Dmitry Gabrilovich (AstraZeneca)

Responses to interview questions from Tal Behar, Precision Medicine World Conference

1. Trial Design

Question:
In myeloid-targeting combination trials in solid tumors, what’s the most common trial-design mistake you see (patient selection, timing, endpoints, assays), and what one change would you make first to avoid false “no-signal” outcomes?

Answer:
The challenges with myeloid cell targeting go beyond trial design alone.

First, cellular and pathway redundancy must be considered. Cellular redundancy refers to the ability of cells from different lineages, primarily granulocytic and monocytic/macrophage populations, to compensate for each other’s function. As a result, simultaneous targeting of multiple myeloid populations is likely necessary.

Pathway redundancy reflects compensation by distinct molecular mechanisms within the same cell. Therefore, identifying and targeting shared pathological pathways is preferable to targeting individual molecules.

Second, approaches must focus on the pathological state of myeloid cells, not on indiscriminate depletion of all myeloid cells. The pathological states of myeloid cells are now well defined and extensively described in the literature.

Third, patient selection should be based on the presence of myeloid expansion and tumor infiltration on a patient-specific basis. A substantial proportion of patients, up to one-quarter in many cancer types, do not exhibit myeloid expansion. In these patients, myeloid-targeted therapies are unlikely to be effective.

Fourth, treatment scheduling should be reconsidered. Current approaches typically rely on continuous and prolonged treatment, which may facilitate compensatory responses due to the high plasticity of myeloid cells. Intermittent therapy at higher doses may prove more beneficial.

2. Translational Biomarkers

Question:
If you could standardize one translational biomarker or assay across myeloid-checkpoint trials, what would it be, and what would you want it to prove early in patients (mechanism and likelihood of response)?

Answer:
The most effective and practical approach for patient selection is assessing the presence of pathological myeloid cells, including macrophages and myeloid-derived suppressor cells (MDSCs), in pre-treatment tumor biopsies.

These biopsies are routinely available for most patients and do not require additional invasive procedures. Relevant markers are well reported, and numerous studies demonstrate a strong association between myeloid infiltration and clinical outcomes.

What remains missing is a direct association between these biomarkers and response to myeloid-targeted therapies.

Blood-based assays are widely used, and markers of MDSCs are well established. However, assay variability can be significant, requiring rigorous control over sample acquisition and analysis. While this is achievable in academic settings, it presents a major challenge for harmonization in large-scale multicenter trials.

On-treatment biopsies and blood samples provide excellent material to assess target engagement and should be incorporated early in clinical development. Demonstrating even a modest association with clinical outcomes in a small initial cohort can provide sufficient confidence to advance a therapy into larger trials.

Without such associations, it becomes difficult to understand why a combination therapy fails or produces inconsistent clinical signals.

3. Message to the PMWC Community

Question:
Are there any other points you would like to share with the PMWC community?

Answer:
Myeloid cells are a critical component of the tumor microenvironment and represent one of the major factors limiting the success of current cancer therapies.

Targeting these cells is one of the most exciting opportunities to enhance clinical benefit, including in patient populations that otherwise respond poorly to treatment. However, success will depend on applying smart, biologically informed approaches such as those outlined above.

Suzanne Topalian’s (Johns Hopkins) responses to interview questions from Tal Behar, Precision World Medicine Conference

James Zou’s (Stanford) responses to interview questions from Tal Behar, Precision World Medicine ConferenceQ1.
When your virtual AI lab started generating and debating its own hypotheses, what was one idea or research direction it proposed that genuinely surprised you, and what did that moment change in how you think about discovery?
A1. The virtual lab AI agents made several decisions, such as choosing to design nanobodies rather than the more common antibodies, that surprised me initially but turned out to work well. Most of the decisions and work done by the virtual lab are quite human-like, actually, but much faster. For example, one virtual lab research meeting might take less than a minute, and the agents typically run five meetings in parallel to discuss each topic so they can explore a broader set of ideas. That efficiency will change what discoveries are possible.Q2.
Where does this agentic AI approach still break today, or make you pause before acting on its output, and where does it already outperform how human teams typically reason or explore the problem space?
A2. The current AI agents are better at leveraging and combining existing tools than at creating entirely new tools from scratch. For example, agents can adapt AlphaFold and combine it with other models to create new computational pipelines for drug discovery—I call this combinatorial creativity. It’s still open how to make truly creative AI agents that can come up with entirely novel yet feasible ideas.Q3.
Looking ahead a few years, what does meaningful “human judgment” look like in an AI-native lab, and what is the first concrete proof point that would convince skeptics that this model can close the loop from discovery to development?A3. Imagine each human researcher being a mini-PI, supported by a customized virtual lab of AI agents. The human manager would delegate tasks to the agents, review critical steps and give feedback. The agents would be running tirelessly, analyzing data, synthesizing literature knowledge, drafting documentation, etc. The technology is getting close to the point where this is feasible. This would require changing human habits and workflows, which would take longer than getting the technology ready.