Cedars-Sinai Develops AI Tool to Predict Spatial Gene Expression
Cedars-Sinai investigators announced a new AI-based tool called Path2Space that predicts spatial gene expression across tumor biopsy slides, according to a May 8 Cedars-Sinai news release and a Newswise report. The team trained Path2Space on breast cancer cases where both slide images and spatial sequencing were available, then validated performance on three additional patient datasets, the coverage says. For each sample the researchers report predicting the spatial expression of almost 5,000 genes, with predictions matching measured expression across cohorts, per Newswise. Cedars-Sinai and Newswise state the method produces predictions in minutes and costs substantially less than conventional spatial profiling, which typically takes weeks and costs thousands of dollars. Cedars-Sinai quotes Eytan Ruppin highlighting the need for clinical validation before patient-care use.
What happened
Cedars-Sinai investigators published a description of an AI-based method named Path2Space (Path2Space) that uses digital images of tumor biopsy slides to predict spatial gene expression across the tissue, according to a Cedars-Sinai news release dated May 8, 2026 and a Newswise report of the same date. The team trained the method on a large breast cancer dataset where both histology slides and spatial sequencing data were available, then tested the tool on three additional patient cohorts, the reporting states. Newswise reports that the researchers predicted the spatial expression of almost 5,000 genes per sample and that the predictions matched measured expression across the three validation groups.
Technical details
Editorial analysis: The publicly reported workflow is a supervised image-to-expression mapping, trained on paired histology and spatial-transcriptomics data, per Cedars-Sinai and Newswise. The research team framed Path2Space as producing per-pixel or per-region estimates of gene expression (spatial transcriptomics) from standard stained slides, which reduces reliance on physical spatial-sequencing assays that are slower and more expensive, according to the press coverage. The releases do not publish model architecture, training hyperparameters, or evaluation metrics beyond the broad claim of cross-cohort agreement; reviewers and practitioners will need access to the paper, code, or benchmark data to assess reproducibility and bias.
Context and significance
Editorial analysis: Computational pathology efforts that infer molecular signals from routine histology images have been accelerating because they can unlock molecular readouts at scale without additional wet-lab assays. If externally validated, image-based spatial-expression prediction could materially lower per-sample cost and increase throughput for spatial biomarker discovery, according to the pattern observed in recent literature on histology-to-genomics prediction. However, image-inferred molecular readouts are subject to dataset shift, staining variability, and confounding signals; these are recurring technical challenges in similar projects.
What to watch
Editorial analysis: Observers should look for the peer-reviewed manuscript, public release of code and models, quantitative metrics (per-gene correlation, calibration, and spatial concordance), and external independent validations. Reports quote Eytan Ruppin saying clinical trials would be needed for patient-care translation; the Cedars-Sinai release and Newswise coverage indicate the group sees clinical validation as the next step. Practitioners should also watch for comparisons versus established spatial-transcriptomics platforms on cost, resolution, and sensitivity.
Scoring Rationale
This is a notable research advance in computational pathology that could lower cost and increase throughput for spatial biomarker discovery. The score reflects potential practitioner relevance combined with the need for peer review, code release, and independent validation.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
