A “simulated cell” is a computational model that mimics how real cells behave by integrating harmonized multi-omics data with a curated universal cellular signaling network. The cell signaling network is trained using large datasets from experiments on actual cells. By simulating various cellular conditions, it predicts how cells will respond to various perturbations (e.g. different drug combinations, KOs, synthetic lethality screens). The model can simulate different cancer cell lines, PDXs, or even patient samples by inputting specific omics data, allowing it to predict the effects of drug therapies in a personalized, cell-specific manner. 

A scientist working on oncology drug development can use the Simulated Cell to predict, for example, how cancer cells will respond to different drug combinations. By inputting molecular data from specific cancer cell lines, PDXs, or even patient samples, the model can identify synergistic drug combinations and biomarkers relevant to those treatments. Hypotheses selected with Turbine have a 2X higher validation rate, enabling your scientists to focus on the experiments that matter, reducing the need for costly and time-consuming lab experiments. 

In addition to faster drug discovery and improved accuracy in predicting efficacy, simulations reveal the causal mechanisms driving disease or resistance, allowing for rapid hypothesis generation and testing at a scale and speed that traditional experimental methods cannot achieve. 

Turbine’s Simulated Cell technology is a game-changer for oncology drug development. Unlike traditional AI or data-driven models that rely purely on existing experimental data, Turbine has constructed a dynamic model that reflects the biological logic of intracellular signaling. This allows us to simulate how a drug will behave in different cancer cell types before running costly and time-consuming lab experiments. 

What sets us apart is our ability to model complex mechanisms and identify causal interactions that others miss. By integrating AI-driven hypothesis generation, detailed mechanistic simulations, and wet-lab validation, we can uncover hidden biology, reducing trial-and-error in the lab and leading to fewer failed experiments and faster progress to the clinic.

Turbine’s Simulated Cell can integrate multiple types of biological data, including: 

• Genomic data (mutations, CNVs from WES or WGS) 

• Pre- and post-treatment transcriptomic data (RNA-seq expression levels) 

• Proteomics data (protein abundance, post-translational modifications) 

• Functional screening data (CRISPR, RNAi) 

• Viability data including multi-dose IC50 values tested across cell lines, PDOs, PDXs.  

• Contextual information including payload/drug mechanism of action and any known resistance or sensitivity data, compound structure information.  

• Imaging data. 

Yes, patient-derived omics data, including transcriptomic, genomic, and therapy response information, can be used to create patient-specific in silico avatars. These avatars enable predictions about personalized drug responses and biomarker-driven treatment strategies. 

Turbine can simulate nearly any human cell type characterized by pre-and post-perturbation transcriptomic data including: 

• 1400+ cancer cell lines from various solid tumors and hematological malignancies 

• 600+ PDX models (NIBR PDX library) 

• 300+ patient samples (BEAT AML 2.0) 

• Organoid cultures 

• Immune cells 

Yes, Turbine’s Simulated Cell models can predict ADC payload effectiveness by: 

• Modeling cellular sensitivity to different payload classes using transcriptomic data. 

• Simulating response to novel payloads in untested cancer subtypes. 

• Predicting genetic factors driving sensitivity and resistance to ADC payloads. 

• Identifying potential patient stratification biomarkers for ADC development. 

• Identifying resistance-driving mutations and compensatory mechanisms. 

Recent studies validated that Turbine’s models accurately rank cancer subtypes by sensitivity to ADC payloads and predict the impact of gene knockouts on payload response. The platform extends these capabilities by modeling payload combinations and highlighting potential mechanisms of resistance. 

Yes, Turbine’s simulations allow overexpression (as well as under expression, knock-in, knock-out or interaction-specific function changes) analysis by modifying the respective simulation parameters and assessing downstream signaling impact. 

Yes, Turbine’s outputs enable the analysis of pathway reactivation, compensatory signaling, and genetic drivers of resistance. This enables the identification of resistance mechanisms and the prediction of new therapeutic strategies. 

Yes, the platform models combinatorial drug perturbations across any models in the database, identifying synergistic and antagonistic interactions. The approach is benchmarked against known synergy datasets and validated through experimental studies. 

Turbine has developed the Benchmark Suite. This carefully composed benchmark set focuses on the models’ ability to identify biologically applicable predictions.  

• Experimental verification through a wet lab to validate key phenotypic or mechanistic predictions. 

• Comparison with public CRISPR and drug response datasets, including benchmarking to phenotypic response as well as post-treatment transcriptomic states. 

• Predictive accuracy metrics (AUC, BAC, correlation coefficients) 

• The model’s predictivity power beyond simple biases in the data (e.g. cell lines from liquid tumors are more sensitive than solid ones) 

• Mechanistic readouts providing causal explanations for predictions 

• Ensemble modeling to assess prediction variability and confidence levels 

Turbine’s models can integrate patient-derived omics data, and clinical trial outcomes to enhance clinical translatability. We build on the understanding that human cells share the same types of proteins, irrespective of whether they are in a patient or a petri dish. Training on both kinds of data enables the model to transfer the logic of protein behavior where applicable, and make the necessary distinctions where the patient omics paints a differing picture 

Turbine uses: 

• An expert-curated signaling network derived from literature and databases 

• Machine learning-driven inference to refine model parameters 

• A graph attention network approach to enhance interpretability and feature importance ranking 

• Iterative validation with experimental data 

Turbine’s approach combines literature-derived knowledge and data-driven machine learning insights. The core signaling network includes manually curated and literature-supported interactions, ensuring biologically plausible outcomes. 

Yes, Turbine can: 

• Identify compensatory pathways driving resistance 

• Predict synergistic drug combinations to overcome resistance 

• Model genetic perturbations’ effects on drug response 

• Rank resistance mechanisms using in silico perturbation screens 

Yes, Turbine generates resistant cell avatars by integrating: 

• Modifying cells in silico. Successful mutations can then be realized in the lab if needed. 

• Mutational signatures linked to therapy resistance 

• Predictions have been shown to achieve high correlation with in vitro screens 

• Limitations arise from gaps in training data, longitudinal modeling, and cell-cell interactions, which are mitigated by targeted data generation. 

• Mechanistic insight generation, linking predictions to pathway alterations 

• Identification of key driver nodes in resistance and drug response 

• Benchmarking against known genetic dependencies and drug perturbations 

Yes, predictions are traceable to: 

• Key pathways and molecular interactions 

• Experimental evidence from public and proprietary datasets 

• Feature importance rankings from graph-based model inference 

• Diverse data sources to reduce over-representation of well-studied targets 

• Rigorous benchmarking to detect biases in dependency predictions 

• Customized model training for specific partner needs