How can probabilistic algorithms solve problems in biology?

One of the key questions for these algorithms is what is the objective function, i.e. how will we be able to test automatically that the output makes enough sense to be worth further investigation. Deep mind did this with Alpha Sort, checking the many variants of output did in fact work. This is much harder for biological outputs were testing without lab work is difficult.

Noise at the molecular and cellular level trickles up to the organismal level. No human beings have exactly the same face, the same gene expression, the same placement of moles on their skin. Making models, in light of this intrinsic randomness, requires probabilistic algorithms.

In Biological modelling, not only is there uncertainty, but there is also irreducible variable - unknowability, if you will. Many relationships in biology do not have a unique set of model parameters modelling them - to accurately portray what is going on, models need to capture the highly probabilistic nature of the problem. For instance, a lot of work has gone into modelling neuronal dynamics in connectomes (maps of neurons in the brain and nervous system). We know now that the same neuron at different times, and that the same neuron in different animals, vary in how they fire, even while causing the same system behaviour. We have to simulate parameter variability to differentiate signal and noise.

They can help explore a space of solutions which would otherwise be hard to compute. What is more difficult is computing the effect in silicon of compounds the algorithms produce. They are likely best used as a guide for new places to search rather than as likely to find a direct workable outcome.

At the atomistic level, we have molecular dynamics simulations that simulate the jiggling of atoms in order to capture the dynamics of molecular motions. At the level of whole molecules, we have the Gillespie simulation which simulates when events happen and which events happen. At even larger scales approximate models including thermodynamic (technically statistical mechanical) models are used to simulate whole organisms. However, given data, a whole set of other probabilistic algorithms are used to learn models from data. These often are broken down into methods that work on continuous or discrete space and continuous or discrete time. For instance, the Hidden Markov model is one such example operating in discrete space & time.

There are a number of critical challenges. For example, there may be many a priori unknown parameters, and even unknown mathematical structures of problems, that must be calibrated in order to use a probabilistic algorithm to realistically simulate a biological event. Next, it is often unclear whether the parameters calibrated are known perfectly or come with associated uncertainty. An alternative, is to use the data itself in order to learn the model or its parameters. Yet this approach (termed "inverse modeling") presents a slew of other challenges. Just as an example, on the practical side, the data sets to be analyzed can be large and unwieldy especially in dealing with volumetric imaging data sets making even data storage an issue.

One of the great difficulties of all large-scale data driven algorithms is how to communicate the results to humans, how to bring enough representation of context and limitations to the answer. There are certainly not enough communication methods for interpreting model outputs so that a decision can be made on viable next steps.