Counterfeit and substandard medicines which contain little or no active ingredients are widespread in low-middle income countries (LMIC’s), often as a result of lack of government oversight over pharmaceutical supply chains, and of laboratory facilities for carrying out quality control. Apart from lacking efficacy, counterfeit and substandard medicines may pose direct health risks by containing toxic ingredients, leading to poisoning and even death of patients. In 2012, more than 100 people died from taking tainted heart medication in Pakistan. In parts of Africa and Asia, it is thought more than 50% of all pharmaceuticals sold are fake.
Needless to say, routinely available, inexpensive and easy-to-use assays for the analysis of active ingredients and contaminants could have significant public health impacts. At the University of Notre Dame, Indiana, Prof. Marya Lieberman and colleagues are working on the Paper Analytical Device (PAD) Project, a research project with the tangible aim of bringing chemical analysis from the lab out to where it is needed. The basic PAD technology is a 12-lane paper card containing reagents to identify specific chemicals or functional groups. When the test drug is “swiped” over the card, which is subsequently placed in water, chemical reactions cause colour changes which are visible to the naked eye. For example, a Cu(II) lane turns forest-green in the presence of beta-lactam type antiobiotics and a different colour for the oral diabetes medication metformin. The combination of different colours can be read like a barcode to determine whether active ingredients are present, and inert fillers and unexpected ingredients are not.
Sandipan Banerjee and other colleagues in prof. Lieberman’s lab have taken the technology a step further by developing a machine learning algorithm for automatically reading analytical results. The authors tested two general approaches: Convolutional Neural Networks (CNN), a type of neural network modelled on the biological visual cortex, which is particularly suited for image recognition; and either nearest-neighbour or support vector machine models. The latter approach requires manual extraction of image features such as colour histograms, while the CNN model extracts features automatically. The CNN model resulted in th highest accuracy, correctly calling just over 94% of samples, while the feature based algorithms ranged in accuracy from 53% to 92%.
The PAD technology is somewhat limited in terms of quantitative detection, i.e. the extent to which reduced amounts of active ingredient are present. Abigail Weaver, also of Prof. Lieberman’s lab, has shown the technology can detect drugs cut with more than 60% inert filler, while samples cut with a lower proportion of filler could not be identified. It is also unclear how multi-sourced pharmaceuticals would impact PAD results, as different generics manufacturers may use different inert compounds, excipients etc. However, as a qualitative analytical method, PAD appears to be a promising and scalable technology with the potential for significant public health impact, particularly in countries with the least existing regulatory oversight and analytical capacity.
More information can be found at the PAD project website.