Researchers behind the development of two free-to-use AI platforms – Path2Omics and Path2Space – are working to bring advanced molecular insights within reach of every pathology laboratory.
The Pathologist asked lead researcher Eytan Ruppin to explain the science behind the ideas, and why the precision medicine community should be excited about these advances.
Why is there a need for AI tools that can infer molecular information from pathology slides?
Sequencing has become an essential part of both cancer research and clinical care, but it remains expensive, labor-intensive, and time-consuming. In many cases, clinicians need to make treatment decisions quickly. The reality is that, even at major medical centers, obtaining sequencing results may take six to eight weeks.
The goal is not to replace sequencing, but to make molecular insights more accessible and available much faster. If validated AI models can extract some of this information directly from routinely acquired pathology slides, they could help clinicians and researchers obtain actionable insights at a fraction of the cost and turnaround time.
Why could AI-inferred spatial biology be a game changer for precision oncology?
Instead of waiting days for complex spatial assays, we could potentially see results the next day. And instead of spending $5,000 or $10,000 per sample, the cost becomes closer to buying someone a cup of coffee.
There are different levels at which we can infer gene expression from pathology images. Bulk tumor gene expression is already moving closer to clinical use, and I expect it will become increasingly important in the coming years as clinicians look for more information to guide treatment decisions. But bulk expression gives you one measurement for the whole tumor. You take the tumor, process it, perform RNA or DNA sequencing, and generate a single readout.
Spatial transcriptomics is different. Instead of one measurement, you may have 1,000 measurements across different spots in the tumor microenvironment. That gives you much more detailed information.
The question is whether that additional detail truly improves decision-making. There is always the risk of “too much information.” Spatial biology gives us an enormous amount of data, but the goal is to show that this more granular information can produce better predictors – and, ultimately, better treatment choices for patients.
How does Path2Omics fit into the bulk omics vision?
Path2Omics is designed to predict bulk, rather than spatial, molecular information from pathology images. The platform will be accessible via a free website that is currently being built at Cedars-Sinai. It will allow pathologists, cancer clinicians, researchers, and academic institutions from anywhere in the world to upload digitized pathology slides and automatically receive inferred bulk omics data.
Initially, the website will focus on inferred bulk transcriptomics and methylation across 30 different tumor types. That is the right place to begin because bulk molecular profiling is already used widely in research and is moving into clinical practice. It gives clinicians and investigators information that can help support treatment decisions, but in a more scalable and accessible way compared with molecular testing.
Once the spatial models have been further validated, the same basic concept could be extended from bulk molecular prediction to inferred spatial biology. The longer-term plan is to build a similar website for Path2Space.
Moving into the spatial biology realm, what unmet need is Path2Space designed to address?
For years, I have been interested in developing new approaches for precision oncology that can better match patients with the therapies most likely to benefit them. In principle, spatial biology could be a powerful enabler of that goal.
Spatial biology is transforming cancer research because it provides a detailed view of the tumor microenvironment and the complex interactions that influence disease progression and treatment response.
However, current spatial omics technologies are even more expensive, labor-intensive, and time-consuming than bulk expression sequencing. They require specialized workflows that can be difficult to scale across large patient populations.
Following on from bulk expression inference, we believed AI might be able to infer at least some spatial biological information from digital pathology images. This could provide even more accurate biomarkers for predicting treatment response and matching patients to the therapies that best fit their individual tumors.
How does Path2Space learn to infer spatial transcriptomics from pathology images?
Each slide is divided into many small image tiles – for example, around 1,000 tiles per slide. For each tile, we have two matched pieces of information: the H&E image itself and the measured spatial transcriptomics data, such as Visium data, from that same region.
If we have slides from 20 or 30 patients, that can generate a training set of 20,000 or 30,000 tile-level samples. Each tile becomes a mini-example linking image morphology to the measured spatial molecular landscape.
We then use a foundation model to extract features from the pathology images and train a supervised machine learning model to predict the spatial transcriptomic information. The first step is therefore to build, test, and validate those models.
The second step is the real conceptual challenge. Once we have a model, we need to apply it to patient cohorts where spatial transcriptomics has not been measured – cohorts where we may only have the pathology images, plus survival or treatment response data. Then we need to ask: can the inferred spatial information predict outcomes better than the current state of the art?
How does Path2Space compare with established multi-omics approaches for predicting treatment response?
We looked at a study published in Nature, in which researchers combined whole-exome sequencing, RNA sequencing, clinical data, and pathology images to build a multi-omics model for predicting treatment response in breast cancer. The technique demonstrated promising predictive performance but required extensive molecular profiling and significant resources for each patient.
In our study, we performed a head-to-head comparison against that approach. Our model relies only on standard H&E pathology slides. There is no additional sequencing, complex molecular assays, or expensive multi-omics testing. Instead, the model uses AI-derived spatial characterization of the tumor microenvironment to generate predictions.
What we found was striking. The pathology-based spatial predictor outperformed the previously published multi-omics approach while requiring only a routinely acquired slide. In practical terms, that means generating a prediction using material that is already available, at a fraction of the cost and with a turnaround time measured in days rather than weeks.
That is the promise of this approach: not simply reproducing information from expensive molecular assays, but potentially delivering clinically useful predictions faster, cheaper, and at larger scale.
How might the tools fit into precision oncology?
Predicting bulk molecular information is a more mature problem. Through Path2Omics, we have developed models that infer bulk transcriptomic and methylation profiles across approximately 30 tumor types. In this setting, each tumor is represented by a single molecular measurement, making the problem comparatively simpler and more scalable.
Spatial biology is a much more complex challenge. Each Path2Space model requires extensive development, large datasets, and close collaboration among computational scientists, pathologists, and clinical researchers. In effect, every cancer type becomes a major project in its own right.
So far, we have developed and validated Path2Space for breast cancer, and we recently completed a large project in head and neck cancer. However, each new tumor type can require one to two years of work. That is why our initial web platform will focus on Path2Omics, while we continue expanding the Path2Space approach to additional cancers over time.
What is your long-term vision for tools like Path2Space?
These types of models could have an immediate impact on clinical trials. Better biomarkers can help researchers identify the patients most likely to benefit from a therapy, improve trial design, and potentially even rescue trials that might otherwise fail because the right patient populations were not selected.
But for me, the larger vision goes well beyond clinical trials. The goal is to rigorously study these tools, validate them prospectively, and ultimately make them part of the everyday workflow of pathologists and oncologists.
Of course, there is still a great deal of work ahead. These technologies must undergo extensive validation before they can be used in routine patient care. It is an ambitious goal, and there is no guarantee of success. But that is ultimately what motivates us.
That is also why we are building publicly available platforms and making these tools accessible to the research community. Our aim is to enable broader evaluation, accelerate discovery, and further democratize precision oncology. But the end goal is, of course, to contribute to better outcomes for patients.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
