Towards evolutionary guided precision medicine of acute myeloid leukemia and Fanconi anemia associated bone marrow failure.

Introduction
Carcinogenesis is a multistep evolutionary process in which a series of genetic changes confer corresponding phenotypes in an initially normal cell, creating a transformed founder cell.1,2 For established cancers, the acquisition of resistance to multiple therapies is also a multistep genetic evolutionary process in those instances where the therapies are non-cross resistant. In this perspective, we review two parsimonious mathematical models relevant to devising personalized strategies for preventing or delaying both carcinogenesis and multidrug resistance, new approaches for their experimental validation and refinement, and emerging possibilities for clinical application. These mathematical models can produce highly adaptive personalized sequences of non-cross resistant therapies as clinical options. The recommendations are proactive, that is, based on risk probabilities, as opposed to reacting to clinical events such as cancer diagnosis, relapse or minimal residual disease observation. By acting in advance, it may be possible to more effectively prevent or delay both of these multistep processes.
We first describe efficiency of carcinogenesis (EOC),3–5 designed to quantify the importance of any variable potentially affecting the ability of transformed founder cells to arise. EOC has been used to examine the role of genetic instability in carcinogenesis.
Next, we discuss dynamic precision medicine (DPM),6–10 designed to produce personalized optimal sequences of two or more non-cross resistant therapies for the delay of a multistep evolutionary process. Simulations indicate that DPM can significantly prolong survival and increase cure rate in a broad range of simulated clinical presentations by delaying the acquisition of multitherapy resistance. DPM can be adapted to delaying carcinogenesis as well.
Laboratory evaluation of these approaches is in progress, and requires the ability to detect rare events that evolve and become apparent over long time periods. We describe new methods to facilitate this experimental validation and refinement.
Two potential applications of EOC and DPM are briefly discussed: (1) slowing carcinogenesis in laboratory models of Fanconi anemia-associated bone marrow failure (FA BMF), while also minimizing transfusion requirements, and (2) optimizing survival in relapsed acute myeloid leukemia (AML).
Finally, we discuss potential obstacles and limitations to these approaches, and their potential significance if successful.

Efficiency of carcinogenesis
Loeb and colleagues11 posited in the 1970s that cancers must be genetically unstable to accumulate the multiple genetic alterations required for cellular transformation to establish a founder cell within a human lifetime (Table 1). “Mutator mutations,” that is, mutations in the genetic machinery that preserves genomic integrity, were considered essential. This led to a debate about whether mutator mutations were in fact essential, based on mathematical modeling attempting to match predicted incidence of cancers with epidemiologic data.15–18 These efforts effectively yielded information about the number of steps in the process based on the shape of the incidence curve,18 but conflicting results regarding essentiality of mutator mutations, based on the uncertainty in various input parameters.28
EOC took a different approach, noting that to account for a large effect of immune surveillance in eliminating cancers before diagnosis, the rate of production of founder cells should greatly exceed that of actual cancer diagnoses. It hypothesized that numerous pathways to cancer existed, but those that could produce founder cells most efficiently would be proportionately more likely to be responsible for overt clinical cancers. The problem was then reduced to computing the relative efficiencies of mutator and nonmutator pathways, and many unmeasurable parameters canceled in the ratio. This was one of several approaches to answering very focused questions with limited information, discussed as Focused Quantitative Modeling (FQM).5
EOC assumes that cells can tolerate considerable genetic instability without overall reduction in their growth potential, the theoretical basis of which was provided in a paper on deleterious mutations22 and validated experimentally in yeast experiments showing that massive increases in genetic instability were tolerated before yeast finally experienced fitness reduction.23
If a given cancer requires C rare events (mutations) for transformation, its efficiency is then compared to that of the mutator pathway that requires an additional rare event (the mutator mutation) while some or all of the other events are speeded up (depending on when the mutator mutation occurs).
Inputs for the model include measurement of mutation rates of intermediate states using highly accurate duplex DNA sequencing29,30 and a dedicated mathematical model for extracting rates,31,32 and measurement of growth rates for these states.
The approach predicted the ubiquity of genetic instability mutations in cancer, recently verified in a large survey.24 The assumption that there are multiple pathways to cancer (i.e., convergent evolution) was validated in renal cancer.19 The model also predicted that cancer types which, based on epidemiology, require fewer mutations to produce a founder cell would not generally exhibit genetic instability since the additional rare mutator mutation would have fewer mutational steps to accelerate, and therefore, not be “worth it” (Supplementary Table S1). Indeed retinoblastoma, for which C was estimated at 2, has a lower mutation burden.33 Subsequently, another similar approach using different computational ­methods yielded similar conclusions.34
EOC can evaluate the effect of various therapies on transformation of a normal stem cell into a founder cell. Consideration of differentiated cells resulting from the stem cell intermediates will need to be added when estimating the effect of therapy on other relevant phenotypes.
The assumptions underlying EOC are described in Table 1. Molecular evidence exists that AML develops through a series of oncogenic mutations and that multiple pathways to forming an AML founder cell exist.12,13

Dynamic precision medicine
DPM6 is potentially a significant enhancement of current precision medicine (CPM) in that it explicitly considers genetic heterogeneity within individual cells, and their evolutionary dynamics in making therapy recommendations. CPM matches optimal therapy to the consensus molecular characteristics of a bulk sample and treats the patient with this same therapy until cancer progression or relapse, repeating the process at that time. However, inevitable therapy resistance leads to diminishing clinical returns with each iteration of this process. Indeed, recent deep sequences studies at very high sensitivity and accuracy and novel mathematical approaches31,32 indicate that, in any cancer with sufficient cell numbers to enable diagnosis, a small subclone with any given mutation of interest exists, unless that mutation substantially reduces the fitness of the cell, leading to the conclusion that resistance to any single therapy or cross-resistant combination therapy is pre-existing. Moreover, to an even greater degree than previously suspected, increasing cancer burden will eventually lead to cells that are simultaneously resistant to multiple non-cross resistant therapies through independent resistance mutations.31,32 These multiply resistant subclones may often arise from “hypermutator” subclones that have acquired one or more additional mutator mutations beyond that seen in the bulk, and thus, evolve more quickly.35
Given two or more non-cross resistant therapies (each of which may be combinations), DPM recommends a personalized highly adaptive treatment sequence for each individual. Adapting as often as every 6 weeks (2 chemotherapy cycles), it bases its recommendation on proactive prediction of the kinetics of future emergence of drug resistance. Rather than always prioritizing immediate cytoreduction, it often prioritizes prevention of multiple drug resistance, leading to decreased medium to long term relapses. The clinician determines when to over-ride this recommendation based on the immediate need for cytoreduction (e.g., in AML induction, an emergent circumstance). By not treating with the same therapy until relapse, DPM confronts the cancer with a more challenging landscape in which to adapt.36 Simulations of several million virtual patients representing a broad range potential presentations across oncology suggest substantial improvements in both median survival and cure rate are achievable (Figure 1).6,8,10 A searchable library of these virtual patients is available, and provides a resource for clinicians for rapid evaluation of any patient in an interactive “digital twin” paradigm involving real time interaction between patient data and therapy recommendations37 (Supplementary Figure S1), in which the simulation is focused on virtual patients (“digital twins”) that best match the patient’s clinical course up to that point in time.
DPM applications require naming at least two non-cross resistant therapies of interest (“drugs 1 and 2”), to attack subclones with different drug sensitivities. These may be single agents or simultaneous combinations. In the event that dose reduction is required to give “drugs 1 and 2” simultaneously, it may recommend bursts of full dose monotherapy, while maintaining high dose intensity of both agents due to frequent adaptation. Safety of any recommended single agent or combination dose must have been previously established for DPM to recommend it.
DPM also requires measurement of growth rates, mutation rates, and drug sensitivities of sensitive and resistant subclones. Based on these measurements, simulations have shown that patients who will benefit from DPM can be selected for clinical study.10 The predictive biomarker is, thus, based on dynamic properties rather than the conventional static ones. Methods development to measure these with rapid turnaround and subclonal resolution is ongoing. A table of the assumptions and limitations of DPM is given in Ref. [10].
Both stem cells and differentiated cells will have resistant subpopulations, and indeed stem cell and differentiated compartments may differ from each other in chemosensitivity. Thus, minority resistant subclones may be from either compartment. Model enhancements will need to account for interconversion between these compartments and their differing kinetics and drug sensitivity profiles.

Experimental evolution
Theoretical studies suggest that rare cells comprising 1 in 100 000 or less of cancer burden can ultimately be an important source of late relapse and mortality.6 Acquisition of multiple independent resistance mutations and subsequent outgrowth of rare subclones takes time, and most laboratory models lack sufficient cell number and experimental duration to study these effects.9,32
Genomic studies of human tissues have revealed the extent to which genetic heterogeneity impacts cancer evolution.31,38 Such studies are typically performed on limited tissues representing brief, temporal snapshots with varying clinical collection methods, disease states, and other parameters. As a parallel approach for quantifying subclonal dynamics, experimental evolution within a programmable bioreactor enables comprehensive data collection from large tumor cell populations (millions to billions of cells) evolving under highly controlled environments and selective pressures. This technology sustains cell populations for weeks to months, thereby providing the ability to “replay” evolution from a defined starting point and facilitating integrated molecular analyses of cancer evolution. Importantly, combining this approach with cellular barcoding39 provides insight on which subclones acquire novel mutations and phenotypes,40 thereby revealing clones that are evolutionary dead-ends versus clones fostering treatment resistance.41

Fanconi anemia-associated bone marrow failure
Fanconi anemia is an inherited disease of BMF and cancer predisposition, such as AML and solid tumors due to defects in DNA repair.25,26 FA BMF is treated by stem cell transplant, which can be toxic, suggesting need for nontoxic alternatives to support hematopoietic function and as a bridge to future interventions (such as gene therapy).
FA gene defects activate intrinsic innate immunity, which retards stem cell fitness, resulting in BMF.27,42 Strategies to improve bone marrow function must not promote AML, yet relaxation of hematopoietic drive might result in diminished risk.
The adult acquired version of BMF is myelodysplasia (MDS), with a heterogeneous array of mutated etiologic factors that underly stem cell failure, subsequent pancytopenia, and leukemia risk. The different genetic characteristics in MDS confer accompanying variability in risk of progression to AML. For example, TP53 mutant MDS has a variable but generally higher risk of AML progression and blast formation, lower remission rate in the context of therapy, and poor survival within 3 years. Conversely, specific MDS marked by SF3B1 mutations may entail a risk of AML of as little as 3%. Given the heterogeneity of MDS14 (Supplementary Figure S2), the ability to predict clonal evolution, and thus, guide therapeutic intervention using conventional therapies is quite challenging. These evolutionary outcomes may also depend in part on the microenvironment, metabolism, and oxygen.
Experimental, computational, and mathematical tools may predict how personalized interventions that enable rational treatment of FA patients with BMF while avoiding deleterious consequences such as leukemogenesis. Experimental bioreactor technologies can define relevant evolutionary pathways and parameters, EOC can isolate factors that may promote or retard leukemogenesis, and DPM modified for this application can define optimal treatment sequences.

Relapsed AML
Significant evidence that leukemia relapse emerges from minor subclones has been previously reviewed.43 The subset of AML cells believed to confer drug resistance and relapse has been named “leukemia stem cells. (LSCs)” This population can be demonstrated to propagate leukemia in secondary and higher transplants of primary AML in xenograft models. A higher proportion of CD34+CD38neg or low leukemia stem cells was associated with a reduced rate of complete remission.44 High throughput drug screening of sorted leukemia stem cells (CD34+CD38-CD123+) demonstrated that the LSCs exhibited relative drug resistance for the agents most often used to treat AML, both newly diagnosed and relapsed, as compared to the entire CD34+ blast population.45 Moreover, a recent publication highlights that for all acute leukemias, a primitive stem/progenitor cell (HPSC) subpopulation identified by integrating cell multi-omics, bulk RNAseq and clinical features is highly drug resistant. These investigators characterized a core transcriptional regulatory network in the HPSCs that can predict responses to therapy.46 Techniques are now available to identify the mutational composition and phenotype of individual AML subclones by single cell mutation analysis47 that allows us to track serial drug resistance with serial testing, and to accurately compare prevalence of selected subclones at diagnosis and relapse by duplex sequencing.48 These methods enable observation of the mutations contributing to the drug resistance of the leukemia stem cell fraction.
Using these techniques as well as high throughput drug screening to identify drugs to treat emergent subclones, pilot studies of personalized treatment of relapsed patients are ongoing. DPM can be used to further optimize these personalized treatments when there are multiple possible treatment options from which to choose, by creating personalized sequences of these treatments based on evolutionary dynamics.

Clinical translation
DPM and EOC are at an early stage of development, requiring laboratory validation and clinical testing. DPM growth and drug sensitivity parameters can be measured with standard cell culture techniques. Mutation rates can be measured using duplex sequencing and accompanying analytical approaches.31,32
Clinical translation in the future will optimally require widespread availability of these techniques and rapid turnaround of results, including standardization and commercialization of laboratory assays. Recently, it has been discovered that most of the benefit of DPM is delivered in the first two 6 week adaptation periods (termed “moves” in analogy to chess), reducing the cost and invasiveness of the approach, and allowing clinical window studies of DPM.10 This discovery points solid tumor DPM research towards the neoadjuvant period, in which tissue availability is optimized (Figure 2), whereas in liquid cancers blood sampling may suffice. This suggests the possibility for community referrals to tertiary centers for a DPM window, followed by return of the patients to community physicians.
Recent work has introduced a randomized stratified clinical trial design for evaluating DPM, and also shown that DPM is surprisingly robust to uncertainty in drug sensitivity estimates.49 DPM can also be enhanced to address sensitivity issues in rare subclone detection by exploiting evolutionary theory to estimate the risk of undetected subclones.50
To provide a practical, user-friendly method for both simulating DPM clinical designs and making treatment decisions for individual patients we developed a user-facing R Shiny application, which does not require any mathematical or computer science background for use, and runs on the physician’s laptop.
One section allows clinical trialists to input various trial designs for testing DPM and provides performance evaluation and expected results for the clinical trial. The ability to simulate hypothetical trials with ease will greatly improve feasibility of testing the DPM concept for biopharmaceutical and academic clinical trialists.
The second section, under development, is designed to produce personalized treatment recommendations for individual patients. It allows the physician to enter 2 or more non-cross resistant therapies (which could be combinations themselves), the DPM input parameters (to the extent feasible and available), and/or more standard clinical measurements from which possible ranges of these parameters may be inferred by matching the patient’s clinical course to virtual patients with similar outcomes (see legend to Supplementary Figure S1). Inference of DPM parameters from standard clinical data may be further enhanced by outcomes data from public databases. The system will determine whether there is a likelihood of benefit from DPM, projected survivals on DPM and CPM, the recommended DPM therapy until the next adaptation point, the degree of confidence in the recommendation, and the evidence supporting it, and report this back to the clinician in real time.
This individualized cancer therapy decision tool has the potential to change oncology practice. The decision tool would allow providers to make an informed decision that accounts for the potential of resistance developing in their patient’s cancer in the future and is not solely based on cross-sectional data on lesion size. The physician would then put this recommendation into context, and would make the ultimate treatment decision (Supplementary Figure S3).

Discussion and conclusions
A variety of other approaches for evolutionary guided precision medicine exist, of which we cite some of the first.51–53 EOC and DPM are approximations not fully accounting for the complexity of cancer or its evolution. CPM is an even simpler approximation, which does not explicitly include dynamics or single cell heterogeneity, but it has provided significant improvement in clinical outcomes. EOC and DPM are parsimonious models that nonetheless have promise to further improve clinical outcomes as indicated by extensive simulations. Parsimonious models are less complete than more complex models but have the advantages of interpretability and potentially greater ­generalizability and ease of translation.54
Laboratory and clinical testing of EOC and DPM in specific clinical applications is essential. Recent advances in experimental evolution, sequencing technology, drug sensitivity prediction, and patient selection using a dynamic evolutionary classifier have made this possible, opening up new future ­possibilities for greater optimization of personalized therapy.

Supplementary Material
szag008_Supplementary_Data