CSTS Healthcare

Personalized Medicine: Developing metrics for comparing treatment protocols, relating treatment protocols to personal biological data; comparing treatments selected by AI with treatments selected by physicians


Cancer is a disease that affects 14M people each year. While we have had some success with specific cancers, many other patients are put on a rollercoaster of emotion through cycles of remission and relapse. At this point, there is abundant scientific evidence that tells us cancer is driven by multiple genes working in concert.

Traditionally in medicine, drugs are designed and approved by testing on large populations. This type of results in only a portion of the population actually responding to a given therapy. However, different patients respond differently to the same drugs. Typically, on a large double-blind clinical trial with thousands of patients and two arms, everyone on one arm receives the same therapy A, while everyone else on the other arm receives exactly the same therapy B.

In the past several decades, we have learned that cancer is a highly heterogenous disease with multiple genes involved in cancer progression, and therefore requires combination therapies tailored to each individual’s life history and genomic profile. Based on multiomic profiles one can model the cancer biology: which genes are driving the cancer, and which of the 10 hallmarks of cancer are active.

This allows the design of personalized therapies that are unique to every individual. However, this also presents a novel statistical problem, called the N-of-One problem, where it is difficult to achieve statistical power. Because each patient in a trial is receiving a differing, unique therapy, these therapies are not obviously comparable. One recent study, the I-PREDICT trial attempted to provide a comparison of personalized combination therapies when the therapy was only partially adopted. They did not directly evaluate personalized therapies, but simply determined that when a given therapy targeted more mutations, it was correlated with better outcomes.

Problem Statement

We have developed a computational system that identifies a personalized cancer therapy for every cancer patient, given their unique set of DNA and RNA. For each patient, our system identifies the best of set of target genes and active hallmarks, each of which have sets of available therapies associated with them. In a clinical setting however, a clinician prescribes the therapy they believe is most appropriate for a given patient. In this context, if there are 100 patients, there are 100 unique therapies. Moreover, the therapy a patient receives may not match what our system identified as the best therapy. This problem then breaks down into the following sub-problems:

  • We would like to construct a set of similarity measures that allow us to compare our Aiomic therapies with those actually given by oncologists.
  • Given the entire set of patients, can we quantify adoption rates of Aiomic therapies
  • Can we measure outcomes for Aiomic therapies as to whether they succeeded, partially succeeded or failed?

In contrast to the I-PREDICT approach, for our problem, we are interested in not just matching gene targets, but evaluating the success of Aiomic therapies when the given therapies are not an exact match. The problem can be stated as follows:

  • A trial consists of N patients.
  • Each patient receives a therapy T, that consists of a set of drugs D that target zero or more Genes, and zero or more Hallmarks.
  • A therapy can be viewed as set of drugs, T = Drugs = $\left\{d_1, \ldots, d_n \right\}$
  • Each drug can be related to zero more targets, where a target is either a gene or a hallmark.
  • Each therapy is associated with one of three Outcome = {succeeds, partially succeed, or fails}
  • Aiomic identifies the best therapy Aiomic-Therapy for each patient
  • A clinician administers a given therapy Given-Therapy for each patient


  • How can we compare Aiomic-Therapies to Given-Therapies? For example, would it simply be the % intersection between the set of targets and hallmarks that are covered by the drugs in each therapy?
  • Across all patients, given the set of Aiomic-Therapies and Given-Therapies, can we say how often Aiomic therapies were adopted? To what degree?
  • Across all patients, given outcomes, when a Given-Therapy is not exactly the same as a Aiomic- Therapy, and has non-zero overlap, can we assign an outcome to the corresponding Aiomic- Therapy?
  • How much of a Given-Therapy outcome can we allocate to the Aiomic-Therapy in cases where there is partial overlap between the recommended therapy and the given therapy?
  • If the adoption rate was 30%, is 30% of the outcome due to the recommendation? That is to say, if only one of the drugs from the recommendation were used in the given therapy, what % is attributable to the recommendation.
  • Can we create a predictive model for partial adoption of our therapies?


See: 123456

  1. Klement et al. Future Paradigms for Precision Oncology. Oncotarget (2016). 7:46813-46831. https://doi.org/10.18632/oncotarget.9488 ↩︎

  2. Hanahan, Douglas et al. Hallmarks of Cancer: The Next Generation. Cell, Volume 144, Issue 5, 646 – 674. https://doi.org/10.1016/j.cell.2011.02.013 ↩︎

  3. Lillie EO, Patay B, Diamant J, Issell B, Topol EJ, Schork NJ. The n-of-1 clinical trial: the ultimate strategy for individualizing medicine? Per Med. (2011) 8:161–73. 10.2217/pme.11.7 ↩︎

  4. N-Of-One Trial- https://en.wikipedia.org/wiki/N_of_1_trial ↩︎

  5. Offord, Catherine. N-of-1 Trials Take on Challenges in Health Care– The Scientist. (2019) July/August. https://www.the-scientist.com/features/n-of-1-trials-take-on-challenges-in- health-car-66071 ↩︎

  6. Sicklick JK, et al. Molecular profiling of cancer patients enables personalized combination therapy: the I-PREDICT study. Nat Med. 2019 May;25(5):744-750. doi: 10.1038/s41591-019- 0407-5. Epub2019 Apr 22. PMID: 31011206; PMCID: PMC6553618. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6553618/ ↩︎

Ali Hashemi
Ali Hashemi
Cofounder CSTS Health Care
Peijun Sang
Peijun Sang
Assistant Professor
Natalia Accomazzo Scotti
Natalia Accomazzo Scotti
Postdoctoral Fellow
Eric Rozon
Eric Rozon
Graduate Student
Ellie Thieu
Ellie Thieu
Math Researcher, Data Analyst