Genialis™ Supermodel

A foundation model of cancer biology

Trained on ~ half a million RNA sequencing samples from over 30 tissue types, this large molecular model (LMM) defines a comprehensive landscape of cancer biology, learned from preclinical experiments and globally diverse patient records. The LMM maps each new patient sample based on its biological state, revealing underlying drivers and vulnerabilities.

The Genialis Supermodel delivers therapeutic intelligence by interpreting the molecular biology of tumor samples. The output are “biomodule” scores, algorithmic representations of over 150 biological phenomena, states, and processes. These include pathway activity, oncogene and transcription factor regulation, gene dependency, tumor microenvironment, cell state, DNA damage response, replication stress, and more.

Technically, the Supermodel is a compendium of phenotypic embeddings that serve as input features for machine-learning modelling of predictors. It is not a black-box, but an intrinsically interpretable biological space.  Practically, the Supermodel enables your translational and clinical teams to interrogate questions about response, mechanism, durability, indication, prognosis, … virtually any prediction task.

Biomodule scores can be used to train predictors that:

  • Distinguish responders from non-responders

  • Generate hypotheses about mechanisms of response and resistance

  • Suggest likely combination approaches, and

  • Uncover novel targets for cancer therapy

Genialis Supermodel uses RNA data to uncover cancer biology, quickly configuring biomarkers for drug targets.

Accessing the Genialis Supermodel

The Supermodel can be accessed on our cloud or deployed on yours, plugged into your existing data and AI ecosystem. Supermodel inference can be done in milliseconds per sample on a standard desktop computer. 

We’ve built an industrial-grade software stack that complements existing solutions, while rich APIs allow the Supermodel to autonomously contribute to established research workflows. 

In addition, Genialis provides AI-enabled expert services to assist with technology integration, data processing, and AI design, modeling, and interpretation.

Genialis Supermodel workflow

A typical workflow has four main phases:

  • Primary analysis analyzes raw RNA sequencing data and produces gene expression profiles

  • Data harmonization includes data quality assurance, data pre-processing, and batch effect detection and mitigation, as well as integration with your proprietary and third-party licensed data

  • Biomodule scores are transcriptomic embeddings, computed with the Genialis Supermodel

  • AI predictions are made by applying existing predictors (AI models) or training new ones on biomodule scores

Genialis Supermodel workflow from Primary analysis, Data harmonization to Biomodule scores and AI predictions

Four key software components support this workflow:

  • Genialis Expressions is a robust and scalable cloud software that captures analysis metadata and handles primary processing of sequencing data via a suite of validated pipelines. Extensive documentation available here.

  • Genialis Precision Medicine SDK prepares data for machine learning using data normalization, a system of preprocessors, and a framework for batch effect detection and removal

  • Genialis Supermodel transforms harmonized sequencing data into a low-dimensional biological space using biomodules

  • Predictors are AI models developed to predict response, prognosis, or other clinically relevant insights; these are trained by the Supermodel licensees or contracted and licensed directly from Genialis

Each of the software components comes with an API allowing bundled or independent deployment and integration of these components within existing architectures.

Supermodel white paper

Technical details and validation

Why RNA

Biomarkers that capture complex cancer biology from gene expression data

Genialis™ krasID

The first and only biomarker that can accurately stratify KRAS patients by clinical response