Blog on CPTPSP paper with Merck group

We are really excited to share our work “Beyond the single average tumor: Understanding IO Combinations using a clinical QSP model that incorporates heterogeneity in patient response” done in collaboration with Merck colleagues and published now in CPTPSP (Thanks to the discerning reviewers & editors!) This has been ongoing work started a few years ago and presented in early form at PAGE 2018. We continue to work on it (presentations at AACR 2020, Keystone 2020, ACoP 2020 (Pg 325)) and are planning more publications on it in the coming months).

There are 2 key concepts presented in this paper that animates a lot of the work done by the joint group and motivate us to think beyond a single average tumor when understanding a patient’s response to treatment.

Cancer progression in response to pembrolizumab is attributable not just to target lesion growth, but also due to non-target growth & appearance of new metastases.

When a patient is committed to a melanoma trial, they often have multiple lesions. The clinician picks a few (median number of lesions picked = 3, range can be 1-12) of these lesions to track carefully by measuring their size at every visit as target lesions. Other lesions present at baseline are considered non-target; the clinician examines them but does not measure them at every visit.

The waterfall plot that cancer researchers are now familiar with, tracks the change in the size ((size = sum of the longest diameters) of target lesions only. The patients that respond are on the right-hand side of that chart with decreasing tumor sizes and the patients on the left have growing target lesions and are considered to have ‘Progressive Disease’ (when change in tumor size >20% from baseline).

Now, clinicians report another score called the RECISTv1.1 and here’s where things can get complicated. Patients who have ‘Progressive Disease’ under RECISTv1.1, not only have growing target lesions (as seen in waterfall), but they may have observable growth in non-target lesions (those lesions that were present at baseline, but doctors hadn’t tracked) OR show the appearance of new metastatic lesions.

As we were trying to fit our models to the reported public data, it was intriguing that from the waterfall of pembrolizumab monotherapy ~15% were progressing (because their target lesion growth was > 20%) but ~40% were declared to be ‘Progressive Disease’ by RECISTv1.1 criteria in 12 months. That means a ~25% of patients had shrinking or stable target lesions but were still considered to have Progressive Disease on treatment! This already suggests that the target lesions alone are not the whole story.

There is another complication in that both RECISTv1.1 & the waterfall capture the ‘best overall response’ and it is possible that while someone seems to be a responder at a given time, the target lesions may actually grow back to progress to disease, if we follow-up for the usual 2-year period or so. In the next stage of our work, we focus on the progression-free survival curves which capture these time-dynamics vs. any measure that captures only a snapshot of the patient’s disease trajectory.

Target lesions themselves respond heterogeneously to pembrolizumab and the measurement of the average change in size obfuscates the reality of growing and shrinking target lesions in a patient. This heterogeneity in response should inform combination strategy.

Figure 6 in the paper shows this essential feature of the model (and data), where a given Virtual Patient can have lesions with varying responses. Consider a Virtual Patient (VP #35 in the pembro waterfall) whose average increase in lesion size (dSLD = change in sum of longest diameters) with pembro alone is ~60%. The goal of combination treatment is to reduce this change rate all the way to 0 and then some to shrink the lesions. Also, this patient has 3 target lesions However, when we look a little closer at their target lesions: 1 of them is already shrinking by -30%, another is growing at 10% and another more than doubles with growth rate >100% to give us a change in size of 60%.

On combination with ipilimumab for this VP, dSLD goes down to ~-4%. However, the outcome for the patient doesn’t change. Why? Tumor 1 was shrinking anyway and just shrunk more at the measured time on combination (-45% on combination vs. -30% on pembro) and Tumor 3 is unresponsive to both pembro and the combination (+60% on combination vs. >100% with pembro alone at the measurement time). The other lesion goes from growing at 10% with pembro alone to decreasing at -20% on combination. We go into more detail on why this may be the case in the paper, based on the tumor-immune interaction within a lesion (it may depend on many things including the fraction of CD8 TCells in the lesion that can be ‘unlocked’ by pembro therapy. Hot tumors have a lot of CD8 TCells to unlock, and cold tumors have few or none).

But the clinical upshot of this is: even if sum of target lesions diameters shows greater reduction with combination, individual lesions may still be growing, and patient outcome will remain unchanged for many, especially those with even a single cold, unresponsive tumor. We don’t get to estimating progression free survival in this paper, but based on these considerations, we recommend therapies that are orthogonal in strategy (immune, to go after hot) & non-immune, to go after cold lesions) in treating cancers, especially cancers where pembrolizumab is already pretty effective.

In our more recent work, we have calibrated models to both heterogeneity in best overall response reported in multiple lesions per patient and progression free survival curves (Fig 6 from ACoP 2020 poster, pg 325) and we hope to use these models to predict effectiveness of combination therapies with pembrolizumab.