Summary
A bioRxiv preprint presents POME, a self-supervised model that represents incomplete mixed-type biomedical data as a graph. The authors report improved missing-data imputation and useful sample representations across three biomedical datasets.
A bioRxiv preprint presents POME, a self-supervised model designed to analyse biomedical datasets that combine different data types with substantial missing values. The researchers report that the method produced strong results for filling in missing data and generated sample representations useful for subgroup discovery, prediction and representation-space analysis.
The work, posted on September 15, 2026, describes POME as a method for partially observed mixed-type data. It was evaluated on three real-world biomedical datasets and is available as a Python package through GitHub and PyPI.
Contents
Turning incomplete records into a graph
Biomedical data frequently combines numerical measurements, categories, labels and other variable types in the same dataset. Individual samples may also have different portions of their records missing. This combination makes it difficult to apply analysis methods that expect a complete table with one uniform data type.
POME addresses this structure by representing the dataset as a bipartite graph. In this type of graph, the two sides contain different kinds of entities: one side can represent samples, such as patients, while the other represents variables or measured features. The observed relationships between samples and variables preserve which information is available without requiring every sample to have a complete record.
The model uses shared concept learning to produce low-dimensional embeddings for both samples and variables. An embedding is a numerical representation that places objects with useful similarities closer together in a learned representation space. Because POME learns from the data itself rather than relying only on externally supplied labels, the authors describe it as self-supervised.
This design is intended to let the model handle mixed data types and missing observations together, rather than treating imputation, representation learning and downstream analysis as entirely separate steps.
What the authors tested
The researchers evaluated POME on three real-world biomedical datasets using several downstream tasks and objectives. They report state-of-the-art imputation performance, meaning that POME performed strongly when estimating values missing from the original data.
The evaluation also examined the quality of the learned sample representations. According to the authors, these representations supported unsupervised discovery of patient subgroups that were both well separated in the learned space and clinically meaningful. Unsupervised discovery means that the grouping process does not require predefined outcome labels for every sample.
The same representations were used for supervised predictive modelling, where known labels or outcomes guide the prediction task. The paper also describes zero-shot representation-space mining, including a use case involving adjuvant therapy modality recommendation. In this setting, the learned space is explored for patterns relevant to a new analytical objective rather than being limited to one original prediction task.
Why the representations matter
A method that can represent incomplete records may reduce the need to discard samples with missing information or force every variable into the same format before analysis. Producing embeddings for both samples and variables could also provide a common basis for comparing patients, measurements and relationships within a biomedical dataset.
The reported therapy-recommendation example illustrates the intended breadth of the approach, but it is an analytical use case of representation learning rather than evidence of improved treatment outcomes. The study is a computational evaluation of a preprint method; the source describes no clinical intervention or prospective patient trial.
The authors’ abstract reports validation across three biomedical datasets but does not give dataset sizes, numerical effect estimates or the comparator methods used for the state-of-the-art claim. Those details are important for judging how broadly the results may generalise. The method’s performance on other biomedical populations, data types and patterns of missingness will require further evaluation.
POME is presented as reusable research software, allowing other researchers to apply the approach to partially observed mixed-type datasets and assess it in their own analyses.