On March 24, 2026, the U.S. Patent and Trademark Office issued US12586391B2 to Tempus AI, Inc., a granted patent titled “Systems and methods for deconvolving cell types in histology slide images, using super-resolution spatial transcriptomics data.” In plain terms, the claim covers a computer method that trains a machine-learning model to read a stained tissue image and infer, cell by cell, what kinds of cells are present — the kind of automated read that a pathologist would otherwise do by eye. For a company whose business has historically rested on sending samples through a sequencing lab, an issued claim on the software layer that interprets the images is a notable addition to what it can enforce.

The grant does not stand alone. The patent record shows Tempus holding 25 issued U.S. grants in total, with the pace of issuance rising each year: 6 in 2023, 10 in 2024, and 9 already attributed to the 2026 indexing window. The same March 24 issue date also carried US12584176B2, an “integrated machine-learning framework to estimate homologous recombination deficiency” — a second AI-plus-genomics grant landing on the same day.

What the estate actually covers

Read across its numbers, the portfolio clusters around a small set of capabilities. On the imaging side, US12524826B2 covers “determining biomarkers from histopathology slide images,” and US12361542B2 covers a deep-learning framework that fuses radiomic, pathology, and molecular image data for biomarker discovery. These three image-analysis grants — the March 24 deconvolution patent, the biomarker-from-slides patent, and the multimodal-fusion patent — describe overlapping pieces of one pipeline: take a digitized slide or scan, and output a structured molecular or diagnostic readout.

The estate also reaches the wet-lab and data-handling ends of the same workflow. US12365943B2 claims a method for “next generation sequencing uniform probe design,” an optimization that tunes the recovery rate of sequencing probes. At the records layer, US12512187B2 covers “sparse N-gram modeling for patient-entity relation extraction” — pulling structured relationships out of free-text electronic health records — and US12536595B2 covers mobile supplementation, extraction, and analysis of health records tied to clinical-trial matching. The footprint, in other words, runs the length of the diagnostic chain: sample preparation, image interpretation, and the clinical-data plumbing that surrounds both.

The classification codes reinforce that reading. Across the 25 grants, the most common CPC tag is G16H 50/20 — healthcare informatics for medical diagnosis — which appears on 9 of them, followed by G16H 10/60 (patient-record handling) on 7 and the imaging codes G06T 7/0012 and G06V 10/82 on multiple records. These are software and informatics classes, not the wet-chemistry classes that dominate a traditional diagnostics filer. The estate is built around computation applied to medical data.

A handful of records sit at the clinical-application end of that computation. US12471827B2 covers an ECG-based system for predicting future atrial fibrillation, taking electrocardiogram data into a trained model that outputs a risk score; US12451250B2 covers predicting therapeutic sensitivity using patient-derived tumor organoid cultures; and US12525343B2 covers an “artificially intelligent routing agent” that routes portions of a clinical task through multiple customized agents. The estate, read end to end, is not confined to one modality — it spans imaging, sequencing, records, and the predictive models that consume all three. That breadth is itself a fact about how the company is building its position: across the inputs a diagnostics platform ingests rather than around a single test.

A computer-implemented method, computing system and computer-readable medium include receiving training data and training a machine learning model to generate a cell expression map.— Systems and methods for deconvolving cell types in histology slide images, using super-resolution spatial transcriptomics data, US12586391B2

The business edge a granted claim buys

A granted patent, unlike a pending application, is enforceable coverage — it defines a boundary the holder can assert against others operating inside it. The March 24 deconvolution grant, paired with the biomarker-from-slides and multimodal-fusion grants, gives Tempus issued claims across the specific step of turning a stained-tissue or scanned image into a molecular readout. That is the same step a widening field of digital-pathology and AI-diagnostics entrants is building toward, which is why the freedom-to-operate question around image-to-readout pipelines now touches granted claims rather than only published applications.

The competitive context is visible in the same week's record. Across the diagnostics-and-assay grants issued the week of March 24, the volume leaders by assignee included a spread of research institutions and tools makers — Washington University with two, and single grants apiece to names such as Memorial Sloan Kettering Cancer Center, Massachusetts Institute of Technology, and other diagnostics filers — with Tempus appearing among them. No single assignee dominated the week's diagnostics grants; the field is fragmented across academic and commercial filers. Tempus issuing two AI-and-genomics grants on one day, against that fragmented backdrop, is a fact about the pace at which it is converting filings into issued coverage relative to a field where most assignees logged a single grant.

The data also frame what the estate is not. With G16H informatics codes dominating and only a handful of records touching the sequencing chemistry itself (the probe-design grant being the clearest example), the issued coverage is concentrated on the analytical and software layers. For a reader tracking the diagnostics business, the signal in the record is directional: Tempus is accumulating enforceable claims on the computation that sits on top of lab data, and the March 24 grant adds the histology-deconvolution piece to a cluster that already covered biomarker calling and multimodal image fusion. The volume trend — grants per year rising rather than flat — indicates the company is investing in building out that software-defined diagnostics position rather than treating it as a one-off.

What the reader does with that is their own call. The record simply shows a diagnostics company that issued two AI-and-genomics grants on a single day in late March, inside a 25-patent estate whose center of gravity has shifted decisively toward informatics and image analysis.