Eric J Ma's Website

At IMED 2016, I heard a lot of talks surrounding surveillance efforts. Most of it has been syndromic surveillance, because that's what's currently collectable. For digital disease surveillance, search queries (a marker of "interest" in a disease) are used as a proxy for probable presence of a disease.

These are good for real-time surveillance, but I think it's not good enough from the standpoint of accuracy and understanding mechanism.

For accuracy, syndromic surveillance and search queries are n^th-degree manifestations of underlying causes. Multiple different pathogens can present the same symptoms. Upticks in searches for a given symptom or pathogen may have multiple underlying causes, though "hearing of the disease" and "being concerned about it" may be related to certain diagnoses and local spread of information.

Found data can help act as a very early sentinel to a problem, and can guide us by providing leads to follow-up, but for understanding mechanism, neither tell us how a new virus is going to be dangerous.

In other words, while fast and rapid and local, syndromic surveillance and digital disease surveillance remains a reactive endeavour, because we can't mechanistically forecast risk from the data.

On the other hand, genomic data is often collected with the intention of understanding pathogen dynamics, as we have well-developed phylogenetic tools for this. However, we still rely on single point mutations, identified in small-scale, one-off, ad-hoc, [insert-some-dashed-phrase-that-expresses-frustration] studies, and basically treat the limited amount of evidence as gospel. I am quite sure that is not the right way to do phenotypic interpretation of genomic data.

If we're going to make disease surveillance a proactive endeavour, I am convinced that the disease surveillance world will need a paradigm shift, and investment in infrastructure. By predictive, this is what I mean:

1. Calculating what has a high probability of jumping into humans.
2. Phenotypically testing those viruses without using live viruses (concept of pseudotyping).
3. Using genomic information to predict phenotype.

Here's how I see it happening, given everything I've learned at IMED 2016:

(1) Sequence everything, in animals that interface with humans

60-70% of new infections are going to be zoonotic in nature. To know what might be coming up, use metagenomics to sequence animal species in geographic areas that humans, prior to the host switch event. To me, this is a no-brainer. Long-term serial sampling is going to be key.

(2) Leverage pseudotyping technologies and genetic systems, and automation, to create systematic phenotyping platforms

Doing so will allow us to phenotypically test components of a virus for a particular step in its life cycle, such as speed of entry into a cell, or resistance to a drug. With rapid DNA synthesis & assembly technologies, we should be able to experimentally measure a virus' epidemiologically-relevant phenotype within 24 hours of an outbreak. We should also be able to rapidly generate large-scale genotype-phenotype data for machine learning.

What would be relevant phenotypes? I can think of a few, broken down basically to "druggability" and "life cycle steps":

1. For druggable proteins, measure drug resistance. It has high epidemiological relevance, and hence, value.
2. For viruses targetable by drugs and not alike, measure cellular entry capability.
3. For identified polymerase complexes, measure replication capacity.
4. For cellular release machinery, if identified, measure release.
5. If there's others already existing for which the case can be made for systematic phenotyping, scale the assay up.

(3) Develop machine learning tools that allow for learning of structural properties that determine epi-relevant protein phenotype

By doing so, we can explicitly model epistasis in a protein, and therefore account for epistatic interactions explicitly when developing learning algorithms that map genotype to phenotype. With this kind of technology coupled to the ability to sequence a virus within hours of sample collection, we'll be able to cut down the time for phenotypic interpretation from the order of a day (because of experimentation required) to seconds (because it's a computational prediction). By going Bayesian, we can even model uncertainty in final predictions, thereby aiding decision-making under uncertainty.

In conjunction with the "sequence everything" approach, systematic phenotyping and deep learning will allow us to move to a "predictive" paradigm because it'll give us the ability to rationally predict the phenotypic risk of a virus before it even jumps into humans. I see this as an infrastructural project: requiring up-front commitment of time and money. Yet, we know that investments in infrastructure always pay off dividends in a myriad of ways, from time saved to lives saved.

Something I learned from IMED 2016, however, is that epidemiology is an old discipline, and with old disciplines comes a certain tunnel vision on how things should be done. It'll be up to a new and creative generation (i.e. us) to convince the funders and old guard that our new way will add value to the discipline.