\u003csup\u003e28\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor visualisation, ROC curves were plotted to illustrate discrimination, and violin plots were used to display the distribution of performance estimates across feature representations and aggregation strategies. The code used for performance evaluation and visualisation was adapted from the best-practice guidance paper’s accompanying GitHub repository.\u003csup\u003e29\u003c/sup\u003eTo assess the difference in pseudoprobabilities of DGF in the paired kidney experiments (Cytokine Filter cohort) we used a two-tailed paired t-test.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe novel computing pipeline is displayed in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA, and a summary of study cohorts is given in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB. Kidney cohort sample sizes were: 34 for the NMP trial samples (9/34 with DGF), 10 for the Cytokine Filter preclinical cohort (5 pairs) and 263 for the QUOD retrieval biopsy cohort (73/263 with DGF). Full details on every tested model is given in Supplementary Table\u0026nbsp;1, and includes all predictive model metrics recommended by Calster et al.\u003csup\u003e27\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eCombining BulkFormer and TabPFN allows DGF prediction\u003c/p\u003e \u003cp\u003eWe initially analysed the NMP trial cohort. Our BulkFormer-TabPFN pipeline allowed RNA-Seq raw count data from biopsies taken at 1 hour of normothermic machine perfusion to be used to predict early post-transplant function, as measured by delayed graft function (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAssessing various options for aggregation of BulkFormer gene-level to sample-level embeddings revealed that maximum absolute aggregation resulted in the best model performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA\u0026amp;B). Maximum absolute and maximum aggregation produced models with C-statistics whose 95% confidence intervals excluded 0.5, indicating discrimination significantly better than chance, whereas mean and median aggregation did not. As a sensitivity analysis we also aggregated embeddings using 99.9th percentile, to see whether maximum embeddings were overly influenced by outliers; these models showed inferior predictive performance (Supplementary Table\u0026nbsp;1).\u003c/p\u003e \u003cp\u003eWe next assessed predictive performance of a pipeline which did not use BulkFormer and instead applied TabPFN directly to the transcriptomic data (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA). When the BulkFormer step was omitted, predictive performance was lost, with TabPFN-only pipelines generating predictions that were worse than chance (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA\u0026amp;B). This was true for both the default TabPFN v2.5 models, and TabPFN models specifically trained for high feature count data (the v2.5 XL models).\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC displays receiver-operating characteristic curves for the top three performing model pipelines, based on c-statistic. The best overall model was the BulkFormer-TabPFN model which used maximum absolute embedding aggregation (c-statistic\u0026thinsp;=\u0026thinsp;0.82, 95% CI\u0026thinsp;=\u0026thinsp;0.67\u0026ndash;0.97; Brier score\u0026thinsp;=\u0026thinsp;0.15, 95% CI\u0026thinsp;=\u0026thinsp;0.08\u0026ndash;0.22). The distribution of predicted probabilities of DGF, stratified by actual DGF status, is displayed in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD.\u003c/p\u003e \u003cp\u003ePredicted DGF probability is dynamic and modulated by a novel therapy\u003c/p\u003e \u003cp\u003eTo assess whether the predicted DGF probability from this pipeline is sensitive to therapeutic modulation of the kidney, we used a cohort of biopsies from a preclinical study of a novel therapy.\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e This consisted of paired kidneys (i.e. both kidneys from the same donor) undergoing normothermic machine perfusion with or without cytokine filter treatment.\u003c/p\u003e \u003cp\u003eUsing embedding and outcome data from the NMP trial cohort as the context, we generated zero-shot predicted probability of DGF for each of the Cytokine Filter kidney biopsies (displayed in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). As the machine perfusion timepoints differed between groups (1 hour in NMP trial cohort; 2 and 4 hours in cytokine filter cohort), we have labelled this as a pseudoprobability. At the end of cytokine filter treatment (4 hours), there was a significant reduction in predicted pseudoprobability of DGF in the Cytokine Filter group (mean difference in DGF pseudoprobability = -21.3%, 95% CI = -38.7% to -4.0%, paired t-test p\u0026thinsp;=\u0026thinsp;0.027) compared to the control group. This occurred despite no difference between groups at the 2-hour timepoint (mean difference in DGF pseudoprobability\u0026thinsp;=\u0026thinsp;+\u0026thinsp;4.7%, 95% CI = -24.8% to 34.2%, p\u0026thinsp;=\u0026thinsp;0.683). Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e displays these results.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBiopsy during NMP allows better predictions than biopsies during cold storage\u003c/p\u003e \u003cp\u003eWe next assessed whether the same pipeline (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA) could be used to predict DGF, in an independent cohort of biopsies taken from kidneys flushed with cold preservation solution (thus at 4\u003csup\u003eo\u003c/sup\u003eC temperature) at the time of organ retrieval (QUOD cohort), as opposed to kidney biopsies taken during normothermic machine perfusion (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB). This analysis was conducted entirely within the QUOD cohort, with no use of NMP trial samples at any stage, including as \u0026lsquo;context\u0026rsquo;. Full model performance metrics in the retrieval biopsy cohort are provided in Supplementary Fig.\u0026nbsp;1 and Supplementary Table\u0026nbsp;1.\u003c/p\u003e \u003cp\u003eMirroring the results in the NMP cohort, the BulkFormer-TabPFN model incorporating maximum absolute embedding aggregation ranked highest by c-statistic (Supplementary Fig.\u0026nbsp;1A). Within the QUOD retrieval biopsy cohort, only BulkFormer-TabPFN using absolute and maximum aggregation produced models with C-statistics whose 95% confidence intervals excluded 0.5. Consistent with the NMP findings, models omitting the BulkFormer step demonstrated no evidence of predictive discrimination.\u003c/p\u003e \u003cp\u003eWith an identical model pipeline, outcome definition, and an identical transcriptomic pipeline performed in the same laboratory,\u003csup\u003e6,18\u003c/sup\u003e predictive performance was significantly higher when using NMP biopsies versus the QUOD kidney biopsies (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). For example, with the maximum absolute aggregation approach, discrimination was strong in the NMP cohort (c-statistic\u0026thinsp;=\u0026thinsp;0.82, 95% CI\u0026thinsp;=\u0026thinsp;0.67\u0026ndash;0.97) but substantially lower in the QUOD cohort (c-statistic\u0026thinsp;=\u0026thinsp;0.59, 95% CI\u0026thinsp;=\u0026thinsp;0.52\u0026ndash;0.67).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe show for the first time the ability of an AI pipeline combining BulkFormer Version 2 and TabPFN v2.5 to generate outcome predictions from raw transcriptomic data. In the setting of the small datasets examined here, the addition of BulkFormer for context-aware distillation of gene data into embeddings significantly improved predictive performance, compared with pipelines using TabPFN alone. Predictive performance was sensitive to the method used to aggregate gene-level into sample-level embeddings; we found that maximum absolute aggregation of BulkFormer embeddings was optimal, a method which we believe has not been described previously.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e All code is freely available on GitHub and includes tools enabling readers to apply the pipeline to their own raw count data.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eDue to the small sample size, and single centre collection/processing of samples, this work is exploratory and is clearly not a clinical prediction model which is ready for deployment. Nevertheless, there are important messages and opportunities for transplantation. Firstly, this adds to the body of evidence for improved organ assessment using NMP;\u003csup\u003e2\u0026ndash;4\u003c/sup\u003e with our pipeline, transcriptomic signatures during NMP led to superior predictions compared with transcriptomic signatures at retrieval. Importantly, the NMP trial and QUOD retrieval biopsy cohorts were analysed completely independently, with no sharing of samples, features, or contextual information at any stage. QUOD cohort predictions were not informed by the NMP cohort, and the observed differences in performance are due to the underlying biological signal themselves.\u003c/p\u003e \u003cp\u003eSecondly, we have provided evidence that these predictions are responsive to the delivery of a novel NMP therapy (Cytokine filters).\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e With larger cohorts of outcome-linked biopsy data, this could be a potential surrogate endpoint in preclinical or early-phase trials, which could potentially reduce the need for larger, more expensive clinical trials in the future. As a continuous variable it offers increased power and lower sample sizes in comparison to binary endpoints, and it is intrinsically tied to post-transplant outcome. This is particularly important given the substantial volume of preclinical NMP therapeutic studies, where robust, translatable efficacy readouts are currently limited.\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e Although larger cohorts of outcome-linked data are required, the code we provide\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e acts as a blueprint for the generation of DGF predicted probabilities from kidney biopsy raw count data.\u003c/p\u003e \u003cp\u003eTranscriptomic assessment of kidney biopsies has entered select clinical practice for post-transplant biopsies, most notably with the Molecular Microscope Diagnostics System (MMDx).\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e Although sequencing costs have reduced substantially (\u0026lt;\u003cspan\u003e$\u003c/span\u003e150 for bulk RNA-Seq), turn-around time remains a hurdle in the time-pressured setting of organ assessment. Advances in next-generation sequencing with nanopore have already been reported to provide sample-to-answer times of just 6 hours.\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e As RNA-Seq workflows continue to improve, and technologies for extending organ preservation time evolve,\u003csup\u003e34\u003c/sup\u003e transcriptomic assessment could one day guide organ accepting decisions.\u003csup\u003e\u003cspan additionalcitationids=\"CR35 CR36\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThis manuscript provides a blueprint for how AI foundation models can be combined to make rapid predictions from raw count data, without manual bioinformatic analysis. A key advantage of our pipeline is that it performs predictions \u0026lsquo;zero-shot\u0026rsquo;, meaning that no model retraining is performed, and none of the model weights are changed. This is possible as both BulkFormer and TabPFN are foundation models. They use transformer-based architectures, and train on massive amounts of data.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e After this training, they can then take new labelled data as a \u0026lsquo;prompt\u0026rsquo; and provide predictions with a single forward pass through the network, akin to how a large language model can take text as a prompt and provide a response (without retraining the model weights).\u003c/p\u003e \u003cp\u003eThis contrasts with standard approaches to generate predictions from RNA-Seq data. Standard statistical methods generally rely on feature selection (selection of \u0026lsquo;significant\u0026rsquo; genes) prior to subsequent modelling. This process is inherently unstable, and in small datasets overfitting and overoptimism are virtually guaranteed.\u003csup\u003e\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e These issues are magnified if variable/feature selection is performed using association with the outcome being predicted.\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e,\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e Even options for dimensionality reduction such as principal components can distort inference by inflating type I error, whilst losing biologically relevant signal.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e,\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e Established machine-learning methods, including random forests and gradient boosting, remain vulnerable to the combined challenges of high feature dimensionality, noise, and small sample size, often resulting in unstable models which are overfitted to noise.\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e,\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e For these reasons, we have deliberately avoided comparison of our zero-shot approach, to any methods which require retraining, as they would not provide a meaningful benchmark.\u003c/p\u003e \u003cp\u003eTabPFN is trained to approximate Bayesian inference, and has demonstrated state-of-the-art performance for tabular data predictions across a wide range of disciplines, outperforming bespoke models using established machine learning approaches.\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e In addition to improved predictive performance, TabPFN offers practical advantages: prediction is rapid as model retraining is entirely avoided, implementation is often simpler as hyperparameter tuning is not required, and performance is particular strong in the setting of low sample-size.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e One limitation of TabPFN is its significantly decreased performance with data that exceeds 2,000 dimensions.\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e This is clearly a challenge for transcriptomic data, where the number of genes exceeds 10,000\u0026ndash;20,000.\u003c/p\u003e \u003cp\u003eThis is where we turned to BulkFormer version 2,\u003csup\u003e25\u003c/sup\u003e which utilises a graph neural-network to incorporate known gene-gene interactions, and a performer-based architecture which was trained on over 500,000 bulk RNA-Seq samples.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e It takes RNA-Seq data on 20,010 genes, and distils this into a compressed set of embeddings which represent biological signal, whilst minimising information loss. In the original BulkFormer manuscript these embeddings were fed into a random forest model to generate predictions.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e BulkFormer outperformed existing single-cell trained models on a range of tasks including patient prognosis prediction.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe benefits of these two models are complementary for transcriptomic data. TabPFN performs well across multiple disciplines, especially in the setting of small sample sizes, however struggles with high feature count data. BulkFormer allows the number of features to be reduced, whilst preserving biological signal. Our findings support these conceptual benefits. TabPFN would also allow very simple integration of donor characteristics into the same model, as additional features, which may further improve predictions.\u003c/p\u003e \u003cp\u003eDespite these advantages, and growing use of both models individually, to our knowledge only a single pre-print article has described the combination of BulkFormer and TabPFN.\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e This found no benefit of integrating BulkFormer with TabPFN. There are several potential reasons for the difference in conclusions with this study. Firstly, they aggregated gene-level embeddings using mean aggregation only; an approach which also failed to improve predictions in our manuscript. They also used older versions of both BulkFormer and TabPFN, and used datasets with over 10,000 samples (which is not where TabPFN shows the most benefit).\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThis study falls outside of the standard paradigm of internal and external validation of clinical prediction models, as it uses zero-shot models.\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e As there is no model retraining, many of the standard approaches for internal validation studies, such as optimism calculation/correction, do not apply. However, this does not represent true external validation, as the study cohort used as the input is still used for generating predictions.\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e The most recent TRIPOD\u0026thinsp;+\u0026thinsp;AI guidelines explicitly excluded foundational models, and the TRIPOD-LLM guidelines do not cover tabular models. In the future new guidelines will be required to cover the application of foundational tabular models for clinical prediction.\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe main limitations relate to the retrospective nature of the study, and reliance on available sample numbers and metadata. Overall, there was a small available sample size of biopsy data linked to post-transplant outcome, with minimal metadata on other important clinical factors (such as donor age). As a result it is unclear how much this transcriptomic data adds to standard clinical factors. However, our BulkFormer-TabPFN pipeline achieved superior predictive discrimination (c-statistic) than DGF predictions models using clinical data, and predictions were modifiable by treatment (in kidney pairs with identical clinical variables).\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e Whilst combining BulkFormer-TabPFN improved performance compared with TabPFN alone, combining two models amplifies the \u0026ldquo;black-box\u0026rdquo; nature of foundation models, making it challenging to attribute predictions to specific genes.\u003c/p\u003e \u003cp\u003eAnother limitation is that all biopsies from all cohorts were processed by the same laboratory and subjected to an identical RNA-Seq generation pipeline, which may limit generalisability and transferability to external cohorts. In addition, NMP biopsies linked to outcome data were only available at the 1-hour timepoint. As transcriptomic profiles evolve over the course of NMP,\u003csup\u003e6\u003c/sup\u003e the applicability of models derived from early biopsies to samples obtained after prolonged perfusion remains uncertain and warrants further investigation. The benefit of our approach is that as larger outcome-linked RNA-Seq datasets become available, they can be incorporated into our pipeline, allowing iterative evaluation of generalisability, consistency, transportability and predictive performance.\u003c/p\u003e \u003cp\u003eIn conclusion, transcriptomic data poses a unique challenge due to the high number of genes, relative to the often small sample size. Standard methods for prediction rely on feature selection, which leads to overfitting and optimism about model performance. Here we combine two state of the art artificial intelligence models, BulkFormer and TabPFN. These transformer-based foundation models can generate predictions \u0026ldquo;zero-shot\u0026rdquo;, without retraining on new data. This manuscript is the first to demonstrate the synergistic benefits of combining these models in terms of predictive performance. Our freely available code offers a blueprint to apply this pipeline to generate predictions from raw RNA-Seq, which could be applied in a wide range of contexts beyond deceased donor kidney transplantation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThank you to Professor Menna Clatworthy, Benjy Tan and John Ferdinand for their help in accessing the raw data used in this study. We would also like to thank Vasilis Kosmoliaptsis for the help of his laboratory team in acquiring raw data, and for providing comments on the manuscript contents.\u003c/p\u003e\n\u003cp\u003eFunding declaration\u003c/p\u003e\n\u003cp\u003eSJT was funded for this work via a Medical Research Council Clinical Research Training Fellowship (MRC/Y000676/1), which was part-funded by Kidney Research UK. This study is funded by the National Institute for Health and Care Research (NIHR) Blood and Transplant Research Unit in Organ Donation and Transplantation (NIHR203332), a partnership between the National Health Service (NHS) Blood and Transplant, the University of Cambridge, and Newcastle University. The views expressed are those of the author(s) and not necessarily those of the NIHR, NHS Blood and Transplant, or the Department of Health and Social Care.\u003c/p\u003e\n\u003cp\u003eChatGPT v5.2 was used for copy editing only, after the first draft was finalised. The authors remain accountable for all aspects of the work.\u003c/p\u003e\n\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eStudy concept: SJT\u003c/li\u003e\n \u003cli\u003eStudy design: SJT, GK\u003c/li\u003e\n \u003cli\u003eCuration of existing data sources: SJT\u003c/li\u003e\n \u003cli\u003eBuilding BulkFormer-TabPFN pipeline: SJT, GK\u003c/li\u003e\n \u003cli\u003eData analysis: SJT, GK\u003c/li\u003e\n \u003cli\u003eInterpretation of results: All authors\u003c/li\u003e\n \u003cli\u003eInitial drafting of manuscript: SJT, GK\u003c/li\u003e\n \u003cli\u003eCritical review and editing of manuscript: All authors\u003c/li\u003e\n \u003cli\u003eReview of final manuscript version: All authors\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eData availability statement\u003c/p\u003e\n\u003cp\u003eAll data and code are freely available open access via our GitHub repository for the project https://github.com/TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCompeting Interests Statement\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eOlawade, D. B., Marinze, S., Qureshi, N., Weerasinghe, K. \u0026amp; Teke, J. Transforming organ donation and transplantation: Strategies for increasing donor participation and system efficiency. \u003cem\u003eEur. J. Intern. Med.\u003c/em\u003e \u003cb\u003e133\u003c/b\u003e, 14\u0026ndash;24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ejim.2024.11.010\u003c/span\u003e\u003cspan address=\"10.1016/j.ejim.2024.11.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJaynes, C. L., Goggins, W. C., Holzner, M. L., Garonzik-Wang, J. \u0026amp; Leuvenink, H. G. D. No Kidney Left Behind: Rescuing Unused Donor Kidneys for Transplant at the First Centralized Assessment and Repair Center. \u003cem\u003eTranspl. Int.\u003c/em\u003e \u003cb\u003e38\u003c/b\u003e, 15424. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/ti.2025.15424\u003c/span\u003e\u003cspan address=\"10.3389/ti.2025.15424\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosgood, S. A., Barlow, A. D., Hunter, J. P. \u0026amp; Nicholson, M. L. Ex vivo normothermic perfusion for quality assessment of marginal donor kidney transplants. \u003cem\u003eBr. J. Surg.\u003c/em\u003e \u003cb\u003e102\u003c/b\u003e (11), 1433\u0026ndash;1440. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/bjs.9894\u003c/span\u003e\u003cspan address=\"10.1002/bjs.9894\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTingle, S. J. et al. Contrast-Enhanced Ultrasound to Assess Kidney Quality During Ex Situ Normothermic Machine Perfusion. \u003cem\u003eTranspl. Int.\u003c/em\u003e \u003cb\u003e38\u003c/b\u003e, 14268. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/ti.2025.14268\u003c/span\u003e\u003cspan address=\"10.3389/ti.2025.14268\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosgood, S. A. et al. Normothermic machine perfusion versus static cold storage in donation after circulatory death kidney transplantation: a randomized controlled trial. \u003cem\u003eNat. Med.\u003c/em\u003e \u003cb\u003e29\u003c/b\u003e (6), 1511\u0026ndash;1519. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-023-02376-7\u003c/span\u003e\u003cspan address=\"10.1038/s41591-023-02376-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerdinand, J. R. et al. Cytokine absorption during human kidney perfusion reduces delayed graft function-associated inflammatory gene signature. \u003cem\u003eAm. J. Transpl.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e (6), 2188\u0026ndash;2199. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/ajt.16371\u003c/span\u003e\u003cspan address=\"10.1111/ajt.16371\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpiers, H. V. M. et al. Transcriptomic profiling during normothermic machine perfusion of human kidneys reveals a pro-inflammatory cellular landscape and gene expression signature associated with severe ischemia-reperfusion injury and delayed graft function. \u003cem\u003eFront. Immunol.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fimmu.2025.1679251\u003c/span\u003e\u003cspan address=\"10.3389/fimmu.2025.1679251\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRao, P. S. et al. A comprehensive risk quantification score for deceased donor kidneys: the kidney donor risk index. \u003cem\u003eTransplantation\u003c/em\u003e \u003cb\u003e88\u003c/b\u003e (2), 231\u0026ndash;236. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/TP.0b013e3181ac620b\u003c/span\u003e\u003cspan address=\"10.1097/TP.0b013e3181ac620b\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClarke, R. et al. The properties of high-dimensional data spaces: implications for exploring gene and protein expression data. \u003cem\u003eNat. Rev. Cancer\u003c/em\u003e. \u003cb\u003e8\u003c/b\u003e (1), 37\u0026ndash;49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nrc2294\u003c/span\u003e\u003cspan address=\"10.1038/nrc2294\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimon, R., Radmacher, M. D., Dobbin, K. \u0026amp; McShane, L. M. Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. \u003cem\u003eJ. Natl. Cancer Inst.\u003c/em\u003e \u003cb\u003e95\u003c/b\u003e (1), 14\u0026ndash;18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/jnci/95.1.14\u003c/span\u003e\u003cspan address=\"10.1093/jnci/95.1.14\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2003).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmbroise, C. \u0026amp; McLachlan, G. J. Selection bias in gene extraction on the basis of microarray gene-expression data. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e. ;99(10):6562\u0026ndash;6566. (2002). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1073/pnas.102102699\u003c/span\u003e\u003cspan address=\"10.1073/pnas.102102699\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHollmann, N. et al. Accurate predictions on small data with a tabular foundation model. \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e637\u003c/b\u003e (8045), 319\u0026ndash;326. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41586-024-08328-6\u003c/span\u003e\u003cspan address=\"10.1038/s41586-024-08328-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGrinsztajn, L. et al. TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models. \u003cem\u003earXiv\u003c/em\u003e. Preprint posted online February 5, 2026:arXiv:2511.08667. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2511.08667\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2511.08667\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang, B., Fan, R., Yi, M., Cui, C. \u0026amp; Cui, Q. A large-scale foundation model for bulk transcriptomes. \u003cem\u003ebioRxiv\u003c/em\u003e. Preprint posted online June 17, 2025:2025.06.11.659222. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2025.06.11.659222\u003c/span\u003e\u003cspan address=\"10.1101/2025.06.11.659222\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Q., Tan, Y. S., Tian, Q. \u0026amp; Li, P. TabPFN: One Model to Rule Them All? \u003cem\u003earXiv\u003c/em\u003e. Preprint posted online November 27, 2025:arXiv:2505.20003. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2505.20003\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2505.20003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYe, H. J., Liu, S. Y. \u0026amp; Chao, W. L. A Closer Look at TabPFN v2: Understanding Its Strengths and Extending Its Capabilities. \u003cem\u003earXiv\u003c/em\u003e. Preprint posted online June 11, 2025:arXiv:2502.17361. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2502.17361\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2502.17361\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTingle, S. J. \u0026amp; Kourounis, G. TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics: Data and code repository for project. GitHub. Accessed February 27, (2026). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics\u003c/span\u003e\u003cspan address=\"https://github.com/TxDataIncubator/BulkFormer_with_TabPFN_RNA-Seq_Kidney_Transcriptomics\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, R. et al. Assessment of biological organ age using molecular pathology in pre-transplant kidney biopsies. \u003cem\u003eKidney Int.\u003c/em\u003e \u003cb\u003e106\u003c/b\u003e (2), 302\u0026ndash;316. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.kint.2024.03.028\u003c/span\u003e\u003cspan address=\"10.1016/j.kint.2024.03.028\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVenkataraman, K., Irish, G. L., Collins, M. G. \u0026amp; Clayton, P. A. The Association Between Early Graft Function, Donor Type and Long-Term Kidney Transplant Outcomes. \u003cem\u003eTranspl. Int.\u003c/em\u003e \u003cb\u003e38\u003c/b\u003e, 14197. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/ti.2025.14197\u003c/span\u003e\u003cspan address=\"10.3389/ti.2025.14197\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDyer, S. C. et al. Ensembl 2025. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cb\u003e53\u003c/b\u003e (D1), D948\u0026ndash;D957. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/nar/gkae1071\u003c/span\u003e\u003cspan address=\"10.1093/nar/gkae1071\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDurinck, S. et al. BioMart and Bioconductor: a powerful link between biological databases and microarray data analysis. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e (16), 3439\u0026ndash;3440. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bioinformatics/bti525\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/bti525\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDurinck, S., Spellman, P. T., Birney, E. \u0026amp; Huber, W. Mapping identifiers for the integration of genomic datasets with the R/Bioconductor package biomaRt. \u003cem\u003eNat. Protoc.\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e (8), 1184\u0026ndash;1191. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nprot.2009.97\u003c/span\u003e\u003cspan address=\"10.1038/nprot.2009.97\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBraschi, B. et al. Genenames.org: the HGNC and VGNC resources in 2019. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cb\u003e47\u003c/b\u003e (D1), D786\u0026ndash;D792. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/nar/gky930\u003c/span\u003e\u003cspan address=\"10.1093/nar/gky930\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang, B. KangBoming/BulkFormer. Published online February 7, 2026. Accessed January 10, (2026). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/KangBoming/BulkFormer\u003c/span\u003e\u003cspan address=\"https://github.com/KangBoming/BulkFormer\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang, B., Fan, R., Cui, C. \u0026amp; Cui, Q. A large-scale foundation model for human bulk transcriptomes (version 2). \u003cem\u003ePublished online Dec.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5281/zenodo.15744294\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.15744294\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrior-Labs/tabpfn_2_5 \u0026middot; Hugging Face. November 11. Accessed February 27, 2026. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/Prior-Labs/tabpfn_2_5\u003c/span\u003e\u003cspan address=\"https://huggingface.co/Prior-Labs/tabpfn_2_5\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalster, B. V. et al. Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance. \u003cem\u003eLancet Digit. Health\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e (12). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.landig.2025.100916\u003c/span\u003e\u003cspan address=\"10.1016/j.landig.2025.100916\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobin, X. et al. pROC: an open-source package for R and S\u0026thinsp;+\u0026thinsp;to analyze and compare ROC curves. \u003cem\u003eBMC Bioinform.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 77 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ebenvancalster. benvancalster/PerfMeasuresOverview. Published online February 5. Accessed February 27, 2026. (2026). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/benvancalster/PerfMeasuresOverview\u003c/span\u003e\u003cspan address=\"https://github.com/benvancalster/PerfMeasuresOverview\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFranchin, B. et al. Normothermic Kidney Perfusion: Current Status and Future Perspectives. \u003cem\u003eKidney Int. Rep.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (9), 2943\u0026ndash;2952. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ekir.2025.06.041\u003c/span\u003e\u003cspan address=\"10.1016/j.ekir.2025.06.041\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRabelink, T. J. et al. Opportunities and challenges with the implementation of normothermic machine perfusion in kidney transplantation. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e (1), 6883. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-025-60410-3\u003c/span\u003e\u003cspan address=\"10.1038/s41467-025-60410-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHalloran, P. F., Madill-Thomsen, K. S. \u0026amp; Reeve, J. The Molecular Phenotype of Kidney Transplants: Insights From the MMDx Project. \u003cem\u003eTransplantation\u003c/em\u003e \u003cb\u003e108\u003c/b\u003e (1), 45\u0026ndash;71. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/TP.0000000000004624\u003c/span\u003e\u003cspan address=\"10.1097/TP.0000000000004624\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu, W. et al. Rapid pathogen detection by metagenomic next-generation sequencing of infected body fluids. \u003cem\u003eNat. Med.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e (1), 115\u0026ndash;124. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-020-1105-z\u003c/span\u003e\u003cspan address=\"10.1038/s41591-020-1105-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDumbill, R. et al. Prolonged normothermic perfusion of the kidney prior to transplantation: a historically controlled, phase 1 cohort study. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e (1), 4584. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-025-59829-5\u003c/span\u003e\u003cspan address=\"10.1038/s41467-025-59829-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaju, A. F., Mukundan, A., Divyashree, M. S., Hariharapura, R. C. \u0026amp; Mahadev Rao, A. RNA diagnostics and therapeutics: a comprehensive review. \u003cem\u003eRNA Biol.\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e (1), 1\u0026ndash;11. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/15476286.2024.2449277\u003c/span\u003e\u003cspan address=\"10.1080/15476286.2024.2449277\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChu, V. T. et al. Implementation and outcomes of a rapid response genomic hospital epidemiology programme at an academic medical centre over 7 years. \u003cem\u003eLancet Microbe\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e (2), 101277. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.lanmic.2025.101277\u003c/span\u003e\u003cspan address=\"10.1016/j.lanmic.2025.101277\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShuja, A. Advances in RNA-Seq Data Analysis: Towards More Reliable Clinical Translation: Advances in RNA-Seq Data Analysis: Towards Reliable Clinical Translation. \u003cem\u003eDev. MEDICO-LIFE-SCIENCES\u003c/em\u003e ;\u003cb\u003e2\u003c/b\u003e(8):1\u0026ndash;2. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.69750/dmls.02.08.0143\u003c/span\u003e\u003cspan address=\"10.69750/dmls.02.08.0143\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S. et al. Evaluation of clinical prediction models (part 1): from development to external validation. \u003cem\u003eBMJ\u003c/em\u003e \u003cb\u003e384\u003c/b\u003e, e074819. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj-2023-074819\u003c/span\u003e\u003cspan address=\"10.1136/bmj-2023-074819\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarrell, F. E. xRegression Modeling Strategies: With Applications, to Linear Models, Logistic and Ordinal Regression, and Survival Analysis, 2nd ed. Heidelberg: Springer. \u003cem\u003eBiometrics\u003c/em\u003e. ;72(3):1006\u0026ndash;1007. (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/biom.12569\u003c/span\u003e\u003cspan address=\"10.1111/biom.12569\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJolliffe, I. T. \u0026amp; Cadima, J. Principal component analysis: a review and recent developments. \u003cem\u003ePhilos Trans A Math Phys Eng Sci\u003c/em\u003e. ;374(2065):20150202. (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1098/rsta.2015.0202\u003c/span\u003e\u003cspan address=\"10.1098/rsta.2015.0202\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNygaard, V., R\u0026oslash;dland, E. A. \u0026amp; Hovig, E. Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses. \u003cem\u003eBiostatistics\u003c/em\u003e \u003cb\u003e17\u003c/b\u003e (1), 29\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/biostatistics/kxv027\u003c/span\u003e\u003cspan address=\"10.1093/biostatistics/kxv027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeldner-Busztin, D. et al. Dealing with dimensionality: the application of machine learning to multi-omics data. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cb\u003e39\u003c/b\u003e (2), btad021. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bioinformatics/btad021\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btad021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar, S., Agarwal, A. \u0026amp; Chatterjee, S. Feature learning augmented with sampling and heuristics (FLASH) improves model performance and biomarker identification. \u003cem\u003eNPJ Syst. Biol. Appl.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 137. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41540-025-00614-x\u003c/span\u003e\u003cspan address=\"10.1038/s41540-025-00614-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, S., Agarwal, V., Gopinath, A. \u0026amp; Kassis, T. The Limitations of TabPFN for High-Dimensional RNA-seq Analysis. \u003cem\u003ebioRxiv\u003c/em\u003e. Preprint posted online September 17, 2025:2025.08.15.670537. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2025.08.15.670537\u003c/span\u003e\u003cspan address=\"10.1101/2025.08.15.670537\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S. et al. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. \u003cem\u003eBMJ\u003c/em\u003e \u003cb\u003e385\u003c/b\u003e, e078378. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj-2023-078378\u003c/span\u003e\u003cspan address=\"10.1136/bmj-2023-078378\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGallifant, J. et al. The TRIPOD-LLM reporting guideline for studies using large language models. \u003cem\u003eNat. Med.\u003c/em\u003e \u003cb\u003e31\u003c/b\u003e (1), 60\u0026ndash;69. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-024-03425-5\u003c/span\u003e\u003cspan address=\"10.1038/s41591-024-03425-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIrish, W. D., Ilsley, J. N., Schnitzler, M. A., Feng, S. \u0026amp; Brennan, D. C. A Risk Prediction Model for Delayed Graft Function in the Current Era of Deceased Donor Renal Transplantation. \u003cem\u003eAm. J. Transplant.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (10), 2279\u0026ndash;2286. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1600-6143.2010.03179.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1600-6143.2010.03179.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2010).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003eTable 1 \u0026ndash; study cohort details, and our aims for each cohort. Normothermic Machine Perfusion (NMP) is a way of keeping an isolated kidney alive outside of the body. Delayed graft function (DGF) was our primary measure of how well the kidney was functioning after transplant.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"652\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11.6564%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCohort\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40.4908%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDetails\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.3988%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSource\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 27.454%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAim\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11.6564%;\"\u003e\n \u003cp\u003eNMP Trial samples\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40.4908%;\"\u003e\n \u003cp\u003eN=34 kidneys biopsied during the United Kingdom NMP randomised trial.\u003csup\u003e5\u003c/sup\u003e All kidneys were subsequently transplanted, with data on early post-transplant function available for all. 9/34 recipients experienced DGF.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.3988%;\"\u003e\n \u003cp\u003eData provided upon request from the authors of Ferdinand 2021\u003csup\u003e6\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 27.454%;\"\u003e\n \u003cp\u003eAssess ability of our pipeline to predict DGF using biopsies taken during NMP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11.6564%;\"\u003e\n \u003cp\u003eCytokine Filter samples\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40.4908%;\"\u003e\n \u003cp\u003eN=10 kidneys (5 pairs). Preclinical normothermic machine perfusion with or without a cytokine filter treatment. The aim of the cytokine filter treatment is to improve post-transplant kidney function.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.3988%;\"\u003e\n \u003cp\u003eFerdinand 2021\u003csup\u003e6\u003c/sup\u003e and associated GEO\u0026nbsp;accession GSE121447\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 27.454%;\"\u003e\n \u003cp\u003eAssess whether DGF predictions are modifiable by therapies delivered during NMP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11.6564%;\"\u003e\n \u003cp\u003eQUOD retrieval biopsy samples\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40.4908%;\"\u003e\n \u003cp\u003eKidneys biopsied at the donor hospital (before transport of the kidney on ice). These biopsies were collected and stored in the Quality in Organ Donation biobank. N=263\u003c/p\u003e\n \u003cp\u003ekidneys with sufficient data on DGF. 73/263 recipients experienced DGF.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.3988%;\"\u003e\n \u003cp\u003eZhang 2024\u003csup\u003e18\u003c/sup\u003e and associated GEO\u0026nbsp;accession GSE215108\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 27.454%;\"\u003e\n \u003cp\u003eAssess ability to predict DGF with retrieval biopsies (in kidneys not receiving NMP).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"BulkFormer, TabPFN, foundation model, transcriptomics, Bulk RNA-Seq, kidney transplantation","lastPublishedDoi":"10.21203/rs.3.rs-9242336/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9242336/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eGenerating predictions from transcriptomic data poses a unique challenge due to the high number of genes, and often small sample size. BulkFormer and TabPFN have emerged as leading transformer-based foundation models for bulk transcriptomic and tabular data respectively. We explore an artificial intelligence pipeline using BulkFormer-TabPFN v2.5 which generates zero-shot predictions from raw RNA-Seq count data without retraining. This was tested on three cohorts of biopsies taken from donated human kidneys. BulkFormer-TabPFN was able to predict delayed kidney function using RNA-Seq counts from kidneys undergoing ex-situ normothermic machine perfusion (NMP; c-statistic=0.82, 95% CI=0.67–0.97). Predictive discrimination was optimised under the following conditions: BulkFormer-TabPFN versus TabPFN alone, maximum absolute aggregation of BulkFormer gene-level embeddings, biopsies taken during ex-situ NMP versus cold storage. BulkFormer-TabPFN predictions were modified by cytokine filter treatment during ex-situ NMP, suggesting they could be a dynamic surrogate endpoint for novel therapeutics, which is intrinsically linked to post-transplant outcome. This demonstrates for the first time synergistic benefits of these foundation models, to generate zero-shot predictions without model retraining. The provided code provides a blueprint to replicate generating predictions from RNA-Seq count data, which could be applied in a wide range of biomedical contexts within and beyond organ transplantation.\u003c/p\u003e","manuscriptTitle":"Combining BulkFormer and TabPFN to predict post- transplant function from kidney biopsies during machine perfusion or cold storage","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-06 13:42:12","doi":"10.21203/rs.3.rs-9242336/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"136901427629089144836681301451277757727","date":"2026-05-15T14:03:59+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-15T00:13:11+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-03T10:13:36+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-28T12:52:46+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-28T12:52:25+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-03-27T08:34:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d2b2867d-28af-4468-8365-3bab8141bbe3","owner":[],"postedDate":"April 6th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"136901427629089144836681301451277757727","date":"2026-05-15T14:03:59+00:00","index":76,"fulltext":""},{"type":"reviewersInvited","content":"4","date":"2026-05-15T00:13:11+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":65743498,"name":"Biological sciences/Biological techniques"},{"id":65743499,"name":"Health sciences/Biomarkers"},{"id":65743500,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":65743501,"name":"Health sciences/Nephrology"}],"tags":[],"updatedAt":"2026-05-15T00:23:27+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-06 13:42:12","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9242336","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9242336","identity":"rs-9242336","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}