Company goal is not to demonstrate one drug or another as more effective, but to define models which use past data and intended dosing to predict future values. This capability is essential for deciding which intended drug choice and dose is best for the patient.
Advances in computability allow for the progression from asking what is safe and effective for a group to detailed prediction of drug impact on the individual patient. This drives a revolution in personalized medicine and the drug approval process. The company is well positioned to be a leader as part of this seismic change.
Company has several years of Dana-Farber MyChart and other health institution data which includes the usual blood tests, PET scans, etc. Dataset is further enhanced by detailed dosing history coming from patient daily logs. There are tens of thousands of images tracking tumor wounds and other features. Camera image data can show drug driven changes on the granularity of a few days — faster than possible from more invasive image scanning results. Faster and more tumor site specific than possible from blood tests.
It is expected that inclusion of 10x Genomics data will enhance several aspects of site dependent disease progression. Because CDK4/6 inhibitors work by putting cancer cells into a suspended state, 10x samples can function like tree cores allowing analysis of a cancer genomics evolutionary history.
Existing tools provide easy and complete access (with permission) to patient MyChart data. Blood lab test results are available in a per institution csv file, facilitating analysis and visualization with Python libraries: NumPy, pandas, matplotlib, Seaborn, etc.
Perturbation Dosing Strategies for two CDK4/6 inhibitors
The CDK4/6 inhibitors ribociclib and abemaciclib drugs are commonly used for the treatment of breast cancer. They provide an example illustrating company methodology.
FDA guidelines for ribociclib1 and abemaciclib2 drugs can be viewed as the choice of a 28 day pattern where each day in the pattern is assigned a number D (for dose) from the set {0, 1, 2, 3}. For ribociclib, dose in milligrams is D * 200 mg. For abemaciclib, dose in milligrams is D * 100 mg.
Guidelines give four pattern choices for each drug. For ribociclib, patterns are of the form: D is the same value for 21 days followed by zero for 7 days. For abemaciclib, patterns are of the form: D is the same value for 28 days.
When both drugs are considered together, there are 7 patterns: 3 non-zero ribociclib patterns, 3 non-zero abemaciclib patterns and the zero pattern.
Assume rules for patterns are relaxed. Daily limits and monthly limit for the sum are retained.
If the D value can be assigned to each day independently, there are 4 raised to the 28th power patterns. This is somewhat more than 10 raised to the 16th power. This is correct for abemaciclib. All ribociclib patterns must end with 7 zeros. There are “only” 10 raised to the 12th power of these.
Intuitively these patterns are safe, as they employ a single drug type and are lower in dose than one of the standard patterns.
Ribociclib dose was dynamically adjusted to keep ALT values low while preventing disease progression. This effort was successful.
There is no claim that this would be generally true for other patients.
The previous perturbation approach can be generalized (or merged). If D = (R, A) where R is the number of ribociclib and A is the number of abemaciclib, then toxicity limits from the 28 day guidelines give:
- R + A <= 3
- (SUM R)/63 + (SUM A)/84 <= 1
- R is zero for last 7 days
The space of allowed (R, A) values is much greater than 10 raised to the 16th power. The exact number is unimportant.
The previous defines a sum rule. Given two patterns: (1) a valid pure R 28 day pattern and (2) a valid pure A 28 day pattern, the sum of the two patterns is a valid merged pattern when all three of the listed merge restrictions hold. More generally, the sum of any two patterns is valid when the merge restrictions hold.
The diference of two patterns can be used to define a perturbation vector. A 2D vector (R, D) defines each day of the perturbation where -3 <= R <= +3 and -3 <= A <= +3. A negative perturbation is when all coordinates are zero or negative and at least one coordinate is a non-zero negative value. A positive perturbation is when all the coordinates are zero or positive and at least one coordinate is non-zero positive.
When a full dose abemaciclib did not prevent disease progression, some abemaciclib was replaced by ribociclib in the pattern. Replacement was consistent with previous formulas. Goal was greater effectiveness with lower ALT. This effort was successful.
There is no claim that this would be generally true for other patients.
Each treatment cycle has two associated vectors: (1) Treatment Vector and (2) Test Result Vector.
Typical for treatment cycle vectors for CDK4/K inhibitors, vectors are as follows:
Treatment Vector: A 28 coorordinate vector (28D vector), one coordinate for each day. Each coordinate is an (R, A) pair as described previously (or 2D vector). The pair represents the total milligrams of ribociclib and the total milligrams of abemaciclib taken by the patient on the particular day.
Test Result Vector: This is a huge and complex vector. It will be organized in different ways depending on context. Here context is both data available and analysis strategy. All blood test results are included: ALT, eGFR, CA 15-3, WBC, MCH, Chloride, Potassium, etc. Typical for Dana-Farber are 60 different blood test readings. There can be many additional readings. For example, a subvector of SUV numbers: each coordinate representing a tumor site or a vector of the three highest, listing highest to lowest. If there are multiple days on which tests were taken, the previous might be a structured subvector associated with a particular day.
It is good to comprehend the complexity, but also good to start simple. The only test results for a treatment cycle might be blood test results, where say 25 are the most interesting.
With the previous simplification, from a treatment cycle starting point, there might be a 28D planned treatment vector producing a 25D test result vector at the end of the cycle.
Central Question: Given knowledge of past dose and lab test vectors, what is the projected lab test perturbation vector 28 days in the future for any choice of 28 day dose perturbation vector?
The traditional approach employs multiple cohort study arms with many individuals per arm. There are multiple years for the study. Test structure is artificial by design. Cambridge Precision Dosing Perturbation Model takes different approach to this familiar problem. Models are derived from disorganized “real life” data. Like a feedforward network, models and weights are updated as new information from any patient is added to the model data. Goal is finding reproducible patterns present in the data.
There is a change in the question. The traditional approach has the focus of determining whether a particular general approach is good for a selected category of patients over the long haul. The Perturbation Model tries to determine the impact to a particular patient if a particular perturbation is done now. Whether change is better or worse is part of the model, but the path to improved outcome is to better predict what the change will be in detail.
The perturbation model allows for the questions of patients, care givers, researchers and other to ask all their questions simultaneiously and obtain immediate analysis results. From the analysis results, choices are made and new data is feed back into the models. Data for the models can contain non-patient data such as in vitro studies of cells and organs grown in the laboratory. Or even computer simulations.
As a simple example, given past pure R patterns, how fast will ALT fall and CEA increase if a particular negative perturbation is applied? Stopping medication is an example of a negative perturbation being applied, but it is not the only choice.
In practice, patients and their care providers make similar but different dosing choices based on preference and past data. Sometimes patients miss doses. In almost all cases, these are reasonable choices. Practical model development must view this variation as a strength, not a weakness.
Accurate pattern models provide the scientific basis for small steps or perturbations from the tested model values. By the iteration of this process, the set of patterns to which the model applies grows with experience. The risk of a tiny step to the individual patient is low, but there is evolution toward optimal patterns.
It was noticed that some dose patterns were more effective against larger tumors; some dose patterns were more effective against smaller tumors. In this comparison, total dose as measured by D value was the same for compared patterns.
FDA has recently published documents encouraging greater dose innovations3.
“The goal of Project Optimus is to educate, innovate, and collaborate with companies, academia, professional societies, international regulatory authorities, and patients to move forward with a dose-finding and dose optimization paradigm across oncology that emphasizes selection of a dose or doses that maximizes not only the efficacy of a drug but the safety and tolerability as well.”4
From the previous, a model for personalized medicine emerges. One patient may have bad ALT numbers, another RDW-CV. The distribution, genomics and size of tumors varies. Patterns vary between patients and the same patient at different times. A patient’s past patterns can be a better guide than the average from a large study (even when that study uses patients of similar medical history). It is useful to estimate how fast ALT falls when drug dose is reduced or stopped. Is the drug only one factor in liver health?
Except for the 7 patterns from the guidelines, all other patterns are technically “off-label”. We are only discussing patterns that meet the safety formulas of the previous relaxed FDA rules. These relaxed rules are consistent with general FDA oncology drug dosing recommendations.
As a statement of law, the FDA does not restrict doctors from off-label prescriptions.
Malpractice lawyers, insurance companies, drug companies and health institutions all take a (justified) keen and watchful interest in off-label prescriptions. Off-label prescriptions are extremely common, but there is always risk to the patient and their care providers to being first. Perturbation methodology keeps that risk small.
A (one of many) useful general background video reference:
“Clinical Insights: Optimizing, Ribociclib, Abemaciclib and Palbociclib”5, Jan 5, 2026, https://www.youtube.com/watch?v=MbEkR3q9mI0
These approaches are not limited to CDK4/6 inhibitors.
For example, the start of pacilitaxel treatment can trigger a rapid rise in eGFR values (larger eGFR values is generally good).
The foundation to Cambridge Precision Dosing is that medical researchers need to think somewhat less like experimental physicists and somewhat more like astronomers. We cannot create a black hole in the laboritory, but must use data from distant stellar objects — physicists are doing this too.
At the other extreme, Scientific American August 2024 article “The End of the Lab Rat”6 shows other ways old test models for establishing drug safety are changing. Cells and organs grown in the laboratory replace animal studies for both greater speed and greater accuracy.
- FDA: KISQALI® (ribociclib) tablets, for oral use, https://www.accessdata.fda.gov/drugsatfda_docs/label/2024/209092s018lbl.pdf ↩︎
- FDA: VERZENIO® (abemaciclib) tablets, for oral use, https://www.accessdata.fda.gov/drugsatfda_docs/label/2024/208716s017lbl.pdf ↩︎
- Project Optimus: Reforming the dose optimization and dose selection paradigm in oncology, https://www.fda.gov/about-fda/oncology-center-excellence/project-optimus ↩︎
- Ibid. ↩︎
- OncUpdates: Latest Oncology Insights, “Clinical Insights: Optimizing, Ribociclib, Abemaciclib and Palbociclib”, Jan 5, 2026, https://www.youtube.com/watch?v=MbEkR3q9mI0 ↩︎
- Rachel Nuwer, “The End of the Lab Rat”, Scientific American, August 20, 2024 ↩︎