Disability progression in multiple sclerosis remains resistant to treatment. The absence of a suitable biomarker to allow for phase 2 clinical trials presents a high barrier for drug development. We propose to enable short proof-of-concept trials by increasing statistical power using a deep-learning predictive enrichment strategy. Specifically, a multi-headed multilayer perceptron is used to estimate the conditional average treatment effect (CATE) using baseline clinical and imaging features, and patients predicted to be most responsive are preferentially randomized into a trial. Leveraging data from six randomized clinical trials (n = 3,830), we first pre-trained the model on the subset of relapsing-remitting MS patients (n = 2,520), then fine-tuned it on a subset of primary progressive MS (PPMS) patients (n = 695). In a separate held-out test set of PPMS patients randomized to anti-CD20 antibodies or placebo (n = 297), the average treatment effect was larger for the 50% (HR, 0.492; 95% CI, 0.266-0.912; p = 0.0218) and 30% (HR, 0.361; 95% CI, 0.165-0.79; p = 0.008) predicted to be most responsive, compared to 0.743 (95% CI, 0.482-1.15; p = 0.179) for the entire group. The same model could also identify responders to laquinimod in another held-out test set of PPMS patients (n = 318). Finally, we show that using this model for predictive enrichment results in important increases in power.

No abstract

Modeling treatment effect could identify a subgroup of individuals who experience greater benefit from disease modifying therapy, allowing for predictive enrichment to increase the power of future clinical trials. We use deep learning to estimate the conditional average treatment effect for individuals taking disease modifying therapies for multiple sclerosis, using their baseline clinical and imaging characteristics. Data were obtained as part of three placebo-controlled randomized clinical trials: ORATORIO, OLYMPUS and ARPEGGIO, investigating the efficacy of ocrelizumab, rituximab and laquinimod, respectively. A shuffled mix of participants having received ocrelizumab or rituximab, anti-CD20-antibodies, was separated into a training (70%) and testing (30%) dataset, but we also performed nested cross-validation to improve the generalization error estimate. Data from ARPEGGIO served as additional external validation. An ensemble of multitask multilayer perceptrons was trained to predict the rate of disability progression on both active treatment and placebo to estimate the conditional average treatment effect. The model was able to separate responders and non-responders across a range of predicted effect sizes. Notably, the average treatment effect for the anti-CD20-antibody test set during nested cross-validation was significantly greater when selecting the model’s prediction for the top 50% (HR 0.625, p=0.008) or the top 25% (HR 0.521, p=0.013) most responsive individuals, compared to HR 0.835 (p=0.154) for the entire group. The model trained on the anti-CD20-antibody dataset could also identify responders to laquinimod, finding a significant treatment effect in the top 30% of individuals (HR 0.352, p=0.043). We observed enrichment across a broad range of baseline features in the responder subgroups: younger, more men, shorter disease duration, higher disability scores, and more lesional activity. By simulating a 1-year study where only the 50% predicted to be most responsive are randomized, we could achieve 80% power to detect a significant difference with 6 times less participants than a clinical trial without enrichment. Subgroups of individuals with primary progressive multiple sclerosis who respond favourably to disease modifying therapies can therefore be identified based on their baseline characteristics, even when no significant treatment effect can be found at the whole-group level. The approach allows for predictive enrichment of future clinical trials, as well as personalized treatment selection in the clinic.