Thread by @cecilejanssens

Why it is so hard to predict complex diseases and traits from DNA? Here is a collection of my go-to slides, a short story about the predictive ability of polygenic risk scores (PRS).

The genetic origins of complex diseases (and traits) is different from monogenic diseases: it is not one gene with a major effect but complex interactions between many genetic variants and nongenetic factors, such as diet, exercise, smoking, housing, etc, with small effect each.

Genetic variants play a bigger role in some diseases than others. This role is indicated as the heritability. For an explainer about heritability: theconversation.com/explainer-what…

Heritability is not a ‘feature’ of a disease. It varies for the same disease between populations and subtypes. For example, genes play a larger role in hip fractures among the young, but most hip fractures occur among the elderly, when poor vision and imbalance come into play.

The genetic variants that are identified to play a role in hip fractures among the young may therefore, at best, only predict fractures in young but not in older people.

The heritability dictates the maximum predictive ability of PRS for a disease, in a specific population. This tells that if the genetic roles to the causes of complex diseases are entirely understood, a DNA scan can predict some with higher discriminative accuracy than others.

What AUC means … see other thread:

https://twitter.com/cecilejanssens/status/1104134423673479169

Whether we *can* completely understand all genetic factors and their interactions depends (among other things) on the complexity of the causes. That is not unique to complex diseases, but also relevant for disasters: can we understand the causes of disasters?

Here, I often share the story of the herald of free enterprise, a ferry that capsized in good condition and a great crew on a clear day. Researchers found out several errors and omissions that together may have contributed to its capsizing.

They summarized these factors in a causal diagram, attributing the cause to a unique (but unprovable) combination of 13 factors that happened in the right sequence and timing to cause the ferry to capsize.

This combination of factors may explain the capsizing of this ship but is unlikely to occur again. This makes intervention on contributing (risk) factors, rather than mechanisms, the best way to prevention. Complex diseases are no different. It's diet, exercise, low stress, ...

A paper worth reading in this context is that of @ken_rothman & Greenland about the causal mechanisms of complex causes.
Each patient who develops a disease may have their own causal mechanism, unknowable before the onset of disease, 'guessable' afterwards.

Thus, the predictive ability of polygenic risk scores is a function of the heritability and complexity, of the quality of the model and the data: even when phenotypes are the same (capsized ships or complex diseases), their causes may be too complex to model or be known.

In sum, if we want to use DNA to predict diseases or traits, the predictive ability will be somewhere between perfect and random, depending on the heritability and complexity of their origins.

If you think about all that can be predicted using DNA, where on the scale would you place these, left or right?

Here are my thoughts, including a synthesis of what is on the left and on the right (Here I always make the same joke that “susceptibility to common complex diseases” is a long label on purpose, there are positive exceptions)

After a discussion of predictive ability, it’s always good to realize that predictive ability does not need to be perfect to be useful. Clinical utility is determines how high predictive ability needs to be or how low is predictive enough.

And that brings us here, a one slide summary of what DNA can and cannot tell.

For more on prediction research, download a plain language manual here: cecilejanssens.org/wp-content/upl….
The slides of this thread can be downloaded here: cecilejanssens.org/slides-from-tw…