inVerse: Using AI to learn from success and failure

The question to be answered for personalised treatment discovery was formulated as follows

Is it possible to analyze clinical trials conducted for multiple sclerosis (MS) and find potential signals which could help to identify candidate personalised therapies?

Volv performed analysis across the studies highlighted above and it produced 22 ‘core’ models with which to predict patient response to 15 important MS therapies and facilitate personalized treatment optimization. In the example below we focus on one of those core models.

• Beyond this we then performed further analysis to provide concise summaries of:
  
• ‘new concepts’ identified via computational analysis of clinical trials (e.g. promising disease-modifying therapies (DMTs), novel biomarkers for responder/non-responder prediction);
• simplified/accessible versions of prediction models.
  

We identified ten ‘core’ sets of models and their main purposes (e.g. optimal therapy selection, assessment of unapproved treatments), and for each core model set, we then specified type/output/predictive features for all component models and, where feasible, derive simplified versions of the models using our clinically interpretable model method from inTrigue.

Contact us for more details.

Using inVerse to perform Predictive Analysis of MSBase Studies

Background:

• personalized therapy is of considerable interest in MS, and many clinical trials have compared safety and efficacy across different MS therapies,
• most work on personalized therapy has focused on predicting patient response to a specific therapy rather than deciding which of several treatments would deliver optimal outcome;

here we examine whether different patients do indeed have different optimal therapies and, if so, which biomarkers would facilitate personalized MS treatment selection.

Observational/experimental data:

• data sources include MSBase (www.msbase.org), a global MS cohort study, and seven clinical trials which post quantitative results;
• data is in aggregate form and includes patient features (demographics, clinical, paraclinical) as well as response to treatment for six different outcome measures.

Analysis Process

Predictive analysis consists of three main steps:

1. Data preprocessing:
   ‘reverse-engineer’ available aggregate datasets to yield estimates for the characteristics and treatment responses of small bags of study patients.
2. Predictive modeling:
   for each of five therapies (interferon b, glatiramer acetate, fingolimod, natalizumab, and mitoxantrone), learn an individual-level prediction model from patient bag data and use models to predict treatment response for 8513 patients in the study dataset.
3. Predictive analysis:
   1. determine the optimal therapy for two outcome measures for each of 8513 patients;
   2. identify biomarkers that signal which of the five target therapies should be used.

Features predictive of response were identified using feature importance and forward/ backward feature selection [HTF 2009] and, interestingly, vary according to therapy type. Predictive features include

• for fingolimod, natalizumab, mitoxantrone:
  
  • 1 example: Expanded Disability Status Scale (EDSS) score;
  • other features have been redacted, contact us for more information.
• for glatiramer acetate,
  
  • 1 example: interferon b: EDSS;
  • other features have been redacted, contact us for more information.
• note: analysis of related dataset suggests lymphocyte subpopulation statistics are predictive for fingolimod response

Note on datasets:

Analysis is informed by seven clinical trials together with data from two associated papers.

We have applied computational analysis to identify clinical trials expected to be of value for personalisation of treatment. From an initial set of ~2000 candidate trials, this procedure resulted in the collection/preprocessing of 45 project-relevant datasets involving 18 unique therapies.

Analysis of these datasets shows:

• accurate patient-level therapy-response models can be learned from aggregate datasets (usually only aggregate data is posted in clinical trial reports and other publicly-available databases);
• it produced 22 ‘core’ models with which to predict patient response to 15 important MS therapies and facilitate personalized treatment optimization;
• identified predictive features/biomarkers for each model, some of which appear novel and almost all of which are measurable in the course of routine MS care.

Predictive modeling yielded a number of interesting findings, including:

• detection of unsuccessful clinical trials which could have been successful had proper cohort selection been performed (e.g. CONCERTO laquinimod trial); crucially, predictive patient selection can be achieved using data available before the trial is initiated;
• demonstration that personalized therapy optimization models can be learned for established first-line and second-line treatments (e.g. interferon b, glatiramer acetate, fingolimod, natalizumab, mitoxantrone) as well as for less commonly-used drugs;
• derivation of responder/non-responder prediction models for the five therapies listed above and also for fampridine (based on clinical biomarkers), dimethyl fumarate (blood-borne biomarkers), intrathecal corticosteroid therapy (CSF/clinical data), rituximab (MRI data), injectables (clinical data), apheresis therapy (brain biopsy data), laquinimod (clinical/MRI data), and teriflunomide (early clinical response).

Volv performed analysis across the studies highlighted above and it produced 22 ‘core’ models with which to predict patient response to 15 important MS therapies and facilitate personalized treatment optimization. In the example below we focus on one of those core models.

• Beyond this we then performed further analysis to provide concise summaries of:
  
• ‘new concepts’ identified via computational analysis of clinical trials (e.g. promising disease-modifying therapies (DMTs), novel biomarkers for responder/non-responder prediction);
• simplified/accessible versions of prediction models.
  

We identified ten ‘core’ sets of models and their main purposes (e.g. optimal therapy selection, assessment of unapproved treatments), and for each core model set, we then specified type/output/predictive features for all component models and, where feasible, derive simplified versions of the models using our clinically interpretable model building from inTrigue.

Find out more about inVerse