our model
Immunova: making precision oncology available to everyone, everywhere.
A diagnostic and detection-supportive suite built on holistic, transparent, and sustainable design principles across the cancer care continuum to enable wider access to sustainable & ethical precision oncology globally.
Immunova AI is a non-profit to expand access to precision oncology by developing an open-source diagnostic and detection-supportive suite.
Non-profit open-sourced diagnostic and detection-supportive suite integrating WSIs, multi-omics, and EHR data with ViTs, GNNs, and Transformer encoders to generate clinically actionable insights across the cancer care continuum.
Our input modalities.
01 Histology Preprocessing
02 Omics normalization and gene signature encoding,
03 Contextual embedding of clinical variables
Current problems
01 Low Predictive Accuracy – Single-biomarker tests (PD-L1) are insufficient, leading to high treatment failure rates.
02 Manual and Subjective Inputs – Critical inputs like Tumor-Infiltrating Lymphocyte quantification are often done manually by pathologists, a process that is slow, subjective, and prone to inter-observer variability
03 Lack of Model Transparency – “Black box” AI is a major barrier to clinical trust and adoption
04 Single Modality Limitations – Models fail to integrate synergistic data from imaging, genomics, and clinical records.
Our solution
01 Synergystic Data Fusion – Includes a drug optimization module with 62% alignment to real-world responder data and a survival prediction module with r = 0.82 correlation to 24-month PFS
02 High Predictive Accuracy – Achieved AUC up to 0.86 in predicting immunotherapy response in a retrospective Yale cohort, identifying ~40% of non-responders.
03 Lack of Model Transparency – “Black box” AI is a major barrier to clinical trust and adoption
04 Single Modality Limitations – Models fail to integrate synergistic data from imaging, genomics, and clinical records.
Our hybrid model fuses
Vision Transformers (ViT)
for spacial tissue analysis
Graph Neural Networks (GNN)
to encode cell-cell interactions
Transformer-based encoders
for omics and clinical features.
5 modules to provide accurate, impactful information
01 TIL Classification
This module automates detection and classification of tumour-infiltrating lymphocytes (TILs) from whole slide images (WSIs). The model operates in a unsupervised setting, requiring no manual annotations.
WSIs are tiled into fixed-size patches (256×256 pixels), followed by stain normalization to standardize tissue appearance.
A Res-Net40 deep learning model is then used to predict the presence and subtype of TILs per tile (CD3, CD34, CD38, CD20, CD68, CDK4, D2-40, Cyclin-D1, Ki67, FAP, P53, SMA) and other immune phenotypes.
Outputs include a CSV file of per-tile predictions and Grad-CAM heatmaps, which provide spatial interpretability and highlight histological regions influencing the prediction.
02 Treatment Prediction
This module integrates Multiomics data (RNA-seq, Methylation, Proteomics, miRNA, CNV) and image-derived immune features to estimate the likelihood of a patient responding to immunotherapy.
TIL features are aggregated across tiles to obtain sample-level immune profiles.
The model then predicts a probabilistic treatment response score for each patient, alongside key marker genes and feature importances driving the prediction
This ensure transparency, making the module particularly suited for guiding treatment plans in precision oncology workflows and reducing ineffective immunotherapy use.
03 Survival Analysis
This module combines transcriptomic and clinical data to predict patient survival risk and stratify outcomes.
It accepts a normalized RNA-seq expression matrix alongside key clinical variables such as age, cancer stage, treatment history, and survival time/status.
Survival time is standardized and paired with status indicators to train a risk estimation model.
The output is a hazard risk score for each patient, indicating the likelihood of poor prognosis visualized via a Kaplan-Meier survival curve.
04 Reversion Module
This module maps transcriptional state changes in tumor cells to characterize progression and plasticity within the tumor microenvironment. It integrates single-cell RNA sequencing data with copy number variation (CNV) profiles generated using CopyKAT to distinguish malignant from normal cell populations.
Highly variable genes are selected from the normalized scRNA-seq matrix, and attractor state analysis is applied to model the dynamic trajectories of transcriptional changes. The module identifies key gene switches and transition boundaries to calculate a “reversion score” for each cell, quantifying its deviation from a normal-like state.
Outputs include per-cell annotations of CNV status, reversion scores, and a list of top genes driving malignant transformation. This module offers insight into intratumoral heterogeneity and may reveal novel therapeutic targets for reversing malignant phenotypes.