GEM-1 Predicts Clinical Trial Response
A virtual human modeling + immune checkpoint blockade case study
Biopharma is awash in the “spectacle” of foundation models, whose impressive benchmarks often fail to move the needle for real-world drug development teams. At Synthesize Bio, we go beyond model benchmarks and tackle actual challenges in drug development. Our Virtual Human framework captures the hidden biological signals that dictate clinical success and applies to real situations.
We’ll unpack how our flagship model, GEM-1, predicts drug response and clinical success.
Challenge: Predict Response to Immune Checkpoint Blockade
Immune Checkpoint Blockade (ICB) has significantly improved treatment options for major cancers like lung and skin cancer by targeting PD-1/PD-L1 signaling, a key axis that cancers leverage to survive immune attacks. The blood-vessel-blocking (anti-angiogenic) drug Bevacizumab hits VEGF to cut off vascularization to prevent tumor growth in a variety of solid tumors. In combination, they successfully treated a difficult tumor type, advanced hepatocellular carcinoma.
This discovery opened up a new arena… Now there are many PD-(L)1/VEGF bispecific programs in development for cancer treatment. Bispecifics combine the ICB approach (e.g. PD-1/PD-L1) with the anti-angiogenic approach (e.g. VEGF) in one molecule to target both these pathways. It’s very promising! They’re being tested widely in trials that, at best, use only a simple immunohistochemistry test to identify patients who may respond.
Leading Bispecifics in Clinical Development
How could this lack of precision medicine be addressed? PD-L1 immunohistochemistry helps to predict efficacy of ICB but doesn’t capture the complex, underlying state of the tumor microenvironment, where the drugs need to act. Fairly simple transcriptomic models do a bit better at predicting clinical response. We show that GEM-1, a more powerful transcriptomic approach, does better still.
Synthesize Bio’s GEM-1 model can better focus on the biology and potentially generalize across cancer types. But how? It moves away from simple gene expression and toward the latent biological factors that GEM-1 learned. Latents are like the “signal” hidden beneath the “noise” of different labs and machines, making a better predictor.
We used IMBrave150, the landmark study that established Atezolizumab + Bevacizumab as the new standard of care for liver cancer, to ask: If the study had used GEM-1’s predictions to inform trial design, would they have seen better response rates to the new treatment? Although this drug combination was a major improvement over standard of care, most patients still didn’t respond to therapy in this trial. So, we started with an analysis of IMBrave150, taking gene expression data of baseline biopsies and passing it through GEM-1 to extract biological latents to be used as variables that could predict response.
The results were striking. Regression with the GEM-1 latents surpassed the predictive performance of RNA-seq counts data and strongly outperformed the FDA-approved PD-L1 immunohistochemistry (IHC) approach or expression of the target gene (CD274, a.k.a. PD-L1) itself.
At a reasonable selectivity threshold, this translates into an estimated 15% gain in response rate from this trial. A 15% gain in response rate is often the difference between a failed trial and an approved drug.
And latent regression in IMBrave150 didn’t just predict response in liver cancer. It translated well to IMvigor210, an ICB trial in bladder cancer. Simple gene expression data alone (rather than the GEM-1 latents) likely fell short here because they aren’t portable measures of biological state across different organs or cancers. Biology in the liver isn’t the same as biology in the bladder, but the immune response is. GEM-1 effectively “cleaned” the data, separating biological signals from technical noise (batch effects, sequencing platforms) and learned to generalize about the immune response.
Predicting Current and Future Trials
Analyzing old trials, like IMBrave150, validates our models and tells us how well they perform. The important next step is predicting the outcomes of new trials to know where ICB+VEGF therapies can succeed and have impact.
First, we applied our latent-based predictor to lung cancer to forecast the success of ongoing trials of the bispecific therapies. Gene expression data are not available for these trials, but GEM-1 can generate data! We made 100 virtual clinical trials of 100 patients each for both lung adenocarcinoma and lung squamous cell carcinoma, then predicted response from aPD-L1/aVEGF treatment. For these indications, we already have measurements of objective response rate (ORR) for three drugs to evaluate model performance. Our predictions line up strikingly closely with observed trial results.
GEM-1 Predicts Response Rates to Bispecifics in Lung Cancer
Scaling Across the Entire Cancer Landscape
GEM-1 makes it possible to scale this approach across hundreds of cancer types, many of which might otherwise not get their own, dedicated Phase 3 trial.
We applied our latent-based predictor of PD-L1/VEGF blockade response to the 27 cancers with GEM-1-generated data in the SYNTH-cancer dataset plus GEM-1-generated expression profiles of 90 additional cancers that are seen at lower incidence in the clinic. So, we’re utilizing our GEM-1 model at two levels: first, via the model latents that predict treatment response; second, to create a broad, harmonized gene expression cancer atlas on which to perform the prediction. This way, we were able to map out success likelihood for ICB+VEGF across a wide swath of the oncology landscape.
Here’s a selection of cancers from this project with patient-level predictions of the likelihood that they’ll benefit from ICB+VEGF treatment. A few cancer types are labeled. Key features worth noting:
Patient response is rather uniform (small boxes, short bars) in some cancers, while response in other cancers is heterogeneous and would likely benefit from a companion diagnostic.
The range of response probabilities across cancer types is huge, suggesting that a precision medicine approach to indication expansion for this treatment combo is critical.
This use of GEM-1 as a reliable predictor represents a shift in how we can and should use such powerful AI models. They inform clinical trial design from a whole new perspective: who and what is going to respond? Specifically, this AI-informed approach enables key clinical development activities:
Prioritize Indications: Identify which cancers (e.g., PDAC or glioblastoma) might show unexpected sensitivity.
Stratify Patients: Refine inclusion/exclusion criteria in silico before the protocol is locked.
Simulate Bispecific Effects: Use “reference conditioning“ to virtually swap out a patient’s current treatment with a bispecific agent, simulating the clinical impact of the one-two punch in molecules like HB0025 or JSKN027.
With GEM-1, we are no longer limited by the physical samples in our freezer. It empowers us to extend cohorts, test hypotheses in a virtual clinic, and move into pivotal trials with AI-powered estimation of patient response.
What’s Next?
Predicting response to ICB+VEGF is just one example of applying our models in a problem-first approach. We’re also using sample generation to forecast liver toxicity, augment sample sets to magnify their power to detect biological signals, and test key drug effects to predict clinical trial success likelihood at scale.
Stay tuned for more blog posts on these applications…
If you are working on these types of questions in drug development, reach out.
This post was authored by Alex Abbas, Director of Computational Biology at Synthesize Bio.
Learn more
You can learn more about the GEM-1 model by reading the preprint and trying it out on our webapp.