Machine Learning With Statistical Imputation for Predicting Drug Approvals
We apply machine-learning techniques to predict drug approvals using drug-development and clinical-trial data from 2003 to 2015 involving several thousand drug-indication pairs with over 140 features across 15 disease groups. To deal with missing data, we use imputation methods that allow us to fully exploit the entire dataset, the largest of its kind. We show that our approach outperforms complete-case analysis, which typically yields biased inferences. We achieve predictive measures of 0.78 and 0.81 AUC (“area under the receiver operating characteristic curve,” the estimated probability that a classifier will rank a positive outcome higher than a negative outcome) for predicting transitions from phase 2 to approval and phase 3 to approval, respectively. Using five-year rolling windows, we document an increasing trend in the predictive power of these models, a consequence of improving data quality and quantity. The most important features for predicting success are trial outcomes, trial status, trial accrual rates, duration, prior approval for another indication, and sponsor track records. We provide estimates of the probability of success for all drugs in the current pipeline.
Keywords: drug development, clinical trials, healthcare, machine learning, imputation
Watch Lo discuss this topic in his talk at the Amazon re:MARS conference in June 2019.
While many recent medical breakthroughs such as immuno-therapies, gene therapies, and gene-editing techniques offer new hope for patients, they have also made biomedical innovation riskier, and more complex and expensive. These breakthroughs generate novel therapies for investigation, each of which requires many years of translational research and clinical testing, costing hundreds of millions to billions of dollars, and yet often face a high likelihood of failure (Fernandez, Stein, & Lo, 2012). In fact, drug development productivity—the ratio of the number of new drugs approved to R&D spending each year—has declined steadily over the past 50 years despite scientific and technical progress. This phenomenon, which Scannell, Blanckley, Boldon, and Warrington (2012) termed “Eroom’s Law,” as the reverse of Moore’s Law, suggests that the cost of developing new drugs has doubled approximately every nine years since the 1950s. In the face of multiple uncertainties, the need to evaluate drug candidates better and allocate capital to high-potential opportunities more efficiently has only intensified.
To address these needs, in this article we apply machine-learning techniques to predict the outcomes of randomized clinical trials. Machine learning is an interdisciplinary field focused on tackling pattern recognition problems and building predictive models to make data-driven decisions, which is well-suited for this context. Successful applications of these techniques have already revolutionized a number of industries (e.g., advertising, marketing, finance and insurance, oil and gas exploration) and are poised for even greater impact via autonomous vehicles, facial-recognition authentication, and general-purpose robotics.
Drug developers have already applied machine-learning tools to the discovery process via high-throughput screening of vast libraries of chemical and biological compounds to identify drug targets. However, in managing their portfolios of investigational drugs, biopharma companies typically use unconditional estimates of regulatory approval rates based on historically observed relative frequencies. Machine learning techniques yield conditional estimates of success, conditioned on a host of predictive factors known to affect the likelihood of approval, including drug compound characteristics, clinical trial design, previous trial outcomes, and the sponsor track record. We show that these features contain useful signals about drug development outcomes that will allow us to forecast the outcome of pipeline developments more accurately.
Our methodology and results have several implications for stakeholders in the biomedical ecosystem. More accurate forecasts of the likelihood of success of clinical trials will reduce the uncertainty surrounding drug development, which will increase the amount of capital that investors and drug developers are willing to allocate to this endeavor. By extension, this would lower the cost of capital and increase the efficiency of the allocation process. Specifically, we predict the probability of success of drug candidates in two scenarios: (1) advancing from phase 2 to regulatory approval and (2) from phase 3 to regulatory approval (see Figure 1). Investors and drug developers may use such predictions to evaluate the risks of different investigational drugs at different clinical stages, providing them with much-needed transparency. Greater risk transparency is one source of improved financial efficiency because it facilitates more accurate matching of investor risk preferences with the risks of biomedical investment opportunities.
Machine-learning models can also offer guidance to scientists, clinicians, and biopharma professionals as to which factors are most important in determining clinical-trial success, suggesting ways to improve the drug development process and decelerate or reverse Eroom’s Law.
Policymakers and regulators would also benefit from machine-learning predictions, particularly for drug-indication pairs that are predicted to fail with high likelihood—these cases highlight the most difficult challenges in biomedicine and underscore the need for greater government and philanthropic support.
To the best of our knowledge, our study is the largest of its kind. We construct two datasets, one for each scenario, from two proprietary pharmaceutical pipeline databases, Pharmaprojects and Trialtrove provided by Informa® (Informa, 2016). The phase-2-to-approval dataset includes more than 6,000 unique drugs for 288 indications and over 14,500 phase 2 trials, and the phase-3-to-approval dataset contains more than 1,400 unique drugs for 253 indications and over 4,500 phase 3 trials. These data cover over 15 indication groups. In contrast, most published research on drug approval prediction have very small sample sizes, are concentrated on specific therapeutic areas, and involve only one or a small number of predictive factors: Malik et al. (2014) examined the trial objective responses of 88 anticancer agents in phase 1; Goffin et al. (2005) studied the tumor response rates of 58 cytotoxic agents in 100 phase 1 trials and 46 agents in 499 phase 2 trials; El-Maraghi and Eisenhauer (2008) looked at the objective responses of 19 phase 2 anticancer drugs in 89 single agent trials; Jardim, Groves, Breitfeld, and Kurzrock (2017) examined the response rates of 80 phase 3 oncology drugs to identify factors associated with failures; and DiMasi et al. (2015) analyzed 62 cancer drugs and proposed an approved new drug index (ANDI) algorithm with four factors to predict approval for lead indications in oncology after phase 2 testing (see Supplementary Materials H for a comparison of our analysis to theirs).
Another key difference in our approach is that we deal with missing data using statistical imputation methods. We explore four common approaches to “missingness” and demonstrate their advantages and disadvantages over discarding incomplete cases. With the FDA Amendments Act of 2007, drug and clinical trial data collection has been rapidly expanding, but these data are often sparse, and our dataset is no exception. Related studies (e.g., DiMasi et al., 2015) have typically used only complete-case observations—discarding clinical trials with any missing information—which typically eliminates large portions of the data and may also lead to certain biases.
We use machine-learning techniques to form our predictions, including cross-validation for training and a held-out testing set for performance evaluation, and use the standard “area under the receiver operating characteristic curve” (AUC) metric to measure model performance (AUC is the estimated probability that a classifier will rank a positive outcome higher than a negative outcome [Fawcett, 2006]). We achieve AUCs of 0.78 for predicting phase 2 to approval (95% confidence interval (CI): [0.75, 0.81]) and 0.81 for predicting phase 3 to approval (95% CI: [0.78, 0.83]). A time-series, walk-forward analysis approach shows similar results. We also apply our models to the current drug pipeline—that is, all drugs still in development as of the end of our dataset—to identify the candidates that have the highest and lowest probabilities of success. We examine the latest development statuses of these pipeline drug-indication pairs—a true “out-of-sample” experiment (validation on data not used in model building)—and find that candidates with higher scores are, indeed, more likely to progress to later clinical stages. This indicates that our classifiers do discriminate between high- and low-potential candidates.
2. Materials and methods
The commercial data vendor Informa® offers two databases that are used in our analysis: Pharmaprojects, which specializes in drug information, and Trialtrove, which specializes in clinical trials information (Informa, 2016). These two databases aggregate drug and trial information from over 30,000 data sources in more than 150 countries, including company press releases, government drug databases (e.g., Drugs@FDA) and trial databases (e.g., Clinicaltrials.gov [Zarin, Tse, Williams, & Carr, 2016], Clinicaltrialsregister.eu [extracted from EudraCT]), and scientific conferences and publications. Using these sources, we construct two datasets of drug-indication pairs: phase 2 to approval (P2APP) and phase 3 to approval (P3APP). We extract clinical trial features from Trialtrove, and augment this data using drug features from Pharmaprojects. Applying machine-learning algorithms to these datasets allows us to estimate: (1) whether a drug-indication pair that has concluded phase 2 testing will be approved eventually; and (2) whether a pair that has concluded phase 3 testing will be approved eventually. Data cleaning procedures are outlined in Supplementary Materials A.
We consider all indications associated with a particular drug, as opposed to only the lead indication. We extract all features that could conceivably correlate with the likelihood of success, from drug compound attributes (31 features from Pharmaprojects profiles) to clinical trial characteristics (113 features from Trialtrove). These features are defined in Table 1 and Supplementary Materials A. In general, each dataset may be partitioned into two disjoint subsets: one with samples that have known outcomes, and another with samples that are still in the pipeline at the time of snapshot of the databases (that is, the outcomes are unknown). To provide intuition for the characteristics of the samples, we describe key summary statistics of each subset.
The P2APP dataset consists of 6,344 drug-indication pairs that have ended phase 2 testing; that is, there are no phase 2 trials in progress or planned in the database. The phase 2 trials in this dataset range from August 8, 1990 to December 15, 2015. In our sample, 4,812 pairs have known outcomes, while 1,532 pairs are still in the pipeline. In the subset with known outcomes, we define the development statuses of suspension, termination, and lack of development as “failures” (86.8%), and registration and launch as “successes” or approvals (13.2%). The P3APP dataset consists of 1,870 pairs that have ended phase 3 testing, of which 1,610 pairs have known outcomes, while 260 pairs are still in the pipeline. For those pairs with known outcomes, we define “failures” (59.1%) and “successes” (40.9%) in the same fashion as the P2APP dataset. The phase 3 trials in P3APP span from January 1, 1988 to November 1, 2015. These figures are summarized in Table 2. Here, the use of terms “success” and “failure” is in the context of achieving approval. We note that our definition of “failures” can include drug development programs that are terminated due to factors unrelated to the performance of the drug (e.g., market conditions, business decisions). In Section 3 Results, we find that this outcome variable has significant associations with trial performance and other factors.
The datasets cover 15 indication groups: alimentary, anti-infective, anti-parasitic, blood and clotting, cardiovascular, dermatological, genitourinary, hormonal, immunological, musculoskeletal, neurological, anti-cancer, rare diseases, respiratory, and sensory products. Anti-cancer agents make up the largest subgroup in P2APP, and the second largest in P3APP (see Table 3). Industry-sponsored trials dominate both datasets (see Table 4). In aggregate, we observe a decreasing trend in success rates over five-year rolling windows from 2003 to 2015 (see Figure 2).
Table 1. Description of parent features extracted from Pharmaprojects and Trialtrove. Some parent features are multi-label (e.g., a trial may be tagged with United States and United Kingdom simultaneously). We transform all multi-label parent features into binary child features (1 or 0). See Supplementary Materials A for specific examples of each feature. Note that drug-indication pairs for the same drug have the same drug features; drug-indication pairs involved in the same trial have the same trial features.
Table 2. Sample sizes of P2APP and P3APP datasets. We consider phase 2 trial information in the P2APP dataset and phase 3 trial information in the P3APP dataset.
Table 3. Breakdown of drug-indication pairs by indication groups. A drug-indication pair may have multiple indication group tags. For instance, renal cancer is tagged as both anti-cancer and rare disease in Pharmaprojects.
Table 4. Breakdown of trials by sponsor types. A trial may be sponsored by more than one party (e.g., collaboration between industry developers and academia).
2.2. Missing data
Prior to the 2007 FDA Amendments Act (FDAAA), it was not uncommon for investigators to release only partial information about pipeline drugs and clinical trials to protect trade secrets or simply because there was no incentive to do more. Even today, some investigators still do not adhere to the FDAAA-mandated registration policy or submit adequate registrations. Therefore, all historical drug development databases have missing data. We note that the “missingness” here is largely related to the post-study reporting of clinical trial data as opposed to in-trial data missingness (e.g., censorship of panel data due to patients terminating trial participation prematurely). In the former case, the data (e.g., trial duration, trial outcomes) are usually available to the investigators but may not be released publicly, and are thus considered “missing” from our standpoint. Therefore, our dataset may be considered an approximation of the actual data available to the FDA and individual drug developers.
Figure 3, Figure 4, Table 5, and Table 6 summarize the patterns of missingness in our dataset (we exclude pipeline drug-indication pairs here because their outcomes are still pending). The missing data patterns are multivariate. When conditioned on the latest level of development, for any indication, we find that successful drugs generally have lower levels of missingness compared to failed drugs. For instance, in the P2APP dataset, 61% of failed drugs have an unknown medium, while only 15% of approved drugs are missing this feature. We also observe that completed trials tend to have greater levels of missingness than terminated trials. Between two datasets, we find that the P3APP dataset, which focuses on phase 3 drugs and trials, generally has less missing data for both drug and trial features than the P2APP dataset which focuses on phase 2 drugs and trials. This is expected since phase 3 trials are primarily used to support registration filings.
Table 5. Missingness in drug features with respect to unique drugs (see Figure 3). The column heading “Unconditional” refers to overall missingness without conditioning on outcome.
Table 6. Missingness in trial features with respect to unique trials (see Figure 4). The column heading “Unconditional” refers to overall missingness without conditioning on trial status.
Most related studies do not report the extent of missing data in their samples, presumably because smaller datasets were used. DiMasi et al. (2015) reported missing data for some of their factors, and addressed it through listwise deletion—deleting all observations with any missing factors. Since statistical estimators often require complete data, this approach is the simplest remedy for missingness. However, it greatly reduces the amount of data available and decreases the statistical power of the resulting statistics. Furthermore, listwise deletion is valid only under strict and unrealistic assumptions (see below), and when such conditions are violated, inferences are biased. In the current study, we make an effort to include in our analysis all observed examples, with or without complete features, through the use of statistical imputation.
Missing data may be classified into three categories (Rubin, 1976): missing completely at random (MCAR), missing at random (MAR), and missing not-at-random (MNAR). MCAR refers to data that are missing for reasons entirely independent of the data; MAR applies when the missingness can be fully accounted for by the observed variables; and MNAR refers to situations when neither MCAR nor MAR is appropriate, in which case the probability of missingness is dependent on the value of an unobserved variable (Van Buuren, 2012). See Supplementary Materials B for the precise definitions of each type of missingness.
If the missingness is MCAR, the observed samples can be viewed as a random subsample of the dataset. Consequently, using listwise deletion should not introduce any bias. While convenient, this assumption is rarely satisfied in practice. In most drug-development databases, failed drugs are more likely to have missing features than successful drugs (see Table 5). Clearly, MCAR does not hold.
Applying listwise deletion when the missingness is not MCAR can lead to severely biased estimates. Moreover, given the nature of drug-development reporting, a large portion of the original data may be discarded if many variables have missing values. For these reasons, the listwise-deletion approach adopted by DiMasi et al. (2015) and others is less than ideal.
Given only the observed data, it is impossible to test for MAR versus MNAR (Enders, 2010). However, our knowledge of the data-collection process suggests that MAR is a plausible starting point, and we hypothesize that the missingness in drug and trial features is mainly accounted for by drug development and trial statuses, respectively. Our observations in Table 5 and Table 6 support this approach, as the missingness proportions for some features differ greatly depending on the outcome.
Our assumption of MAR is consistent with the data-collection methodology in the Informa® databases. Drug profiles are built up over time in Pharmaprojects. As a drug advances to later phases, information about its characteristics becomes more readily available because investigators release more data about pipeline drugs after each phase of clinical testing. Informa® inputs this information into its databases as they become available in the public domain or through primary research. Approved drugs are more likely to have more complete profiles, while information about failed drugs tends to stay stagnant because no further studies are conducted. It is very plausible that the MAR nature of our datasets is an artifact of data collection, and by extension, so are similar pharmaceutical datasets extracted from the public domain and maintained in the same fashion. Originally intended to track drug and trial activities, Pharmaprojects and Trialtrove are not structured to keep track of information updates over time since there was no use for it. Without timestamps of the updates, we are not able to eliminate the MAR artifact from our datasets.
In our analysis, we impute the missing data under the more plausible MAR assumption to obtain complete datasets. In contrast to listwise deletion, we fill in missing values using information in the observed variables. This allows us to utilize data that would otherwise be discarded. Thereafter, we can apply all the usual statistical estimators to this imputation-completed data.
Our analysis consists of two parts that we perform in R version 3.2.3. First, we impute missing values to generate complete datasets. Next, we apply a range of machine-learning algorithms to build predictive models based on the imputed data. Illustration of the specific components of our analysis appear in Figure 5.
We formulate our two scenarios as supervised bipartite ranking problems, where the goal is to predict the outcome—success or failure—of a drug-indication pair given a set of input features. Initially, we split each dataset into training and testing sets. For each scenario, we train various classifiers based on the corresponding training set, and compute the expected error of our predictive models by testing them on the held-out testing set.
We create feature matrices from the datasets by representing drug and trial features for each drug-indication pair as vectors (see Figure 6). Drug-indication pairs associated with multiple trials are represented by the same number of feature vectors, e.g., a pair with two trials has two rows. We give a concrete example in Figure 6. Consider the drug-indication pair Analiptin-diabetes type 2 in the P2APP dataset. We represent it using two vector rows since it has two phase 2 trials in Trialtrove. Note that the feature matrix is incomplete due to missing drug and trial features. We also construct a column vector of labels, which contains the outcomes of the drug-indication pairs. Labels are not available for pipeline drug-indication pairs because they are still in development and their outcomes are still uncertain, hence these observations are not used to train our classifiers. However, with the trained classifiers, we can generate predictions for pipeline data.
We split each dataset (excluding pipeline drugs-indication pairs) into two disjoint sets, one training set and one testing set, and form feature matrices for both according to the drug-indication pairs in each set. The testing sets serve as out-of-sample datasets to evaluate our models. Therefore, we mask their outcomes (that is, we treat them as unknown) and access them only at the very end to check our performance.
To deal with missing data in both training and testing sets, we consider listwise deletion and four statistical imputation techniques commonly used in social science research and biostatistics: unconditional mean imputation, k-nearest neighbor (kNN) imputation, multiple imputation (MI), and decision-tree algorithms (see Supplementary Materials C for details). We follow best practices of the missing-data literature by including as many relevant auxiliary variables as possible, as well as all variables used in subsequent models (Collins, Schafer, & Kam, 2001; Enders, 2010; Rubin, 1996; Schafer & Graham, 2002). This makes the assumption of MAR more plausible in our datasets, and helps to reduce bias in subsequent analyses (Schafer, 1997). In particular, it is necessary to include our target variable—the drug-indication development status—in our imputation model because we hypothesize that missingness is mainly accounted for by it. This is not an issue for the training sets. However, the outcomes in the testing sets are masked, and not supposed to be known. Therefore, we treat the testing set outcomes as though they were missing and impute them together with all the other missing features. After imputation, we discard the imputed testing-set outcomes, and use only the imputed feature values for predictions. We do the same when evaluating pipeline datasets.
With respect to the machine-learning algorithm, we explore several linear and non-linear classifiers commonly used in this literature: penalized logistic regression (PLR), random forests (RF), neural networks (NN), gradient boosting trees (GBT), support vector machines with radial basis functions (SVM), and decision trees C5.0. We implement the first five algorithms using the scikit-learn package in Python (Pedregosa et al., 2011) and the sixth using the C50 package in R (Kuhn, Weston, Coulter, & Quinlan, 2014). For training, we weight each feature matrix row example according to the number of trials of the corresponding drug-indication pair. In our earlier example, the drug-indication pair Analiptin-diabetes type 2 was involved in two phase 2 trials. It is represented by two vector rows in the feature matrix (see Figure 6). Both rows are used as training examples, and each is weighted equally during training (0.5, since there are two trials in total). To obtain predictions for a drug-indication pair, we average the output probabilities and scores of the corresponding feature vector rows that are used as inputs to the classifier.
All machine-learning algorithms have hyper-parameters that affect the flexibility of the model and must be tuned to each dataset to optimize goodness of fit. Poorly-chosen hyper-parameters can lead to overfitting (attributing signal to noise) or underfitting (attributing noise to signal). We tune our parameters using k-fold cross-validation (with k = 5 or 10, depending on the sample size). Since the cross-validation process should emulate the testing process as closely as possible, we include imputation in the cross-validation loop as well. We split the training set into validation and non-validation folds. Then we treat validation fold outcomes as missing, and impute them as we would for a testing set. From here, we ignore the imputed validation fold outcomes and proceed with the standard validation process.
In the final step, we test the trained classifiers on the unseen testing sets for out-of-sample model validation. This gives the expected performance of our predictive models for each of the scenarios, using the standard AUC metric to measure model performance.
3.1. Simulation of listwise deletion versus imputation
We study the effects of imputation using a “gold-standard” dataset derived from the complete cases of the P2APP dataset (see Table 7). To simulate the missingness present in the original dataset, we introduce missingness in the gold-standard dataset based on our MAR assumption and the missingness patterns observed in the P2APP dataset. We randomly split the drug-indication pairs into a training set (70%) and a testing set (30%), and use five different missing data approaches, as described in Supplementary Materials C, to generate complete training sets from the MAR training set. We use each imputed training set to build six different predictive models (PLR, RF, NN, GBT, SVM, and C5.0) according to the methodology outlined in Section 2. We repeat this experiment 100 times for robustness. Table 8 summarizes the AUC performance of the classifiers on the gold-standard testing sets. See Supplementary Materials E for a more detailed description and results.
Table 7. Sample size of the gold-standard dataset (derived from complete cases of P2APP).
Table 8. AUC of different classifiers under different missing data approaches. Abbreviations: Avg: average; Sd: standard deviation; 5%: 5th percentile; 50%: median; 95%: 95th percentile; m: number of imputations generated.
For all six machine-learning algorithms, we find that gold-standard classifiers—that is, the models derived from complete data—consistently outperform their complete-case analysis and imputation counterparts. This is logical because useful information is invariably lost when we introduce missingness in the datasets. In contrast, complete-case analysis often leads to inferior performance. The AUCs of classifiers trained on complete-cases training sets tend to be smaller than those trained on imputed training sets. This suggests that imputation does indeed offer improved fit and predictive power over listwise deletion.
Overall, we find kNN imputation to be most compatible with our datasets. It provides the least biased imputations among all missing data methods (see Supplementary Materials E). In particular, the combination of kNN imputation (k = 5) with RF gives one of the highest gold-standard testing set AUCs (0.81). We note a few other MI combinations that yield comparable or marginally better performance but focus on the 5NN-RF approach in subsequent analyses on the main datasets due to its ease of implementation and application. We find that SVM has the worst performance among all machine-learning models. This is not surprising because SVMs are aimed only at learning binary classifiers, and do not generally produce good class probability estimates. Consequently, such models do not necessarily give high AUCs.
We also compare our approach with the ANDI algorithm (DiMasi et al., 2015) by applying a modified version of the index on oncology drugs in the gold-standard testing sets (see Supplementary Materials H for a more in-depth description). We find that our 5NN-RF model achieves significantly higher AUC than the modified ANDI, with an average improvement of 0.1 in AUC over 100 simulations (see Figure 7). We believe that this gain can be attributed to a larger training set with a wider range of features, a nonlinear model that can capture the complex relationships in the data, and a proper model validation methodology.
3.2. Predicting approvals
We analyze the two datasets (P2APP and P3APP) by first splitting each into a training set (70%) and a testing set (30%) randomly (pipeline drug-indication pairs are omitted since their outcomes have yet to be determined). Subsequently, we train 5NN-RF models for each scenario according to the methodology outlined in Section 2. We repeat this experiment 100 times for robustness. Table 9 summarizes the AUC performance metrics for the testing sets. On average, we achieve 0.78 AUC for P2APP and 0.81 AUC for P3APP.
Table 9. Comparison of the general and indication-group specific classifiers for selected indication groups. Abbreviations: Avg: average; Sd: standard deviation; 5%: 5th percentile; 50%: median; 95%: 95th percentile.
The observed performance is essentially the MAR testing set AUC, since backfilling has already affected the datasets used. In Supplementary Materials E, we highlight the perils of relying on the MAR testing set for model validation, and suggest that the AUCs for the gold-standard and MCAR testing sets are more reflective of a classifier’s real performance. Unfortunately, we have access to neither the gold-standard nor the MCAR testing sets, because we do not know the true, underlying values of the missing features. However, our experiments indicate that the AUCs for the MAR and MCAR testing sets of the 5NN-RF combination are very close (a difference of 0.002 on average). This means that we may use the former, the only observed figure, as a reasonable estimate of the latter, which reflects real performance.
Next, we train classifiers based on the union of the training and testing sets, and use them to generate predictions for pipeline drug-indication pairs. We generate predictions for P2APP using only information from phase 2 trials and for P3APP using only information from phase 3 trials. While we cannot compute AUC scores for these samples because their outcomes are still pending, we can compare their prediction scores with their development status at the time of this writing. These pipeline drug-indication pairs may still be in the same clinical stage (no change, i.e., phase 2 for P2APP; phase 3 for P3APP), be terminated (failed), or have progressed to higher phases (advanced).
Figure 8, Table 10, and Table 11 summarize the distributions of pipeline prediction scores. We find that pairs that fail generally have lower scores than those that advance to later phases of development. In Figure 8, we observe peaks at the lower end of the score spectrum for failed pairs (red) for both datasets. In contrast, pairs that advance tend to have peaks at higher scores (green). We observe the same patterns when we disaggregate the distributions by indication groups: the green parts tend to cluster above the distribution median while the red parts cluster below. However, there are also some indication groups for which there are too few samples to make any useful remarks (e.g., hormonal products in P2APP). From Table 10, we see that the average scores of failed pairs are indeed lower than those that advance (differences ranging from 0.05 to 0.15). In Table 11, we bin drug-indication pairs that have new developments (whether failure or advancement) into four groupings, depending on their prediction scores. For each bin, we compute the proportion of samples that advance to later development stages. We find that the proportions generally increase with the score magnitude, suggesting that pairs with higher scores are more likely to advance than those with lower scores. We note that progress to later clinical stages does not always lead to approval. However, the results are still promising because advancement is a necessary condition for approval. Our experiments indicate that our trained classifiers are able to discriminate between high- and low-potential candidates.
Table 10. Distributions of prediction scores for all indication groups in aggregate (see Figure 8). Advanced refers to progress to a higher phase from the original phase. Original phase for P2APP is phase 2; for P3APP is phase 3. For instance, out of 1,511 drug-indication pairs in the P2APP testing set, 859 pairs are still pending decision in phase 2, 244 pairs have failed, and 408 pairs have successfully advanced to phase 3 testing. Abbreviations: n: sample size; Avg: average; Sd: standard deviation; 5%: 5th percentile; 50%: median; 95%: 95th percentile.
Table 11. Distributions of prediction scores for all indication groups in aggregate (see Figure 8). Proportion refers to the fraction of samples that advanced to a later phase from the original phase. Abbreviations: n: sample size.
To gain insight into the logic of our trained predictive models, we compute the average importance of features used in the 5NN-RF classifiers over all the experiments, and extract the top ten most informative variables. The RF classifier (Pedregosa et al., 2011) we used computes the importance of a variable by finding the decrease in node impurity for all nodes that split on that variable, weighted by the probability of reaching that node (as estimated by the proportion of samples reaching that node), averaged over all trees in the forest ensemble (Breiman, Friedman, Stone, & Olshen, 1984). Table 12 summarizes the results.
Table 12. Top ten important variables of 5NN-RF classifiers for P2APP and P3APP. Average and standard deviation taken across all experiments. Abbreviations: Avg: average; Sd: standard deviation.
We find that trial outcome (whether the trial was completed with its primary endpoints met) and trial status (whether the trial was completed or terminated) have significant associations with success. These two features were consistently ranked the top two out of all variables and across both datasets. It is easy to imagine that a drug-indication pair whose trials were terminated has a low probability of success in terms of advancing from phases 2/3 to approval. In contrast, candidates that achieve positive outcomes certainly have a better shot at success. We also observe that prior approval of a drug has an effect on success for new indications or patient segmentation. It is plausible that developing an approved drug for a new indication has a greater likelihood of success than a new candidate.
In addition, trial characteristics such as accrual, duration, and the number of identified sites frequently appear in the top ten important variables. There are several possible explanations. For example, trials that end quickly without achieving primary endpoints may undermine the likelihood of success, and drugs with trials that have small accrual—and thus low statistical power—may have a lower probability of being approved.
We also find sponsor track records—quantified by the number of past successful trials (trials that achieve positive results or meet primary endpoints)—to be a useful factor for prediction. This factor has not been considered in previous related studies, but the intuition for its predictive power is clear: strong track records are likely associated with greater expertise in drug development.
Since drugs developed for different indication groups may have very different characteristics, we might expect classifiers trained on indication-group-specific data to outperform general classifiers. We build and analyze such specialized classifiers by filtering the datasets by indication group before performing the experiment described in the previous section. As a comparison, we also break down the performance of the general classifiers by indication group. Table 9 shows the results for selected indication groups. In general, we find specialized models to give poorer performance than general models. This is likely because the former are trained on less data, which makes them less accurate and more susceptible to overfitting.
We note that the approach adopted in this section—splitting drug-indication pairs into training and testing sets randomly without considering the dates of development—may be less than ideal because of look-ahead bias. For example, if the results of a 2008 trial are included in the training set for predicting the outcome of a 2004 development path for a drug-indication pair, our model will be using future information during validation, which can yield misleading and impractical inferences. To address this issue, in the next section we apply our machine-learning framework to time-series data using rolling windows that account for temporal ordering in the construction of training and testing sets. Although this process makes use of less data within each estimation window than when the entire dataset is used, it minimizes the impact of look-ahead bias and yields more realistic inferences. We study the effects of random splitting versus temporal ordering in Supplementary Materials J.
3.3. Predictions over time
Drug development has changed substantially over time, thanks to new scientific discoveries and technological improvements. To reflect these changes in our predictive analytics, we adopt a time series, walk-forward approach to create training and testing sets for each of the two datasets, P2APP and P3APP (see Figure 9). We sample five-year rolling windows between 2004 and 2014 from each dataset. Each window consists of a training set of drug-indication pairs whose outcomes become finalized within the window, and an out-of-sample, out-of-time testing set of drug-indication pairs that ended phase 2 or phase 3 testing, but are still in the pipeline with undetermined outcomes within the window. For example, consider the P2APP dataset. We draw the first window from 2004–2008, train our algorithm on drug-indication pairs that failed or were approved within this period as the training set, and apply the trained model to predict the outcomes of drug-indications that just ended phase 2 testing within the same window as the testing set.
We evaluate the resulting classifier by comparing its predictions with outcomes that are realized in the future (2009–2015). This rolling-window approach yields a total of eight overlapping training and testing periods where a new 5NN-RF model is trained for each period. The eighth testing period consists of drug-indication pairs in the pipeline at the time of snapshot of the databases. Unlike the first seven periods, their outcomes are still pending current development, and therefore we cannot compute a testing AUC for this window. However, we can examine the predictions and compare the scores with their development statuses at the time of this writing.
Figure 10 summarizes the results of the time-series analysis for the first seven windows. We observe an increasing trend over the years for both P2APP (0.67 in the first and 0.80 in the last window) and P3APP (0.77 in the first and 0.88 in the last window). Interestingly, we note that the proportions of complete cases in the training sets correlate well with the time series AUC (correlation coefficient 0.95 for P2APP and 0.90 for P3APP). We compute the proportion of complete cases by taking the number of feature vector rows with complete information over the total number of rows. As is apparent from Figure 10, the proportions have been increasing over the years for both datasets. This is likely due to better data reporting practices by drug developers, a possible consequence of FDAAA.
Next, we examine the 2011–2015 window. Figure 11, Table 13, and Table 14 summarize the distributions of prediction scores for the P2APP and P3APP datasets. We observe very similar patterns to the static pipeline predictions above. The histograms, average scores, and binning of samples indicate that pairs that fail tend to have lower prediction scores than those that advance. This shows that our classifiers are indeed able to differentiate successful candidates.
Table 13. Distributions of prediction scores for all indication groups in aggregate (see Figure 11). Advanced refers to progress to a higher phase from the original phase. Original phase for P2APP is phase 2; for P3APP is phase 3. Abbreviations: n: sample size; Avg: average; Sd: standard deviation; 5%: 5th percentile; 50%: median; 95%: 95th percentile.
Table 14. Distribution of prediction scores for all indication groups in aggregate (see Figure 11). Proportion refers to the fraction of samples that advanced to a higher phase from the original phase. Abbreviations: n: sample size.
Table 15 summarizes the top ten most informative variables in the 5NN-RF classifiers over the eight rolling windows. We find them to be largely consistent with those observed in the static case: the trial outcome and trial status are significantly associated with success; trial characteristics (such as accrual, duration, and number of identified sites), sponsor track record, and drug medium appear frequently in both scenarios.
Table 15. Top ten important variables in 5NN-RF classifiers for P2APP and P3APP. Average and standard deviation taken across the eight rolling windows. Abbreviations: Avg: average; Sd: standard deviation.
As in the static case, we also train indication-group specific classifiers using rolling windows. Table 16 and Table 17 summarize the results for selected indication groups in P2APP and P3APP, respectively (see Supplementary Materials G for results of all other indication groups). Indication groups with small sample sizes tend to produce poor and unstable specialized classifiers (e.g., the musculoskeletal indication group in P2APP). This is expected because models trained on small training sets are more susceptible to overfitting, especially when non-linear algorithms such as RF are used. In contrast, indication groups with larger sample sizes tend to give rise to rather good classifiers (e.g., anti-cancer in P2APP).
Table 16. Comparison of the general and indication-group specific classifiers for selected indication groups in P2APP. We use bootstrapping to determine the 95% CI for AUC.
Table 17. Comparison of the general and indication-group specific classifiers for selected indication groups in P3APP. We use bootstrapping to determine the 95% CI for AUC.
For comparison, we disaggregate performance by indication group. We find that these classifiers do not lose out to their specialized counterparts. In fact, our results show that the former tend to exhibit more stable performance across the seven windows, particularly on indication groups with small sample sizes. We hypothesize that classifiers trained on all data benefit from having access to larger datasets with greater diversity, and are thus able to make more informed predictions. This suggests that it may be more appropriate to rely on general classifiers, rather than specialized ones, for predictions over time where samples are spread out over multiple windows, since further filtering by indication group results in even smaller sample sizes.
Finally, we extract the top five P2APP pipeline drug candidates with the highest scores in each indication group as predicted by the 2011–2015 rolling-window model. Table 18 summarizes the results. We include only candidates that are still outstanding at the time of writing (neither discontinued nor approved). It is encouraging that many of these candidates (indicated in italics) have advanced beyond phase 2 testing since our analysis, indicating the predictive power of our models. We include an interactive version in Figure 12 where readers can filter our pipeline predictions by indication group and probability of approval. Ultimately, all biopharma stakeholders can use such scores to rank and evaluate the potential risks and rewards of drug candidates.
3.4. Interactive Data Visualization: Network graph of P2APP pipeline drug candidates
The information presented through this research and in the included figure are made available solely for general informational purposes. The authors do not warrant the accuracy, completeness or usefulness of this information. Any reliance you place on such information is strictly at your own risk. The authors expressly disclaim all liability and responsibility arising from any reliance placed on such information by you, or by anyone who may be informed of this information.
Table 18. Top five P2APP pipeline drug candidates with the highest scores in each indication group as predicted by our model. We include only candidates that are still outstanding at the time of writing (neither discontinued nor approved). Drug-indication pairs in italics are those that have advanced beyond phase 2 testing since our analysis.
Drug development is an extremely costly process and the accurate evaluation of a candidate drug’s likelihood of approval is critical to the efficient allocation of capital. Historical successes and failures contain valuable insights on the characteristics of high-potential candidates. Unfortunately, such data are often incomplete due to partial reporting by investigators and developers. Most analytic methods require complete data, however, and prior studies on estimating approval rates and predicting approvals are typically based on a small number of examples that have complete information for just a few features.
In this paper, we extract two datasets, P2APP and P3APP, from Informa® databases and apply 5NN statistical imputation to make efficient use of all available data. We use machine-learning techniques to train and validate our RF predictive models and achieve promising levels of predictive power for both datasets. When applied to pipeline drugs, we find that candidates with higher scores are indeed more likely to advance to higher clinical phases, indicating that our 5NN-RF classifiers are able to discriminate between high- and low-potential candidates.
A time-series analysis of the datasets shows generally increasing trends in performance over five-year rolling windows from 2004 to 2014. We find that the classifiers’ performance correlates well with the proportions of complete cases in the training sets: as completeness increases, the classifier learns better and achieves higher AUCs. This highlights the importance of data quality in building more accurate predictive algorithms for drug development.
Finally, we compute feature importance in the predictive models and find that trial outcomes, trial status, trial accrual rate, duration, prior approval for another indication, and sponsor track record are the most critical features for predicting success. Because the 5NN-RF classifiers are non-linear, there is no simple interpretation of the incremental contribution of each predictor to the forecast. However, the intuition behind some of these factors is clear: drug-indication pairs with trials that achieve positive outcomes certainly have a better chance of approval; candidates sponsored by companies with strong track records and greater expertise in drug development should have higher likelihood of success; and approved drugs may have higher chances of approval for a second related indication. Many of these factors contain useful signals about drug development outcomes but have not been considered in prior studies.
These results are promising and raise the possibility of even more powerful drug development prediction models with access to better quality data. This can be driven by programs such as Project Data Sphere (Green et al., 2015) and Vivli (Bierer, Li, Barnes, & Sim, 2016) that promote and facilitate public sharing of patient-level clinical trial data. Ultimately, such predictive analytics can be used to make more informed data-driven decisions in the risk assessment and portfolio management of investigational drugs at all clinical stages.
The information presented through this research and in the included figure are made available solely for general informational purposes. The authors do not warrant the accuracy, completeness or usefulness of this information. Any reliance you place on such information is strictly at your own risk. The authors expressly disclaim all liability and responsibility arising from any reliance placed on such information by you, or by anyone who may be informed of this information.
We thank Informa® for providing us access to their data and expertise and are particularly grateful to Christine Blazynski, Mark Gordon, and Michael Hay for many helpful comments and discussion throughout this project. We also thank them and Anna Barker, Linda Blackerby, Lara Boro, Steve Finch, Howard Fingert, Xiao-Li Meng, Ellen Moore, Jim Reddoch, Lara Sullivan, Marty Tenenbaum, James Wade, and three anonymous referees for specific comments on this manuscript, and Jayna Cummings and Paige Sammartino for editorial assistance. Research support from the MIT Laboratory for Financial Engineering is gratefully acknowledged.
The views and opinions expressed in this article are those of the authors only, and do not necessarily represent the views and opinions of any institution or agency, any of their affiliates or employees, or any of the individuals acknowledged above.
Bierer, B. E., Li, R., Barnes, M., & Sim, I. (2016). A global, neutral platform for sharing trial data. New England Journal of Medicine, 374(25), 2411–2413. https://doi.org/10.1056/nejmp1605348
Breiman, L., Friedman, J., Stone, C. J., & Olshen, R. A. (1984). Classification and regression trees. CRC Press.
Collins, L. M., Schafer, J. L., & Kam, C. M. (2001). A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychological Methods, 6(4), 330–351. https://doi.org/10.1037//1082-989x.6.4.330
DiMasi, J. A., Hermann, J. C., Twyman, K., Kondru, R. K., Stergiopoulos, S., Getz, K. A., & Rackoff, W. (2015). A tool for predicting regulatory approval after phase II testing of new oncology compounds. Clinical Pharmacology & Therapeutics, 98(5), 506–513. https://doi.org/10.1002/cpt.194
El-Maraghi, R. H., & Eisenhauer, E. A. (2008). Review of phase II trial designs used in studies of molecular targeted agents: outcomes and predictors of success in phase III. Journal of Clinical Oncology, 26(8), 1346–1354. https://doi.org/10.1200/jco.2007.13.5913
Enders, C. K. (2010). Applied missing data analysis. Guilford Press.
Fernandez, J. M., Stein, R. M., & Lo, A. W. (2012). Commercializing biomedical research through securitization techniques. Nature Biotechnology, 30(10), 964–975. https://doi.org/10.1038/nbt.2374
Goffin, J., Baral, S., Tu, D., Nomikos, D., & Seymour, L. (2005). Objective responses in patients with malignant melanoma or renal cell cancer in early clinical studies do not predict regulatory approval. Clinical Cancer Research, 11(16), 5928–5934. https://doi.org/10.1158/1078-0432.ccr-05-0130
Green, A. K., Reeder-Hayes, K. E., Corty, R. W., Basch, E., Milowsky, M. I., Dusetzina, S. B., Bennett, A.V., & Wood, W. A. (2015). The project data sphere initiative: accelerating cancer research by sharing data. The Oncologist, 20(5), 464-e20. https://doi.org/10.1634/theoncologist.2014-0431
Jardim, D. L., Groves, E. S., Breitfeld, P. P., & Kurzrock, R. (2017). Factors associated with failure of oncology drugs in late-stage clinical development: A systematic review. Cancer Treatment Reviews, 52, 12–21. https://doi.org/10.1016/j.ctrv.2016.10.009
Kuhn, M., Weston, S., Coulter, N., & Quinlan, R. (2014). C50: C5. 0 decision trees and rule-based models. R package version 0.1. 0-21.
Malik, L., Mejia, A., Parsons, H., Ehler, B., Mahalingam, D., Brenner, A., Sarantopoulos J., & Weitman, S. (2014). Predicting success in regulatory approval from Phase I results. Cancer Chemotherapy and Pharmacology, 74(5), 1099–1103. https://doi.org/10.1007/s00280-014-2596-4
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel M., Prettenhofer P., Weiss R., Dubourg V., & Vanderplas, J. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. http://www.jmlr.org/papers/v12/pedregosa11a
Van Buuren, S. (2012). Flexible imputation of missing data. CRC Press.
Zarin, D. A., Tse, T., Williams, R. J., & Carr, S. (2016). Trial reporting in clinicaltrials.gov—the final rule. New England Journal of Medicine, 375(20), 1998–2004. https://doi.org/10.1056/nejmsr1611785
I am glad to see this brilliant post. All the details are very helpful and good for us, keep up to good work. I found some useful information in your forum, it was awesome to read, thanks for sharing this great content to my vision, keep sharing. We apply machine-learning techniques to predict drug approvals using drug development and clinical-trial data from 2003 to 2015 involving several thousand drug-indication pairs with over 140 features across 15 disease groups. To deal with missing data, we use imputation methods that allow us to fully exploit the entire dataset, the largest of its kind. We show that our approach outperforms complete-case analysis, which typically yields biased inferences. Click to know more: Machine Learning