Validation of an FMRI Based Classification Pipeline for Detecting Analgesic Efficacy from Neuroimaging Data
WANIGASEKERA V., DUFF E., TRACEY I.
Aim: Functional magnetic resonance imaging (FMRI) can provide objective evidence of target engagement and potential efficacy of analgesics in the early clinical phases of drug development. Such evidence-based go/no-go decisions on compound selection has the potential to improve the efficiency of the analgesic drug discovery process. We have previously developed a classification-based protocol for assessing the effects of analgesics on brain responses to pain as measured by FMRI1. It utilises machine learning to derive signatures of analgesic response across many brain regions from existing FMRI studies of analgesics. This protocol can detect evidence of pharmacodynamic effects where a compound shows consistent effects on brain responses across individuals, and efficacy where the compound’s effects correspond to an analgesic signature derived from the classification database. To further validate the protocol for drug development, we set out to test this protocol in a completely independent FMRI data set (n=24) from a double-blind, randomised, placebo-controlled, three-way crossover study in a healthy volunteer model of central sensitization (CS), a crucial mechanism underpinning neuropathic pain states. This study assessed whether brain responses to painful stimuli can differentiate a clinically effective analgesic in neuropathic pain (gabapentin) from an ineffective analgesic (ibuprofen) and both from placebo2. We expected our protocol to detect gabapentin to show evidence of effective analgesia but not ibuprofen. Methods: We separately assessed the gabapentin and ibuprofen arms of the study with the protocol. The pharmacodynamic effects were determined by assessing (using leave-one-out method) whether a trained classifier could discriminate the drug arm of the study from placebo based on brain responses to painful stimuli. The signature of efficacy was derived from five double-blind, randomised, placebo-controlled FMRI studies with a cross over design. These studies compared the brain effects of different classes of various analgesics (remifentanil, tramadol, pregabalin, and delta-9-tetrahydrocannabinol) to placebo and did not include the drugs under assessment. We used independent component analysis for dimensionality reduction, and a support vector machine with a linear kernel for classification. We assessed discrimination accuracy (drug vs placebo) as a measure of the presence of pharmacodynamic and analgesic signature effects, where chance p = 0.5. Results: Gabapentin, but not ibuprofen, showed evidence of a reliable pharmacodynamic effect. Using a classifier based on the brain responses to painful stimuli observed in other individuals, the brain response to painful stimuli with gabapentin could be correctly distinguished from placebo in 79% of individuals (p=0.003). In contrast, responses following ibuprofen could not be distinguished from placebo (45%, p=0.72). In the assessment for analgesic efficacy, the classifier could correctly identify the gabapentin arm in 17 of 24 subjects (p=0.03), indicating that the compound shows brain effects resembling those found in our training set of efficacious compounds. Ibuprofen showed no such evidence, with discrimination below chance (p=0.92). At a study level, the gabapentin arm showed robust evidence for our signature of analgesic efficacy while ibuprofen did not. We also found evidence that specific analgesic compounds had distinct signatures. A classifier trained on a separate study of gabapentin could reliably identify the gabapentin conditions in this study (p=0.03). However, classifiers trained on studies of the opioid remifentanil failed to identify gabapentin in this data (p=0.5), while successfully identifying effects in remifentanil studies. Conclusion: Here we show that a machine learning protocol generated to detect analgesic efficacy is able to detect brain changes related to analgesic efficacy of a known effective analgesic (gabapentin) in a data set independent of the training data set; however, it failed to detect analgesic effects of ibuprofen. This is in keeping with the known clinical efficacy of these compounds in neuropathic pain, as well as our original study that independently showed gabapentin and not ibuprofen or placebo suppressed neural activity evoked by painful stimuli2. Our classifier failed to detect any ibuprofen induced pharmacodynamic effects on the brain. Ibuprofen is not a centrally acting compound. It has predominantly peripherally mediated analgesic effects; therefore, ibuprofen is unlikely to contribute to measurable effects on pain related brain activity across individuals. Additionally, we are able to show that machine learning protocols can identify brain activity that is specific to the drug class. In the future, we aim to apply our protocol to assess compounds with potential yet clinically unproven analgesic efficacy in early drug development. 1. Duff EP, Vennart W, Wise RG, et al. Learning to identify CNS drug action and efficacy using multistudy fMRI data. Sci Transl Med 2015; 7(274): 274ra16. 2. Wanigasekera V, Mezue M, Andersson J, Kong Y, Tracey I. Disambiguating Pharmacodynamic Efficacy from Behavior with Neuroimaging: Implications for Analgesic Drug Development. Anesthesiology 2016; 124(1): 159-68.