EVA

EVA is a foundation model developed by Scienta Lab, designed specifically to predict human immune response in the context of immunology and inflammation research. It is built for drug development teams working across discovery, preclinical, and early clinical stages who need to bridge the gap between preclinical data and clinical outcomes in immuno-inflammatory diseases.

Scienta Lab developed EVA in response to a persistent challenge in immunology drug development: approximately 70% of drug candidates fail to translate into clinical benefit, largely because signals are lost between laboratory models and human biology. EVA provides a predictive layer that converts early biological data into clinical insight, supporting decisions from target identification through to trial design.

Core Capabilities

• Target and mechanism prioritization: Ranks targets and mechanisms of action by predicted probability of clinical success, using integrated multi-omic data, disease biology, and early experimental signals. This allows teams to identify which programs warrant further investment before major resources are committed.
• Preclinical-to-human translation: Predicts human efficacy from preclinical and target-level data by simulating drug impact on immune pathways and translating those effects into expected clinical outcomes. Supports objective candidate comparison and early go/no-go decisions ahead of IND filing.
• Patient stratification and trial optimization: Predicts patient-level response from baseline and post-treatment data to identify biomarkers and support trial enrichment strategies. Intended to improve signal detection and accelerate decision-making in Phase II–III development.

Therapeutic Areas Covered

• Immunology
• Rheumatology
• Endocrinology
• Dermatology
• Immuno-inflammation
• Gastroenterology
• Neurology
• Respiratory

Model Architecture and Training Data

• EVA is a cross-species, multimodal foundation model trained on more than 500,000 transcriptomics samples and 20 million histology tiles.
• It learns biology across multiple autoimmune and inflammatory diseases simultaneously, capturing shared immune mechanisms across conditions, tissues, and patient populations.
• EVA is trained using ImmunAtlas, Scienta Lab's proprietary dataset combining multi-omic, clinical, and histopathology data from more than 96,000 patient and biosample profiles, with ongoing integration of new datasets.
• The model covers more than 60 immuno-inflammatory diseases and integrates data from more than 70 tissues, drawing on over 200,000 patient and biosample profiles across diseases.
• Reported accuracy for clinical trial success predictions is 65%, with the model described as doubling clinical trial success rates compared to existing benchmarks.

Data Requirements and Compatibility

• At the discovery stage, EVA can generate predictions without asset-specific data by drawing on immune biology learned across diseases and populations.
• For assets in preclinical development, relevant experimental data is required; prediction confidence increases as more data is provided.
• Compatible data types include transcriptomics (bulk or single-cell RNA-seq), proteomics, clinical scores, histopathology, target information, and preclinical readouts.
• EVA can model first-in-class assets, novel targets, bispecifics, and combination therapies by simulating the joint perturbation of multiple targets or mechanisms on immune networks.

Differentiation from Generic AI Models

• EVA is purpose-built for immunology and inflammation, trained on immune-specific biological and clinical data rather than general-purpose datasets.
• It applies transfer learning across diseases, tissues, and patient populations, enabling biologically grounded predictions even from limited new data.
• Standard dataset analysis and generic AI models do not apply this cross-disease transfer learning approach.

Each project using EVA is supported by Scienta Lab's team of immunologists and data scientists, who help translate model predictions into concrete development decisions spanning target selection and trial strategy. Proprietary data submitted by users is used only to fine-tune a private instance of the model for that specific project and is not shared or incorporated into EVA's global training.