Please Note: Blog posts are not selected, edited or screened by Seeking Alpha editors.

Northwest Biotherapeutics: Model Indicates Efficacy For DCVax-L

|About: Northwest Biotherapeutics, Inc. (NWBO)

Summary

Modeling the NWBO clinical trial using publicly available information demonstrates DCVax-L treatment's potential efficacy

Multiple control arms from previous clinical trials were used to model the DCVax-L control arm

Model Treatment mPFS range: 17 to 24 months; Model Treatment mOS range: 27 to 32 months;

Model Treatment PFS efficacy advantage: 7 to 12 months

Model Treatment OS efficacy advantage: 6 to 12 months

Northwest Biotherapeutics (OTCQB:NWBO) is in the process of concluding its Phase 3 clinical trial which is testing DCVax-L on GBM, the most common and deadliest of brain cancers. The stock has been decimated over the past couple of years, and the reasons for the stock dismal performance stem from numerous sources including:

  • questionable business structure and practices,
  • a prolonged and costly Phase 3 clinical trial,
  • inflammatory one-sided articles and social media slander,
  • a temporary FDA enrollment halt, which has since been lifted but not explained, and
  • toxic financing at depressed stock prices.

There have been numerous other articles addressing the said bullet points, and the reader is urged to perform their own due diligence with regards to these and other matters.

With that said, a model has been developed using publicly available information that demonstrates the NWBO DCVax-L treatment’s potential efficacy. The model uses the following publicly available or assumed information:  

  • The DCVax-L enrollment curve released by the NWBO
  • 331 patient clinical trial, with a 3-month enrollment period from surgery
  • Progression Free Survival ("PFS") clinical trial end-point: 248 events and assumed to be reached in October 2016
  • Overall Survival ("OS") clinical trial end-point: 233 events, and assumed to be reached sometime in July 2017
  • Kaplan-Meier (KM) Survival curves from previous clinical trials which are used to model the NWBO control arm.

The reader is cautioned that there is no way to know exactly how the actual DCVax-L control arm is performing, because it is a double-blinded clinical trial. Additionally, NWBO did not specifically give the exact date the 248 PFS events were reached. A NWBO press release in February of 2017 stated: “The Company announced that the partial clinical hold on the Trial has been lifted by the FDA, and that the Trial has accumulated a sufficient number of events toward the progression-free survival (“PFS”) endpoint, but not yet for the overall survival (“OS”) endpoint.”  Many optimists have speculated the PFS endpoint was reached sometime in December or January.  To appease the critics, a slightly earlier date was used as explained henceforth.

NWBO provided guidance in another PR that the 233 OS event endpoint would be reached in mid-July to mid-summer, and a NWBO representative stated that 100 patients were alive (i.e. approximately 231 OS events) in a June 2017 conference.

The historical difference between OS and PFS is approximately 9 months. There are 9 months between October 1st and July 1st, therefore October 2016 was selected as the PFS model trial end date; and per NWBO’s guidance, July 2017 was selected as the OS model trial end date. These dates could be off a month or so in either direction.

NWBO has not yet confirmed that the 233 OS events threshold has been reached, but in all likelihood, it has. Some have speculated that NWBO is allowing the clinical data to mature before moving to data-lock.

The number of events and the dates those events were reached is important because this information is being used by the model to ascertain the performance of the clinical trial.  And although these are only two data points, the data points are extremely useful pieces of information, because the events are accumulative, which means the total value gives us a good indication how the trial is performing as a whole.

The model relies heavily on the clinical trial data from previous trials to model the performance of the NWBO control arm. The reader is cautioned that the inclusion and exclusion criteria from the previous clinical trials is different than the actual NWBO trial. The model will produce less-favorable results if the control arm is performing better than expected.

Additionally, the actual NWBO clinical trial has a “cross-over” feature which allows the placebo patients to switch to the DCVax-L treatment upon disease progression. The model presented herein does not account for the effect of the cross-over. The cross-over feature will make it more difficult for the trial to detect a statistical difference between the control arm and the treatment arm if the treatment is efficacious on tumors that have progressed. It is speculated that the NWBO treatment is less effective on large tumor burdens which is why one of the NWBO trial inclusion criteria requires a total or nearly total resection of the tumor. If this is true, the cross-over effect may be limited; assuming the treatment is efficacious at all. This issue is only applicable to the OS analysis.

The reader should carefully consider these and other limitations while evaluating this model.

There were several control arms used to model the NWBO control arm including those from Genentech (Avastin), ImmunoCellular (ICT-107), and Novocure (Optune). These particular clinical trials were chosen, in part, because the Kaplan-Meier survival curves were publicly accessible.

In a nut-shell, the model enrolls patients into a simulated trial at the same rate specified by the NWBO enrollment curve. The enrollment curve is very important information, because it affects when the target number of events are reached. The model uses the control arm data from a previous clinical trial to calculate the survival time of each patient. A “treatment effect” (i.e. a multiplier) is applied to each treatment arm patient’s survival time.  The month each patient events is calculated and stored, and the events are tallied as the model processes the data on a month-by-month basis for each simulated trial. The treatment effect is adjusted after each simulation until the number of events produced by the model matches the 248 PFS events on October 2016 or the 233 OS events on July 2017.

The model was developed in Excel using a macro to perform the grueling number crunching tasks. Before the model can be ran, the Kaplan-Meier curve control arm data from a previous clinical trial must be manually extracted and charted in Excel. A polynomial formula is then easily generated by using Excel’s built-in “Trendline” feature which generates a polynomial equation to approximate the actual clinical trial KM survival curve.  The equation is used to calculate a “Survival Time” using the “Proportion” on the Kaplan Meier curve as the input. The survival equation for each historical clinical trial is programmed into the macro as a selectable model control arm.

The model is initially tested for accuracy by assuming there is no treatment effect. The model results are compared to the published control arm median survival values to ensure the values match. For example, if the clinical trial indicates that the control arm mPFS is 10 months, then the mPFS generated by the model should be approximately 10 months as well. This accuracy check is also a confirmation that the model design is not flawed in some obvious way.

The results of the accuracy test are listed below in Table 1.0 and 1.1, and the median survival times are high-lighted.  Any treatment benefit shown in the tables is due to sampling error (i.e. chance).

Table 1.0 - Control Arm Accuracy Check for Overall Survival

Table 1.1 - Control Arm Accuracy Check for Progression Free Survival

Notice that the Hazard Ratios are all nearly 1.0 which is expected since a treatment effect was not applied for the accuracy check. Also keep in mind that the table reflects the analytical results from 100 individual simulated trials.  The median survival data in Tables 1.0 and 1.1 is directly comparable to the published values. The model generated accurate median survival times that closely agreed with the published clinical trial control arm survival times as shown in Table 1.2.

          Table 1.2 - Control Arm Accuracy Check

The model was designed so that either PFS or OS could be analyzed. The type of analysis (PFS or OS) and the model control arm is chosen before running the model which performs the following functions:

  1. Model patients are enrolled into a simulated trial at the same rate as specified by the NWBO enrollment curve.
  2. Each patient is assigned to either the treatment or the control arm on a 2-to-1 basis which mimics the actual NWBO trial.
  3. Each patient is randomly assigned a value between 0 and 1 by the macro which represents the "proportion" on the Kaplan-Meier Curve.
  4. As each patient is enrolled, the “Survival Time” (in months) for the patient is generated by the trendline polynomial equation by using the randomly assigned “proportion” as the input to the equation.  The Survival Time is adjusted to yield a survival time that is relative to surgery by adding the “Control Arm Enrollment Time”. This allows apples-to-apples comparisons between the different control arms. For example, the ICT-107 Kaplan-Meier curve data is relative to the enrollment date, therefore approximately 2.7 months, which represents the ICT-107 enrollment time, is added to the Survival Time which produces a Survival Time that is relative to surgery.
  5. The Survival Time is multiplied by the "Treatment Arm Multiplier" (i.e. the Treatment Effect) if the patient is assigned to the treatment arm. For example, if the Survival Time is calculated to be 6 months, and the multiplier is 1.5, then the treatment Survival Time is calculated to be 9 months.
  6. If the Survival Time for the patient is less than the “Model Enrollment Time” (3 months), then the model patient is randomized again. This has the effect of excluding patients who event before being enrolled into the trial, and can result in the model producing a different median Survival Time than the published clinical trial. (The “Model Enrollment Time” and the “Control Arm Enrollment Time” were both set to zero to produce the data for the accuracy check in Tables 1.0 and 1.1)
  7. A future event is generated as follows: 
     
     Event Month = Enrollment Month + Survival Time – Model Enrollment Time

    For example, if a patient is enrolled in Month 30 and the patient's survival time is 13 months and the model enrollment time is 3 months, then a future event is triggered in month 40 (30+13-3).
  8. If the event month exceeds the designated trial end date (October 2016 or July 2017), then the Survival Time is right censored. (Right censored data is used in all of the “At Risk” survival probability calculations for all survival times that are less than the censored survival time, but is ignored in all calculations thereafter. After being censored, the censored data is neither advantageous nor disadvantageous with regards to calculating the survival probability of survival times that are longer than the censored survival time.)
  9. The macro is performing these functions on a month-by-month basis. It begins by processing the first month enrollees, and proceeds thru all additional months. For each month processed, the macro is also tallying the number of events for the control and treatment arms for that month. When the last month of the trial is reached, the macro captures the total number of events for the control and treatment arms as well as the total events for both arms.
  10. At the end of each simulated trial, the Kaplan-Meier Curve data is calculated. The median survival (mOS or mPFS) is extracted directly from the KM curve data. The log-rank P-value and Hazard Ratio with confidence intervals are calculated, and the results are stored in a table for later analysis.
  11. Another simulated trial is automatically initiated, and the macro is repeated until 100 individual simulated trials have been processed which only takes a few seconds to complete.

The median number of events (captured at the end of each simulated trial) from all 100 trials is calculated via the Excel median function, and the total is compared to the target number events for PFS or OS (248,233). The Treatment Arm Multiplier is manually adjusted, and the macro is executed again. This process is repeated until the total number of events generated by the model consistently agrees with the target number of events. This process only takes a few minutes to determine the correct multiplier. As you can see, the Treatment Arm Multiplier is used to "tune" the model to generate the specified number of events on the specified dates.

The results from all 100 model trials is stored in a spreadsheet by the macro, and an analysis is performed on the results.

Overall Survival ("OS") Analysis

The Overall Survival (OS) analysis from all 100 individual model trials is shown in Table 2.0 for each model control arm tested. The Survival time data in Tables 2.0 and 2.1 is relative to surgery, and consequently, the control arm survival time shown is not directly comparable with the original clinical trial. The model also excludes patients that event before being enrolled which also effects the median survival time.

Table 2.0 – Overall Survival (OS) results from all 100 trials

The model Control Arm mOS range was 18 to 21 months and the model mOS DCVax-L range was 27 to 32 months. The treatment benefit for each model control arm analyzed is highlighted and ranges between 6.4-months to 12.0-months advantage for the model treatment arm.  The median P-values were all 0.01 or less, and the hazard ratios range from 0.70 to 0.48 which are significant reductions in risk of death.

To keep these values in perspective; before Optune was approved, the scientific and medical community was only able to improve the GBM median overall survival by about 3 months over the past 25 years. The Optune clinical trial resulted in a 4.9-month improvement and was recently approved by the FDA.

The University of California at Los Angeles (UCLA) is a premier cancer treatment and research institution, and its survival data is particularly interesting. Although it was not a clinical trial, UCLA has tracked the historical survival data for 587 patients treated at its facilities between 2015 and 2017 with numerous cutting-edge therapies including Immunotherapy, TTF Therapy, Ketruda, and many more. In other words, UCLA is fighting GBM with every tool in its oncology arsenal, and yet this model analysis seems to indicate that the NWBO treatment could potentially produce superior results compared to the very best UCLA offers.  

But before you get too excited, notice that the UCLA median P-value for all 100 trials is 0.01 which is very good; however, the average P-value is 0.08 which indicates that not all of the 100 simulated trials resulted in a statistically significant result which typically requires a P-Value of 0.05 or less. A known treatment effect was applied via the “Treatment Arm Multiplier”, yet some of the trials resulted in a statistically insignificant result. This was caused, in this particular case, by sampling error which occurs when the pool of randomly selected patients is not representative of the larger overall population. This demonstrates, that obtaining statistically significant results becomes increasingly more difficult as the treatment benefit declines. NWBO may struggle with this scenario as well, and furthermore, NWBO must achieve a P-value of 0.02 or less to meet its primary endpoint goal. This may also explain why NWBO may choose to allow the clinical data to mature before concluding the trial.

And before moving on to the PFS analysis, the reader should consider the following thought exercise.  For this simple thought experiment, let’s assume all of the 331 patients in the clinical trial were enrolled on November 1st, 2015 which is the approximate date the very last patient was enrolled in the NWBO clinical trial. This would also require 331 surgeries on August 1st, 2015 due to the 3-month enrollment period. Let’s also assume 98 (331-233) patients were alive on August 1st, 2017, which is certainly not a wild stretch of the imagination considering NWBO’s guidance.  This scenario would result in almost 30% of the 331 patients enrolled in the trial having an overall survival of two-years relative to surgery.   Coincidently, the two-year historical overall survival for GBM is also around 30%.  

In reality, the clinical trial took over 5 years to fully enroll. So how could almost 30% of the group still be alive on August 1st, 2017 if the trial has been enrolling for over 5 years when the two-year overall survival is only 30%, and the last surgery was two-years ago?  

Think about this scenario and this question for just a few minutes

 

Progression Free Survival ("PFS") Analysis

The Progression Free Survival analysis from all 100 individual model trials is shown in Table 2.1 below. Keep in mind that a successful NWBO trial outcome requires a 4-month PFS advantage and a P-value of 0.02 or less.

Table 2.1 – Progression Free Survival ("PFS") results from all 100 trials

The model Control Arm mPFS range was 8 to 12 months, and the model mOS DCVax-L range was 17 to 24 months. The treatment benefit for each model control arm analyzed is highlighted and ranges between 7.3-months to 12.5-months advantage for the model treatment arm.  The P-values were all less than 0.005 and the hazard ratios range from 0.53 to 0.27 which indicates extensive reductions in risk of progression.

As a final thought exercise, let’s take another giant leap of faith, and assume all 331 patients in the clinical trial were enrolled on November 1st, 2015 which would also require 331 surgeries on August 1st, 2015. Let’s also take a tiny leap of faith, and assume 83 (331-248) patients had not progressed as of October 1st, 2016 which represents 25% of the 331 patients enrolled in the trial.  This scenario would result in 25% of the 331 patients enrolled in the trial having a PFS of 14-months relative to surgery.  And coincidently, the 14-month historical PFS for GBM is also around 25%.   

So how could approximately 25% of the group be disease free on October 1st, 2015 if the trial has actually been enrolling for over 5 years when the historical 14-month PFS is only around 25%?  

To match the historical standard in this scenario, would require none of the 331 patients to progress while being enrolled; and simultaneously require the trial blended efficacy to be equivalent to the historical standard of care. Needless to say, the first requirement is non-viable, and your cranial light bulb should be glowing by now as the implications of this and the previous exercise are fully realized.

Conclusion

The model demonstrates that the NWBO Phase 3 clinical trial has the potential to produce positive results. If the control arm performance is anything like historical past performance, then the efficacy could be substantial.  When this analysis is viewed as a whole, it provides valuable insight on how the NWBO clinical trial could actually be performing.

In every scenario, the model indicates that NWBO’s trial is performing exceptionally well when compared to other clinical trials, and a successful outcome could result in a block-buster success for NWBO. However, the actual NWBO clinical trial may struggle to achieve statistical significance if the treatment efficacy is relatively small.  

The critics will rightly point out that using historical data to predict or compare the outcomes of clinical trials is problematic for numerous reasons, which is why virtually all Phase 3 clinical trials are blinded and placebo controlled. However, that does not mean that the analysis should not be performed to gain insight on how the trials are performing. 

A few of these same critics, have clairvoyant powers that have allowed them to predict the outcome of this clinical trial (and others) before the results are even known to the principle investigators or DMC members. While it is certainly possible this Phase 3 trial could fail, the vocal critics do not know this with any degree of certainty, and some of their baseless and inflammatory criticism has made it substantially more difficult for NWBO to raise capital to fund the current trial and its clinical pipeline.

NWBO is financially weak and if the Phase 3 trial fails, investors will likely be wiped out either by more toxic financing or bankruptcy. If the Phase 3 trial succeeds, on the other hand, NWBO might become one of those legendary stocks everyone wished they had bought when it was only 20 cents a share.

References

  1. “DCVax Update on Clinical Programs”, ASCO Conference, June 5, 2017, Home - Northwest Biotherapeutics
  2. “A Randomized Trial of Bevacizumab for Newly Diagnosed Glioblastoma”, The New England Journal of Medicine, Vol 370 No. 8, pgs. 699 – 708
  3. “A randomized, double-blind, placebo-controlled phase 2 trial of dendritic cell (DC) vaccination with ICT-107 in newly diagnosed glioblastoma (GBM) patients, June 1 2014, ASCO 2014 Annual Meeting.
  4. “Overall Survival (5-year survival ITT analysis)”, “Progression-free Survival (5-year survival ITT analysis)”, Optune for Healthcare Professionals: Efficacy
  5. “Kaplan-Meier Survival Estimate: Initial Diagnosis (GBM)”, (Age Range: 18-93 chart) http://www.neurooncology.ucla.edu/performance/glioblastomamultiforme.aspx

Disclosure: I am/we are long NWBO.

Additional disclosure: The reader is cautioned that there are numerous limitations to the model. Additionally, the design of the model could be flawed; the data used to model the trial could be inaccurate or uncorrelated to the actual trial; and the assumptions used in the model could be incorrect.