Deciphex CipherX Engine Translates AI Embeddings Into Pathology

Deciphex Brings Model-Agnostic AI Infrastructure to Clinical Pathology

Deciphex has publicly launched CipherX, an artificial intelligence engine designed to bridge the gap between raw foundation model outputs and clinical pathology workflows. The platform has been running in production since January 2026, powering both the company's Diagnexia clinical diagnostic service and its Patholytix pharmaceutical research operations.

According to Deciphex, the system addresses a fundamental challenge in pathology AI: while foundation models convert whole-slide image tiles into high-dimensional numerical embeddings, they struggle to trace predictions back to visible morphological features that pathologists can verify on glass slides. Regulatory bodies increasingly require auditable systems that remain stable when underlying models are updated or replaced.

"There are strong pathology foundation models available today, and more coming from well-resourced teams," Dr. Donal O'Shea of Deciphex stated. "What the field has been missing is a stable layer above the models that speaks in pathology terms, holds up as models change underneath, and lets a pathologist see how a conclusion was assembled. CipherX is that layer, and it has been running our clinical and research work for months."

Two-Layer Architecture Separates Models From Clinical Vocabulary

CipherX operates through a dual-component design. The foundation model layer incorporates multiple neural network encoders, including Deciphex's proprietary DCX-3 fusion model, and selects encoders dynamically based on diagnostic tasks and data-use rights. On THUNDER and EVA public benchmarks, DCX-3 performs alongside prominent pathology foundation models including Virchow2, UNI2-H, H-Optimus-1, RudolfV2, and Midnight.

The distinguishing element is the semantic layer positioned above these encoders. This layer decomposes mathematical embeddings into discrete units called "glyphs"—recurring structural tissue elements confirmed and named by subspecialist pathologists. These glyphs assemble into recognizable histological signatures, enabling the system to construct higher-order spatial arrangements such as tumor-infiltrating lymphocytes or tertiary lymphoid structures without requiring new model training cycles. Because outputs link to named structures that reviewing pathologists can inspect and reject, swapping the underlying foundation model does not disrupt the validated clinical vocabulary established with laboratory customers.

Clinical Performance and Deployment Model

In routine clinical operations, Deciphex deploys CipherX through three Diagnexia Assist tools focused on pre-analytical image quality assessment, complexity-based case triage, and post-authorization quality review. Detection algorithms built on CipherX signatures have achieved 99.85% negative predictive value for adenocarcinoma and 98.76% for melanoma in routine casework, according to company performance data.

For pharmaceutical research, the engine supports digital tissue biomarker development and digital companion diagnostics through Patholytix Research Services. Biomarkers generated using the platform maintain the same underlying framework, ensuring analytical findings can be reviewed against identifiable tissue morphology.

Deciphex is not offering CipherX as standalone licensed software. Instead, the engine is delivered within Diagnexia clinical workflows, embedded in the Patholytix platform, or deployed through custom development programs for study sponsors. The company emphasized that all active production tools function as adjuncts to pathologist decision-making, not substitutes for professional medical judgment.

Why it matters

Pathology AI has concentrated on building larger foundation models, but regulatory approval and clinical adoption require systems that pathologists can audit and that remain stable as underlying models evolve. By creating a semantic layer that translates model outputs into validated tissue structures, Deciphex addresses the gap between computational performance and clinical usability—a requirement for both diagnostic laboratories seeking regulatory clearance and pharmaceutical companies developing companion diagnostics.

Details were first reported by CLP Magazine.