Patent Application
20150177250
The disclosure relates to planning of chemotherapy treatment.
Over the last several years, there has been an increasing understanding that the reasons for the unrealized potential of cancer immunotherapeutics may lay in the state of the immune system in patients with cancer. Most solid tumors contain many non-malignant cells which make up the inflammatory tumor microenvironment. These cells express an immunosuppressive phenotype and act to support cancer growth, invasion, and metastasis, while effectively “shielding” the tumor from the surrounding immune system. An illustrative example are tumor infiltrating regulatory T-cells (Treg) that have been shown to significantly suppress tumor-specific immune responses, thereby promoting rather than suppressing cancer development.
The relationship of cancer and immunity (inflammation) has yielded a number of efforts to correlate measured variables of inflammation with clinical outcomes in patients with advanced malignancies. Measurement of plasma concentration of inflammatory, “acute-phase reaction”, proteins (e.g. C-reactive protein) has been investigated as both a risk factor and a prognostic variable in various human malignancies. Elevated serum levels of several acute phase reactants has been shown to be associated with risk of recurrence, tumor burden, disease progression, presence of anorexia-cachexia syndrome and decreased overall survival in many cancers. The most extensively studied acute phase reactant is C-reactive protein (CRP). Since its discovery in 1930, CRP has been extensively used as a sensitive, albeit nonspecific biomarker of inflammation. In humans, plasma CRP is a positive acute-phase protein, the levels of which rise more than 100-fold in the setting on an inflammatory stimulus. This reflects increased synthesis of CRP, mainly in hepatocytes, induced by pro-inflammatory cytokines such as interleukin-6 (IL-6). After the onset of an acute inflammatory stimulus, CRP can be detected in plasma within 4 to 6 hours with a peak at around 48 hours. CRP half-life is approximately 19 hours and it is fairly constant; therefore the main determinant of the circulating plasma levels is the production rate. Once the inflammation resolves, the CRP plasma level quickly return to normal; unless it is kept elevated by continued production in response to ongoing inflammation and/or tissue damage. Thus, the “acute phase response” is a dynamic process of “up” and “down” regulation of the immune system that fluctuates over time.
In general, the disclosure relates to planning delivery of chemotherapy treatment. In general the systems and/or methods described herein utilize concentration measurements of at least one biological variable to judge the level of systemic inflammation in patients with metastatic melanoma. The systems and/or methods analyze time-dependent fluctuations of at least one biological variable measured in blood samples obtained from clinical patients and determine one or more favorable times for the pharmacological treatment of the patient.
In one example, a method comprises receiving time series data of immune variable concentration for an observed time period for each of a plurality of identified immune variables, detecting one or more periodical patterns in the time series data, fitting a periodic function to the time series data corresponding to each of the plurality of identified immune variables in which a periodical pattern was detected, defining a relative concentration of the fitted periodic function based on a maximum immune variable concentration within the observed time period, a minimum immune variable concentration within the observed time period, and an extrapolated immune variable concentration on a proposed treatment date, defining a relative derivative of the fitted periodic function based on a maximum derivative within the observed time period, a minimum derivative within the same period, and an extrapolated derivative on the proposed treatment date, calculating a treatment prediction parameter based on the relative concentration and the relative derivative, choosing the proposed treatment date such that the treatment prediction parameter is maximized and reporting the proposed date of treatment that maximizes the treatment prediction parameter.
In another example, a computer-readable medium comprises instructions. The instructions cause a programmable processor to receive time series data of immune variable concentration for an observed time period for each of a plurality of identified immune variables, detect one or more periodical patterns in the time series data, fit a periodic function to the time series data corresponding to each of the plurality of identified immune variables in which a periodical pattern was detected, define a relative concentration of the fitted periodic function based on a maximum immune variable concentration within the observed time period, a minimum immune variable concentration within the observed time period, and an extrapolated immune variable concentration on a proposed treatment date, define a relative derivative of the fitted periodic function based on a maximum derivative within the observed time period, a minimum derivative within the same period, and an extrapolated derivative on the proposed treatment date, calculate a treatment prediction parameter based on the relative concentration and the relative differential, choose the proposed treatment date such that the treatment prediction parameter is maximized and report the proposed date of treatment that maximizes the treatment prediction parameter.
In another example, a system comprises a controller that receives time series data of immune variable concentration for an observed time period for each of a plurality of identified immune variables, a periodicity detection module that detects one or more periodical patterns in the time series data, a curve-fitting module executed by the controller that fits a periodic function to the time series data corresponding to each of the plurality of identified immune variables in which a periodical pattern was detected, a treatment prediction parameter module executed by the controller that calculates a treatment prediction parameter based on a relative concentration and a relative derivative, wherein the treatment prediction module further defines the relative concentration of the fitted periodic function based on a maximum immune variable concentration within the observed time period, a minimum immune variable concentration within the observed time period, and an extrapolated immune variable concentration on a proposed treatment date and defines the relative derivative of the fitted periodic function based on a maximum derivative within the observed time period, a minimum derivative within the same period, and an extrapolated derivative on the proposed treatment date, a proposed treatment date module executed by the controller that chooses the proposed treatment date such that the treatment prediction parameter is maximized and a reporting module executed by the controller that generates a report concerning the proposed date of treatment that maximizes the treatment prediction parameter.
The details of one or more examples are set forth in the accompanying drawings and the description below. Other features and advantages will be apparent from the description and drawings, and from the claims.
In general, the example systems and/or methods described herein analyze time-dependent fluctuations of at least one biological variable measured in blood samples obtained from clinical patients and determine one or more favorable times for the pharmacological treatment of the patient. The systems and/or methods determine optimal time(s) for chemotherapy delivery based on serial measurements of the one or more biological variables. In some examples, the biological variables are immune variables. The determination may be patient-specific in the sense that only those biological variables satisfying desired threshold values may be used to determine optimal treatment times for each individual patient.
The measurements of the one or more biological variables may be indicative of the level of systemic inflammation in cancer patients. In the examples described herein, the techniques are described with respect to patients with metastatic melanoma. However, the techniques may also be applied to patients with other types of cancer.
To identify which of the biological variables are indicative of optimal time(s) to deliver treatment to these patients, the systems and/or methods ascertain whether or not one or more biological variables are stable or variable over time in patients with advanced melanoma, and if variable, in what systemic immune context. Presence of any periodical pattern in the data is identified. If a periodical pattern is detected, then curve-fitting is applied to the time series data for each patient to determine the best fit variable function for each of the measured biological variables.
Once the best fit variable function is established, the treatment planning techniques described herein therapeutically utilize the variation of one or more biological variables over time information and devise a treatment strategy which, by using timed administration of conventional cytotoxic therapy (chemotherapy), may augment anti-tumor immunity and affect clinical outcomes.
In an example clinical trial described herein, the patient population included patients with unresectable stage IV malignant melanoma. Eligible patients had unresectable, histologically confirmed stage IV disease, age over 18 years, measurable disease as defined by the Response Evaluation Criteria in Solid Tumors (RECIST), Eastern Cooperative Oncology Group (ECOG) performance status (PS) of 0-2, and life expectancy ≧3 months. Both newly diagnosed, previously untreated patients, as well as patients who have had prior therapy for their metastatic disease were enrolled.
Treatment was initiated with temozolomide (TMZ) 150 mg/m2 on days 1-5 on cycle 1 and the dose was increased to 200 mg/m2 for all subsequent cycles if tolerated. Patients were treated every 4 weeks until progression, unacceptable toxicity or patient refusal. Prior to initiation of first chemotherapy cycle, eligible patients underwent peripheral blood testing for immunological biomarkers (immune variables) every day for 5 days followed by a 2-day rest and then continued blood testing for 5 more days. The blood samples were tested for a total of 70 variables; that is, 70 measurements of cytokine concentrations and cell counts in blood samples. The 70 variables are listed in Table 1.
Peripheral blood samples were obtained at baseline and every day for 5 days followed by a 2-day rest and then continued blood testing for 5 more days prior to the first cycle of TMZ chemotherapy. In order to study the global behavior of the anti-tumor immune response, the samples were further analyzed for plasma concentration of 42 different cytokines/chemokines/growth factors and the percentage of 28 immune cell subsets. All biospecimens were collected, processed, and stored in uniform fashion following established standard operating procedures in our laboratory. To reduce inter-assay variability, all assays were batch-analyzed after study completion.
The data was obtained as follows. Peripheral blood mononuclear cell (PBMC) immunophenotyping for immune cell subset analysis. Blood was separated into plasma and PBMC using a density gradient (Ficol-hypaque, Amersham, Uppsala, Sweden). Plasma samples were stored at −70° C., and PBMC were stored in liquid nitrogen. PBMC bio-specimens were analyzed for the frequencies of T cells (CD3+), T helper cells (CD3+4+), CTL (CD3+8+), natural killer cells (NK, CD16+56+), T helper 1 (Th1) cells (CD4+TIM3+), Th2 cells (CD4+294+), T regulatory cells (Treg, CD4+25+FoxP3+), type 1 dendritic cells (DC1, CD11c+HLA-DR+), DC2 (CD123+HLA-DR+), type 1 macrophages (M1, CD14+197+), type 2 macrophages (M2, CD14+206+) and for the activation status of these cell types. Immunophenotyping of PBMC was performed by flow cytometry using FITC- and PE-conjugated antibodies to CD3, CD4, CD8, CD16, CD56, CD62L, CD69, TIM3, CD294, HLA-DR, CD11c, CD123, CD14, CD197, CD206, and B7-H1 (Becton-Dickinson, Franklin Lakes, N.J.). In addition, intracellular staining for FoxP3 (BioLegend, San Diego, Calif.) was performed according to the manufacturer's published instructions. Data were processed using Cellquest® software (Becton-Dickinson, Franklin Lakes, N.J.). In order to access the Th1/Th2 balance we stained PBMC with anti-human CD4, CD294, and TIM-3. The stained cells were analyzed on the LSRII (Becton Dickinson Franklin Lakes, N.J.). The CD4 positive population was gated and the percent of CD4 cells positive for either CD294 or TIM-3 was determined. GLP validation was performed of anti-CD294 and anti-TIM-3 cell-surface immunostaining for the distinction of Th2 vs Th1 cells in ex vivo (unstimulated) frozen PBMC, respectively. The results show that CD4/CD294 positive Th2 cells exclusively produce IL-4 and not IFN-γ upon PMA and ionomycin stimulation. Conversely, CD4/TIM-3 positive Th1 cells exclusively produce IFN-γ and not IL-4 following the same in vitro stimulation. Enumeration of Treg was performed using intracellular staining for FoxP3 of CD4/25 positive lymphocytes.
Protein levels for 42 cytokines, chemokines, and growth factors, including IL-1β, IL-1rα, IL-2, IL-3, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IL-10, IL-12(p70), IL-12(p40), IL-13, IL-15, IL-17, basic fibroblast growth factor (FGF), Eotaxin, granulocyte colony-stimulating factor (G-CSF), granulocyte-macrophage colony-stimulating factor (GM-CSF), interferon γ (IFN-γ), interferon alpha (IFNa2) 10 kDa interferon-gamma-induced protein (IP-10), macrophage chemoattractant protein 1 and 3 (MCP-1, MCP-3), migration inhibitory protein 1α (MIP-1α), MIP-1β, platelet-derived growth factor (PDGF), Regulated upon Activation Normal T-cell Expressed and Secreted (RANTES), tumor necrosis factor α (TNF-α), tumor necrosis factor beta (TNF-b), vascular endothelial growth factor (VEGF), CRP, transforming growth factor beta (TGF-β1), transforming growth factor alpha (TGF-α), growth-related oncogene-α (GRO), macrophage derived chemokine (MDC), CD40 ligand (sCD40L) were measured using the BioRad human 27-plex cytokine panel (Cat #171-A11127, Bio-Rad, San Diego Calif.) as per the manufacturer's instructions. Plasma levels of TGF-β1 were determined using the duoset capture and detection antibodies (R and D Systems Minneapolis, Minn.) as per manufacturer's instructions. Briefly, plasma samples were treated with 2.5 N Acetic acid and 10M urea to activate latent TGF-β1 followed by neutralization with NaOH and HEPES. The activated samples were added to plates, which had been coated with a mouse anti-human TGF-β1. After incubation the wells were washed and biotinylated chicken anti-human TGF-β1 detection antibody was added. The color was developed using streptavidin-HRP and R and D systems substrate kit. Plasma levels of TGF-β1 were calculated using a standard curve from 0-2000 pg/ml.
All plasma cytokine measurements were performed in duplicate. Normal values for plasma cytokine concentrations were generated by analyzing 30 plasma samples from healthy donors (blood donors at the Mayo Clinic Dept. of Transfusion Medicine). A set of three normal plasma samples (standards) were run along side all batches of plasma analysis in this study. If the cytokine concentrations of the “standard” samples differed by more than 20%, results were rejected and the plasma samples re-analyzed.
A periodicity detection process was applied to the time series data for each of the variables to determine whether each cyctokine concentration/cell count follows a periodical variation over time. In one example, a coherence function analysis was used to perform this test. In another example, a periodogram was constructed for each variable, and the significance of peaks may be assessed with Fisher g-statistic. In another example, the test was performed using permutated time test (Pt-test). However, it shall be understood that any one or more of these tests may be applied, and also that many other tests for periodicity may also be applied, and that the disclosure is not limited in this respect.
If a periodical pattern in the data was detected, then the data was applied to a curve fitting process. In one example, CurveExpert 1.4 software (Daniel G. Hyams Hixson, Tenn.) and GraphPad Prizm 4.0 software (GraphPad Software Inc. La Jolla Calif.) were used to construct time-dependent profiles of plasma cytokine concentrations and immune cell counts by fitting data points to selected mathematical functions. Examples of selected mathematical functions are listed below in Table 2. However, it shall be understood that other periodical mathematical functions may also be used, and that the disclosure is not limited in this respect. Both software packages use Levenberg-Maquart (LM) algorithm to solve nonlinear regressions to fit experimental data to a model curve. The correlation coefficient r=√(St−Sr)/St calculated by CurveExpert may be used as the first criterion for goodness of fit, where St considers the distribution around a constant line and is calculated as St=Σ(y-yi)2 and Sr considers the deviation from the fitting curve and is calculated as St=Σ(yi−f(xi))2. GraphPad Prizm was used to obtain R2 values, 95% confidence intervals for the variables of the fitted functions, and 95% confidence bands for the fitted curves. R2 is calculated as R2=1−Sr/St. These parameters may be used as selection criteria in different steps of the analysis as described below. In another example, a Kolmogorov-Smirnov test (K-S test) may be used to assess the goodness of fit.
In one example, Fourier analysis may be applied to select initial parameters in the LM algorithm which best fit the harmonic which represents the strongest signal. In another example, initial parameters may be selected which best fit an average of a number of the harmonics which represent the strongest signals.
Although specific commercially available software packages are described herein to perform the curve fitting analysis, it shall be understood that other software packages or custom software could also be used to perform the curve fitting analysis, and that the disclosure is not limited in this respect.
The purpose of the periodicity detection analysis is to determine whether any of the measured immune variables change in a predictable fashion following a cyclical pattern (dynamic equilibrium of immunity and cancer). Therefore, the goal of the periodicity detection analysis is to assess whether concentrations of plasma cytokines/chemokines and immune cells fluctuate, and if so, to determine whether these fluctuations follow a mathematically predictable cyclical pattern. To that end, the plasma levels for the 70 immune variables (42 different cytokines/chemokines/growth factors and the percentage of 28 immune cell subsets) in serial blood samples collected every day prior to initiation of TMZ therapy were measured in 10 patients with metastatic malignant melanoma. Of the 12 enrolled patients in this example, the number of data points was inadequate for periodicity detection in two patients. Technical reproducibility was assessed by the coefficient of variation among duplicates (average coefficient of variation was 5.13% for 1593 data points).
The curve fitting analysis was performed based on 10 or 9 sequential measurements (time points) for each variable/patient over a period of 12-days. The “goodness of fit” of the measured variables with a mathematically predicted function was estimated statistically using the correlation coefficient calculated by CurveExpert 1.4 software (REF/source). The cut-off criteria for good fit were computed as follows: (a) the frequency distribution of the correlation coefficient was computed across all profiles and all patients; and (b) the value of the 75th percentile (0.86) was accepted as a cut-off to eliminate profiles which did not fit a model well.
As shown in
If the number of data points satisfies the user input cut-off criteria (108), a periodicity test is performed to determine whether the data points in the time series have a periodical pattern (111). If the periodicity criterion are satisfied (11), the process fits the time series data for each immune variable to sine/cosine function (112).
If the data points fit a function (114), the process may compute various parameters indicative of the “goodness” of the fit of the time series data to each of the functions (116). For example, the process may compute a correlation coefficient (R), a Kolmogorov-Smirnov statistic, a standard deviation of the residuals (Syx), 95 and 99% confidence (CI) band of the curve, 99 and 95% CI of the function parameters. The process may further compute the ratios (Standard Deviation)/(Amplitude) and/or (maximum width of the CI band)/(Amplitude); compute the distribution of frequencies of these two ratios; and/or compute the distribution of frequencies of R, Syx, maximum CI band width.
As shown in
The process may next prompt user for input (122). For example, the process may prompt the user to input one or more of the following: (1) Automatic cut-off for R (Yes/No)?; or (2) Automatic cut-off for (maximum width of the CI band)/(Amplitude) ratio (Yes/No)? If the user inputs selection (1), the process may select curves with R≧75 percentile (126). If the user inputs selection (2), the process may select curves with ratio ≧75 percentile (128).
If the user does not enter automatic cut-offs, the process may prompt user for input (130). For example, the process may prompt the user to: (1) Enter cut-off for R; and/or (2) Enter cut-off for (maximum width of the CI band)/(Amplitude) ratio.
The process may then select the immune variables corresponding to the data series which pass the cut-off criteria (132). The process may then compare the list of selected immune variables with lists of pre-defined variables (determined by, for example, the ranked list of immune variables) (134). The process may then create a list with maximum variables present in an intersection of the two lists (134). The resulting list contains those immune variables having the highest correlation with PFS for that particular patient.
Detection of periodical patterns in the data may be performed using algorithms specifically designed to discover periodical trends in short and noisy data series. These algorithms are not required to report periodicity with mathematical accuracy, such as sine curve fitting. Rather, they are expected to detect an overall periodical trend. This requirement stems from the intrinsic properties of clinical data sets and also from our observation that periods and amplitudes of the immune biorhythms vary even within a single time-dependent profile of a single cytokine Intuitively, stochastic methods would be more efficient for the analysis of such data. Example methods that may be used include coherence function analysis, validation of periodograms' peaks with Fisher g-statistic, and/or permutated time test (Pt-test). Other methods may also exist with potentially better or complimentary characteristics. Data series determined to have a periodical pattern may then be subjected to curve fitting in order to do the extrapolation and computation of an S index (described below) at a time point in the future.
In order to establish whether an ordered pattern of fluctuation correlates with clinical outcome (progression free survival or PFS), an index of fitness may be assigned to each variable. Patients may be ranked by the sum of indices. The correlation coefficient between this rank and the PFS may be calculated. In one example, the assigned index was 1 if the profile fitted a function well (correlation coefficient≧0.86) and the function was biologically possible. In this example, functions with infinite growth or infinite decline were considered biologically implausible as their extrapolation produces biologically impossible values (e.g. <0) for plasma cytokine concentrations or cell count frequencies and were assigned an index of zero (0). The index was −1 if a profile did not fit any function. Using these criteria, the sum of these indices was then calculated for each immune variable per individual patient in this example.
For example, if IL-10 concentration dynamically fitted to cosine, rational or logistic functions in 7 patients and fitted an exponential growth (biologically impossible) function in one patient, this would produce a score of 7 (7×1+0=7). Table 3 shows the rank for each of the 70 immune variables in the example clinical trial.
Separate analysis of the remaining eight patients with metastatic cutaneous melanoma resulted in a correlation coefficient between the total individual score and PFS of 0.72. In a similar way scores (sum or indices) were assigned to each variable. In this case indices were summed across patients per individual variable. Table 3 shows the resulting rank for 52 of the example immune variables.
Determining which immune variables correlate with clinical outcome. In order to understand if certain of the measured immune variables of immune function had a greater/lesser impact on survival, as measured by cyclical function, additional analyses may be performed on the 14 variables assigned a score of 5 or greater in the 8 patients with metastatic cutaneous melanoma (see, e.g., Table 3).
As described above for this example, the index assigned to each variable was 1 if the profile fits a function, 0 for time dependent profiles of variables which fitted biologically impossible functions, and −1 if a profile did not fit any function. As the maximum theoretical score of an immune variable was 8 in this example (8 patients), the cut-off of 5 was chosen because it eliminated those variables which fit a function in <50% of patients. In the case of larger trials (more patients) the cutoff may be chosen appropriately. The maximum score obtained for the remaining variables was 7. These included IL-1rα, IL-9, IL-10, IL-12(p70), IL-13, IL-15, IL-17, G-CSF, VEGF, Th2 T-helper lymphocyte subset (CD4/294), CD11c-positive monocytes (CD11c/14), the ratio of polarized M1/M2 macrophages (DD197/CD206) and DR(hi).
The plasma cytokine concentration or the cell count was extrapolated on the day of treatment for the 14 selected variables in the eight patients on variables with a score of 5 or greater. The first derivative of the fitted function on the day of treatment was calculated. The first derivative shows whether the function at that point is increasing (positive value), decreasing (negative value) or is not changing (zero) and the magnitude of the first derivative reflects the magnitude of the trend.
The range of plasma cytokine concentrations/cell counts varied significantly across patients. In order to be able to compare these concentrations in different patients, the concentrations/cell counts may be converted into relative values by using the formula:
relative conc (“conc”)=(Cmax−Cex)/(Cmax−Cmin), where
Cmax is the maximum concentration within the observed time period,
Cmin is the minimum concentration within the same period, and
Cex is the extrapolated concentration on the day of treatment.
The same conversion may be applied to first derivative values. In the cases when both maximum and minimum first derivative were negative the following formula may be applied:
In one example, an S parameter may be used to characterize both the magnitude of change and the trend of the fluctuation for a given variable. The S parameter may be used to find a relationship between the fluctuation of plasma cytokines/immune cellular elements and may be used to predict clinical outcome and guide personalized “timed” chemotherapy delivery. In some examples, this parameter (S) may be obtained by calculating a sum of the relative concentration and the first derivative with the formula:
S=der+conc where
der is relative derivative, and
conc is relative concentration.
The parameter S, as a sum based on both the relative concentration and the relative derivative, takes into consideration both the magnitude of the concentration and the dynamic trend of a given variable at a precise time point in the immune response cycle. This may help to describe the time-dependent fluctuation of a certain immune biomarker more accurately than the protein concentration or cell count alone.
In other examples, a parameter (Π) may be obtained by introducing weight to the concentration and to the first derivative and then taking a sum of weighted values. For example, the parameter Π may be calculated using the following formula:
Π=(der×N)+(conc×M), where
der is the relative derivative,
conc is the relative concentration,
N is weight of the concentration, and
M is weight of the first derivative.
To determine a relationship between concentration and dynamic trend of the variable at the day of treatment with clinical outcome, the process may find the variables with the highest correlation between the parameter S on the day of treatment and PFS. In order to do that, the parameter S may be ranked in descending order for each measured immune variable. For example, if an immune variable does not fit a biologically possible function, then the sum could not be calculated and since 14 immune variables were analyzed and the lowest rank for a sum was 14, it follows was the next lowest rank for a sum which could not be calculated was 15. Because this rank is weighted by the proportion of non-fitted variables in a given patient, a weighted rank calculated as 15*(number of immune variables which do not fit a function)/(total number of measured variables) may be used. The correlation coefficient may be used to assess the association between the rank of each of these 14 variables and the patients' PFS. In this example, two immune variables, the concentration of IL12p70 and the ratio of CD197/CD206 positive cells (ratio of polarized M1/M2 macrophages) had the highest correlation coefficients of −0.73 and −0.62, respectively. This was further supported by a correlation coefficient of −0.83 between the sum of the ranks for these two variables and PFS. Four patients (50%) with the sum of ranks of these two variables below 15 had average PFS of 466, whereas the other four with sum of ranks above 15 had average PFS of 68 (see, e.g.,
To better understand how the concentration of a cytokine or cell count and the trend for increase or decrease of these variables (first derivative of the fitted function) are related to the clinical outcome, the values of these variables in patients with different PFS may be compared. For example, a fitted cosine curve may be computed where all four parameters of the cosine function (a, b, c and d) were average values of the corresponding parameter across patients being compared and a variable being analyzed. The resulting curve represented averaged concentration/cell count dynamics for several patients on a relative concentration scale (calculation of relative concentration is described above). First derivatives of the fitted function on the treatment day were also plotted on a relative scale (
In attempt to further generalize this observation, we compared concentration/cell count ratio and first derivative of on the day of treatment across 8 patients for IL-12p70, CRP and CD197/CD206. The values were compared as relative values for a given variable in each patient.
Patterns of periodicity of sinusoidally fluctuating immune variables. Since a large proportion of time dependent profiles were fitted to cosine curves when a rather non-stringent criterion (the correlation coefficient) was used, we further selected only those data which fitted cosine curves with the value of R2 greater than the 75 percentile. The cut-off value of the correlation coefficient may be calculated by, for example: (a) the frequency distribution of the correlation coefficient was computed across profiles of all 14 variables analyzed; and (b) the value of the 75th percentile (0.91) was accepted as a cut-off to eliminate profiles which did not fit a model well. As a result, seven profiles were eliminated where the cosine function period was longer than the observation time (14 days). Distinct rhythms were evident for the time-dependent fluctuation (days) of the corresponding plasma cytokine concentrations/cell counts. Table 4 shows the periods in days of the eight cosine curves which satisfied the selection criteria in this example. The shortest period is 3 days and all other periods except one are multiples of 3: 6, 9 and 12. One exception in this example is a 4 day period of IL12p70 in patient 1.
The data in Table 4 show that distinct rhythms were evident for the time-dependent fluctuation (days) of the corresponding plasma cytokine concentrations/cell counts, specifically the ratio of polarized M1/M2 macrophages (CD197/CD206) (30), Interleukin-12 (IL-12p70), Interleukin-17 (IL-17), C-reactive protein CRP), CD11c-positive monocytes (CD11c/14) and Th2 helper T lymphocyte cell subset (CD4/294). For the majority of patients/variables, these rhythms followed a predictable pattern which was a multiple of 3 days (3, 6, 9 and 12 days, respectively) for most of plasma cytokines and cell counts. A few patients demonstrated a 4 day periodicity for IL-12p70, IL-1ra and CD4/294.
Determining the number and frequency of blood draws needed to accurately detect sinusoidal fluctuations in immune variables. The extrapolation of the obtained curves (e.g., ref num. 140 in
The process may also output various data (148). For example, the process may output a report/table/plot of extrapolated and/or maximum values of parameter S per variable for a period of 24 days after the last measurement, output a table of ranks or products per immune variable and output a plot of maximum values of parameter S per variable for a period of 24 days after the last measurement.
The memory may also include programming modules such as a periodicity detection module 215, a curve fitting module 216, a reporting module 220, a treatment prediction parameter S module 212 and a proposed treatment date module 218. Periodicity detection module 216 receives time series data of immune variable concentration for an observed time period for each of a plurality of identified immune variables. Curve fitting module 216 receives data which passed periodicity test and fits a periodic function to the time series data corresponding to each of the plurality of identified immune variables. Treatment prediction parameter module 212 performs all of the calculations necessary to determine the treatment prediction parameter S, such as defining a relative concentration of the fitted periodic function, defining a relative derivative of the fitted periodic function and calculating the treatment prediction parameter based on the relative concentration and the relative differential.
Proposed treatment date module 218 may choose the proposed treatment date such that the treatment prediction parameter S is maximized (or the parameter Π). Reporting module 220 may generate screen displays or printable reports including the proposed date of treatment that maximizes the treatment prediction parameter and/or other presentations of the raw data, intermediate data, or final results. The reporting module 220 may allow the user to create customized reports depending upon the format and/or data the user wishes to view.
The system shown in
The example study discussed herein describes the time-dependent (kinetic) relationship between the tumor and host immune response in 10 patients with metastatic malignant melanoma. In that example, the data analysis suggested that most biomarkers show a temporal variation, implying that these immune variables oscillate repeatedly, in an apparent predictable fashion. The temporal variation of measured biomarkers and the pattern of change (and not only the degree of change itself) may help to define an individual's response to illness.
The techniques described herein may provide evidence that rhythms exist in immune responses to malignant disease and suggest the possibility that such rhythms may be relevant to therapeutic success. Disruption of such biorhythms may have clinical consequences. These observations are consistent with our findings that patients with disorganized (non-curve-fitting) anti-tumor immune responses (see, e.g.,
In the example study, distinct infradian rhythms in the fluctuations of most variables fitted to cosine functions which were in fact multiples of 3-4 days. The contribution of circadian variation to the fluctuation of immune variables was minimized by collection of blood samples at approximately the same time of day (between 8 and 10 AM); therefore the rhythms observed in the example study may be unlikely to be influenced by daytime/nighttime schedule.
By extrapolating the principle of chronotherapy to the anti-tumor immune response, it is possible that coupling treatment with these rhythms may improve the therapeutic index of cancer chemotherapy. In order to accurately predict the fluctuation of the immune response and successfully time chemotherapy administration, the magnitude of change in concentration or immune cell frequency and also the dynamic change of a particular immune variable may be considered. In order to better characterize this time-dependent change, the analysis was extended to 42 other cytokines/chemokines/growth factors and 28 immune cell subsets and studied 1593 additional data points measured over 15 days in 10 patients with metastatic melanoma. By using mathematical modeling and curve fitting analysis, an S parameter that describes both the magnitude of change in concentration and the trend for increase or decrease of a given immune biomarker (S=D+C). This S parameter may be used to identify the variables for which application of chemotherapy at a distinct time-point in the immune cycle correlated with improved PFS.
In this example, two variables, concentration of IL12p70 and the ratio of CD197/CD206 positive cells (ratio of polarized M1/M2 macrophages) exhibited satisfactory correlation with PFS in these examples, emerging as potential candidate biomarkers for timed administration of chemotherapy. Other biological variables, including some of those described herein, may also be appropriate biomarkers, depending at least in part upon the patient.
It shall therefore be understood that other immune variables may also, upon further study, exhibit satisfactory correlation with PFS, and that the disclosure is not limited in this respect.
The additional immunomodulatory properties of TMZ (in addition to its anti-tumor activity) may augment immunological responsiveness through destruction of regulatory T cells, disruption of homeostatic T cell regulation, or abrogation of other inhibitory mechanisms. Timed administration of this agent at a particular time-point in the immune response cycle when IL-12 shows a positive trend (2 out of the 4 day period), may selectively suppress Treg who lag behind T effectors in their clonotypic expansion. By that time, effector T cells may have proliferated and become activated and may be therefore less susceptible to the effects of TMZ chemotherapy.
In the example described herein, curve simulations using function parameters obtained in nonlinear regression fitting of cosine curves to the sample data with periods of 3 to 4 days. This simulation sought to (a) further assess the significance of curve fitting to experimental data; and (b) get a more accurate estimate of the minimum number of data points sufficient for reliable curve fitting, which may allow better planning for a future clinical trial.
Based on the extended example simulation data, a list of candidate biomarkers may include, for example, CRP, IL-10, IL-12p70, G-CSF, IL-9, VEGF, IL-1ra, IL-13, IL-15, IL-17, and/or immune cell subsets such as CD4/294, CD11c/14, CD197/CD206, CD206 and DR(hi).
In summary the data suggests that: (a) patients with stage IV melanoma exhibit a dynamic, not static, anti-tumor immune response; (b) an ordered pattern of change in plasma concentration of various cytokines/chemokines/growth factors and immune cell subsets was observed in patients with the longest PFS; (c) the fluctuations of most variables fit cosine functions with periods which are multiples of 3-4 days; and (d) delivery of cytotoxic therapy (TMZ) at a defined time in the biorhythmic immune oscillation appears to correlate with improved clinical outcome. The sum of the relative concentration of an immune variable and the first derivative takes into consideration both the magnitude of the concentration and the dynamic trend of a given variable and could be used to guide personalized “timed” drug delivery. The data presented herein provide the basis for the design of optimal experimental conditions for testing the hypothesis of timed chemotherapy delivery at a specific phase of the immune cycle.
In a more specific example, a cosine curve simulator (CCS) software module generates simulated cosine/sine curves using function parameters obtained in experiments measuring time-dependent concentration of a selected group of proteins in human blood samples. As discussed above, the simulator takes as an input time series measurements of concentrations of biological variables samples drawn from a number of patients. The other input is distribution of frequencies of technical errors of various magnitudes which was also measured in the experiment. The software outputs curves corresponding to 9 mathematical functions fitted to the input data series. Each fitted curve is supplemented with goodness of fit parameters. The software also outputs a table and a plot of probabilities of cosine curve detection as related to the amplitude, function period, frequency of sampling and length of the observation period.
One purpose of the CCS is to assess confidence bounds of the parameters of the data sets (period of observation, frequency of blood sampling, range of detectable periods of concentration fluctuation, range of detectable amplitudes of concentration fluctuation) for detection of data fitting to 9 mathematical functions.
The CCS algorithm may receive input as described above. The average value and standard deviation is calculated for each biological variable (concentration of a cytokine, chemokine, growth factor or a cell count of a specific cell type) across samples. A range of average+/−2 standard deviations may be calculated for each parameter in the cosine function. There are 4 parameters in the cosine function f(x)=A+B*cos(C*x+D): parameter A determines the vertical shift of the curve, parameter B determines the amplitude, parameter C determines the period, and parameter D defines phase shift.
In one example, the range for parameter B is divided into 100 increments, and range for parameter C is divided into 20 increments to produce periods in the range from 1 to 20 days with 1 day increment. The CCS simulates a set of data points (which correspond to concentration of a protein or cell count) for all possible combinations of period and amplitude for each variable. Further, data may be simulated for three periods of observation: 10 day, 15 days and 20 days and for three frequencies of blood sampling: every day, every other day and with 1 to 2 day interval. Such a simulation will generate 936,000 data sets in total (52 variables*100 amplitudes*20 periods*3 observation periods*3 sampling frequencies). Collectively these data sets may be referred to as “Series A”. A signed experimental error is be added to the ideal value of the function. The error value and frequency follows the distribution of error values obtained in the experiment and the sign is random.
R squared (R2) and standard error may be calculated for each simulated data set. The CCS generates a table and a histogram of distribution of frequencies of R2. Further, CSS may generate another series of data sets—“Series B”. Each set of data points in this series may have the same combination of parameters (52 combinations of amplitude, period, observation period, sampling frequency. One combination per biological variable). However, in this example, the value of the function is not calculated by the cosine formula, but rather is a random number. This random number satisfies all above named parameters.
The curve-fitting as described above may then be applied to the simulated data. For example, curve-fitting may be applied to each data set to 9 mathematical functions (linear function, exponential function, exponential association, logistic model, Morgan-Mercer-Flodin (MMF) model, quadratic function, cosine function, rational function, Gaussian model) and reports which data sets fit any of the functions with R squared above 75th percentile cut-off. The list of these data sets (IDs) may then uploaded into the CCS. Using “Series B” the CCS computes p-value for each simulated data set from the uploaded list. CSS outputs a table of simulated datasets with their parameters and associated p-values. These p-values represent the probability that a data set with a given combination of parameters is fitted uniquely to a cosine curve by chance alone.
The techniques described herein also determine sampling frequency, observation period, curve amplitude and period for one or more biological parameters that fit a function to within a desired goodness of fit. These sampling parameters may then be used to determine a schedule for the real-world collection of blood or tissue samples from patients that will be sufficient to adequately determine desired treatment times. Such a sample collection schedule results in a sufficient number of time points to arrive at a sufficiently accurate determination of desired treatment times while keeping the burden for patients as low as possible. In other words, given the maximum possible number of data points, determine sampling frequency, observation period, curve amplitude and period (for periodical function) which fit a function with high probability and not by chance alone.
To do this, a time series of data points with input parameters derived from our clinical data may be simulated.
Since the most of time series of measurements in original experiment fitted cosine curve, more details the properties of R2 distribution for cosine function will now be described. The analysis of the R2 distribution allows identification of conditions (period, amplitude, sampling frequency, observation period, etc.) which predominantly produce true positive and true negative solutions as well as those which produce false positive and false negative solutions. A solution is the conclusion whether or not a time series of data points fits a cosine curve based on the value of R2. Simulated profiles computed by the cosine formula produced true positive and false negative solutions when R2 was high or low correspondingly. Likewise, random profiles produced false positive and true negative solutions. As a result, ranges of R2 values corresponding to high sensitivity and specificity of the solutions can be determined. One of the goals of the simulation study was to determine the cutoff values of R2 which allow to achieve best combination of specificity and sensitivity.
A small number of time series (10185 profiles=0.0088% of the total number of profiles) formed straight lines and were excluded from further analysis. For the cosine profiles, about 81.7% (461998 out of 565821) of R2 values lie in the range 0.980-1.0 (
As expected, the proportion of R2 above the 90th percentile cut-off obtained from fitting cosine profiles is higher for profiles with greater number of time points, that is, longer observation period or frequent blood sampling. This is a limiting factor in a clinical trial because blood samples cannot be practically taken during a long period of time with high frequency. This may call for an experimental design which would be a compromise between clinical requirements and demands of the curve fitting methods. Such a design is an example collection schedule which permits a sufficient number of time points to be obtained but keep the burden for patients as low as possible. An example schedule satisfying these conditions is 5 sequential days when blood samples are collected, then 2 days of rest followed by another 5 days of sample collection. Such a collection schedule will be referred to herein as the “5-2-5 schedule.”
The example 5-2-5 schedule gives 6 degrees of freedom for data fitting to a cosine function.
Since our hypothesis was that maximums of parameter S indicate active state of the immune response to tumorigenesis which is favorable for therapeutic treatment, we sought to identify time periods when maximum number of variables have maximum cumulative value of parameter S. A clustering algorithm, such as modified K-means clustering or other cluster analysis methods (for example EM-clustering, principal components analysis (PCA), self-organizing maps (SOM), or DBSCAN) may be applied to find these time intervals for the time series generated in the 5-2-5 simulation. Machine learning approaches, such as Support Vector Machines (SVM) may be used to select immune variables which are used in clustering. Otherwise, known biological roles (immune activators or suppressors) of the immune variables can be used as selection criteria. In this example, this method identified two days within a 12 day observation period when the cumulative index had maximum value. The same analysis was then performed on the data obtained from patients with long PFS (916 days; Patient #1 and 841 days; patient #4) and short PFS (68 days; Patient #7 and 70 days Patient #10). Time series of three variables were clustered: concentration profiles of IL-1ra, IL-12p70 and counts of CD206+ cells for these four patients. Since time series obtained from the clinical trial had only 7 or 6 data points, we extrapolated 3 or 4 additional data points to match the same number of points (10) as were analyzed in the simulated 5+2+5 data set. The extrapolated values were computed using Fourier analysis. Clustering produced 1-3 days with maximum cumulative value of parameter S for each patient, as shown in Table 6.
It is noteworthy that treatment days were very close to the days identified by clustering in patients who had long PFS (Patients #1 and #4 in
The techniques described herein for selecting one or more immune variables which may be as predictors of patient's response to pharmaceutical treatment, such as chemotherapy. The basic principle of the method is to accumulate and analyze the knowledge on performance of each of the measured variables in each patient in whom the measurements and the treatment were performed. This accumulation is achieved through creation of a database in which time series of measurements and progression-free survival (RFS) time are recorded. In some examples, the algorithm computes and enters into the database the R2 value of the fitting of each time series to the cosine function. Next, frequency distribution of R2 values may be computed and the R2 value of the 75th percentile may be defined. This value may serve as a cut-off for selecting variables in the next steps of the algorithm. Depending on required stringency of variable selection, a higher (or lower) R2 cut-off level may be selected, for example, 80th or 90th percentile (or lower than 75th percentile).
In another example, in order to select immune variables to be used as discriminators in the clustering algorithm, the algorithm may divide the whole range of PFS longevities into the number of bins ten times less than the number of patients. For each bin the algorithm counts profiles of each variable with R2 above the cut-off value and the sum of Π indices on the treatment start date for these variables (see, e.g., Table 7 and Table 8). Next, the linear regression analysis may be performed both on the counts of each variable with R2 above the cut-off value and on the sums of parameter S and the slope of the regression line may be computed. Variables with high positive value of the sum of the slopes (for example, IL-12, IL-1ra and CD206 in Table 7) have positive correlation (PC) with PFS (see, e.g., the graph for IL-12p70 in
Tables 7-9 illustrate data corresponding to example procedures that may be used to select immune variables that will may used as discriminators in the clustering algorithm. The range of PFS time is divided into a number of bins (clusters) 10 times less than the number of patients. In this example there were 100 patients and so the PFS times were divided into 10 PFS bins (see, e.g., the last row of Table 7).
Temporal profiles which fit the cosine function with R2 greater than selected cut-off are counted for each RFS bin and the slope of the regression curve of the counts is computed. Table 7 shows the mean and standard deviation (SD) of the slope values for all variables. These are used to calculate the 75th percentile (mean+0.67×Standard Deviation) and the 25th percentile (mean−0.67×Standard Deviation) of the slope values. In this example, variables for which the slope values were above the 75th percentile include IL-12, IL-1ra, and CD206. Variables for which the slope values were below the 25th percentile include IL-17 and IL-10.
Table 8 shows the sums of parameter S on the first treatment day for temporal profiles which fit the cosine function with R2 greater than selected cut-off are computed for each RFS bin and the slope of the regression curve of the sums is computed. The mean and standard deviation (SD) of the slope values for all variables are computed and are used to calculate the 75th percentile (mean+0.67×Standard Deviation) and the 25th percentile (mean−0.67×Standard Deviation) of the slope values. In this example, variables for which the slope values were above the 75th percentile include IL-12, IL-13 and IL-1ra. Variables for which the slope was below the 25th percentile include IL-17.
Table 9 shows the sum of the slope values computed in Table 7 and Table 8 for each variable. The mean and standard deviation (SD) of the sums for all variables are computed and are used to calculate the 75th percentile (pink) and the 25th percentile (blue) of the slope values. In this example, variables for which the sum of the two slope values were above the 75th percentile include IL-12 and IL-1ra. Variables for which the sum of the two slopes that were below the 25th percentile include IL-17.
Variables with slopes above the cut-off value(s) identified in any one or more of the sums shown in Table 7, Table 8 or Table 9 may be used as discriminators in the clustering algorithm.
Time-dependent fluctuations' profiles of the selected immune variables are used to determine the optimum time of chemotherapy delivery by using the following method. Cosine profiles of the fluctuations may be clustered with the aim to find time window, during which the frequency of peak values of the parameter S is the highest. The clustering is done by the K-means method with modifications. K-means clustering requires a priori knowledge of the number of clusters in which the objects (profiles) will be grouped. By our method, the number of groups is determined from the number of full function periods which fit into one observation period. The maximum possible number of groups equals the maximum number of function periods and the minimum number of groups equals the minimum number of function periods which fit into one observation period. The algorithm computes the number of clusters for the whole range of integers from the maximum to the minimum numbers. For each iteration (number of clusters) and for each variable the algorithm calculates the dates when the parameter S has maximum value. These dates are used as centroids for K-means clustering. Since the result of K-means clustering depends on the order of initial centroids, our modification performs clustering for all possible combinations of centroids and then computes the date when the sum of indices for all clustered cosine profiles was maximal. Next, the algorithm computes the dates with maximum sum of relative parameter S across all possible combination of centroids and all numbers of clusters. These dates are outputted as optimal dates for chemotherapy application for a given patient and a given set of immune variables (
Although in
The example systems and/or methods described herein analyze time-dependent fluctuations of at least one biological variable measured in blood samples obtained from clinical patients and determine one or more favorable times for the pharmacological treatment of the patient. The systems and/or methods determine optimal time(s) for chemotherapy delivery based on serial measurements of one or more biological variables. In some examples, the biological variables are immune variables.
Each new series of experimental measurements may be processed according to the described workflow. This iterative computation of simulated parameters based on ever growing experimental evidence may iteratively enhance statistical power accuracy of p-values and overall precision in detecting functions to which the data fits. This, in turn, may enhance the accuracy of prediction of the date for most efficient chemotherapy treatment.
The techniques described in this disclosure, including functions performed by a processor, controller, control unit, or control system, may be implemented within one or more of a general purpose microprocessor, digital signal processor (DSP), application specific integrated circuit (ASIC), field programmable gate array (FPGA), programmable logic devices (PLDs), or other equivalent logic devices. Accordingly, the terms “processor” “processing unit” or “controller,” as used herein, may refer to any one or more of the foregoing structures or any other structure suitable for implementation of the techniques described herein.
The various components illustrated herein may be realized by any suitable combination of hardware, firmware, and/or software. In the figures, various components are depicted as separate units or modules. However, all or several of the various components described with reference to these figures may be integrated into combined units or modules within common hardware, firmware, and/or software. Accordingly, the representation of features as components, units or modules is intended to highlight particular functional features for ease of illustration, and does not necessarily require realization of such features by separate hardware, firmware, or software components. In some cases, various units may be implemented as programmable processes performed by one or more processors or controllers.
Any features described herein as modules, devices, or components may be implemented together in an integrated logic device or separately as discrete but interoperable logic devices. In various aspects, such components may be formed at least in part as one or more integrated circuit devices, which may be referred to collectively as an integrated circuit device, such as an integrated circuit chip or chipset. Such circuitry may be provided in a single integrated circuit chip device or in multiple, interoperable integrated circuit chip devices, and may be used in any of a variety of applications and devices.
If implemented in part by software, the techniques may be realized at least in part by a computer-readable data storage medium comprising code with instructions that, when executed by one or more processors or controllers, performs one or more of the methods described in this disclosure. The computer-readable storage medium may form part of a computer program product, which may include packaging materials. The computer-readable medium may comprise random access memory (RAM) such as synchronous dynamic random access memory (SDRAM), read-only memory (ROM), non-volatile random access memory (NVRAM), electrically erasable programmable read-only memory (EEPROM), embedded dynamic random access memory (eDRAM), static random access memory (SRAM), flash memory, magnetic or optical data storage media. Any software that is utilized may be executed by one or more processors, such as one or more DSP's, general purpose microprocessors, ASIC's, FPGA's, or other equivalent integrated or discrete logic circuitry.
Various examples have been described. These and other examples are within the scope of the following claims.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2013/046339 | 6/18/2013 | WO | 00 |
| Number | Date | Country | |
|---|---|---|---|
| 61660927 | Jun 2012 | US |
