Model Overview
SLIM is a probabilistic, statistical epidemiological simulator. Compared to IBM approaches we simulate overall populations and associated distributions and perform statistical operations on them rather than emulating individual lice or fish.
At the same time, SLIM is also a game simulator where each agent (farmers) seeks to maximise their own profit and may choose to collaborate or defect (more on this later), which treatment to apply and so on.
This simulation includes explicit encoding of genomic effects into the evolution of the pandemic. Treatments may encourage resistant traits to proliferate and thus make the former less and less effective at each cycle.
Note
This section assumes you have some basic knowledge of salmon aquaculture. If you don’t we suggest having a look at Rationale or 1 first.
The project models an environment in which an slim.simulation.organisation.Organisation
of salmon farmers  which reside
and operate on the same loch  run their own slim.simulation.farm.Farm
s which are in turn divided into
slim.simulation.cage.Cage
s.
A cage is the physical location of salmons. Initially all cages within the same farm are filled with smolts which are then left to grow up and harvested after 1.52 years.
Sea lice stay in the reservoir and, either thanks to water currents or wild salmon, leak into salmon cages. Once they find a host to attach to they evolve from recruitment or copepopid to chalimus and beyond. SLIM models cagespecific, stagespecific lice populations and assumes concentration and pollution effects as farms are operative for long periods of time.
As lice attach to fish they cause a number of diseases which result in fish death. If the lice aggregation ratio exceeds a given threshold treatment must be performed. However, other managers not reaching that threshold may be faring better and opt to defect by failing to treat.
Whether egoism backfires or not is not an easy question as it depends on the current stage of the epidemics, lake pollution and how water currents work in the specific simulation environment.
Note that farm managers are individual agents, meaning they have full freedom to stock cages in any season they prefer and treat when needed, and also to decide which treatment to apply.
Note
The exact formulae behind this are documented in our paper. When it comes out please check it out!
Sea Lice Population
Our model is built on 3 main papers 2 3 4:
Just like in 2, we model the lice lifecycle as a compartmentalised model of 6 stages: recruitment (R),
copepopids (CO), chalimus (CH), preadults (PA) and adults, the latter divided in males (AM) and females (AF).
Differently from 2 we omit the explicit age encoding but instead group the population
by its genotype. Nevertheless, the age distribution can be simulated at will via
slim.simulation.cage.Cage.get_stage_ages_distrib()
.
A genotype distribution of the lice population as related to resistance to treatment is modelled by the
slim.simulation.lice_population.GenoDistrib
class. A genotype distribution
is, ultimately, a discretised dictionary with where the keys are allele combinations (see
slim.simulation.lice_population.GenoDistrib.alleles
) and the values are the actual number of lice in that group.
In a similar way to 2, each day we estimate how many lice die or evolve at every stage.
All stages are subject to the following:
background mortality;
treatmentinduced mortality;
cleaner fish mortality (not yet implemented);
evolution (can be either additive from the previous stage or subtractive to the successive stage);
age mortality (i.e. lice that could not evolve will die).
In all stages but AM or AF lice can evolve. Evolution from PA will result in roughly 50% split between AM and AF.
Gene encoding
Let \(\mathbb{G}\) be the set of genes. Each gene \(g\) can appear in at 3 forms (alleles): dominant (represented as \(G\)), recessive (\(g\)) and heterozygous dominant (\(Gg\)), typically grouped in the set \(\mathbb{A_g}\). \(A2G(G) = g\) gives the representative gene of an allele.
We represent the genotype distribution of the population \(m\), time \(t\), farm \(f\) and cage \(c\) as \(G^m_{tfc}\).
This function is defined as \(\mathbb{Allele} \rightarrow \mathbb{N}\), that is it associate alleles to a discrete number.
The frequency of a single allele \(x\) is represented by \(G^m_{tfc}(X=x)\), with the following constraints:
All the frequencies of the alleles belonging to the same gene must sum to the same gross count
All the frequencies of distinct genes must coincide  i.e. they provide different yet consistent views on the same distribution
All the frequencies must be positive integers.
In symbols:
One can see the different genes as independently identically distributed (i.i.d.) variables with no correlation or whatsoever.
Treatment
Treatment affects the lice population in different ways, depending on the genotype of the population and the
type of treatment. We model two types of treatment: chemical and nonchemical treatments. The full list
of supported treatments can be found in slim.types.treatment_types.Treatment
.
When a treatment is administered some delay occurs before effects are noticeable (nonchemical
treatments have a virtual delay of one day). The mortality rate (a multiplicative factor of the given
population) is computed in
slim.simulation.cage.Cage.get_lice_treatment_mortality_rate()
. In the case of EMB
it is the following:
where:
\(t, f, c\) represent the current time, farm and cage;
\(g\) is the chosen genotype;
\(\phi^EMB\) is the phenotype resistance corresponding to the given genotype. The codomain is in \([0,1]\)
\(t_{fcb}\) is the time when a treatment was started;
\(\delta^{EMB}\) is the delay of the treatment;
\(\Delta^{dur}\) is the efficacy duration, computed as \(\delta^{dur} / T_{t^0}\) where \(T_{t^0}\) is the average water temperature when the treatment is applied and \(\delta^{dur}\) is a constant.
In other words, if the current time falls within the efficacy timeframe of a treatment the mortality rate is computed as the inverse of the resistance rate provided by \(\phi^{EMB}\).
Once the mortality rates are computed for each genotype, we use a Poisson distribution to generate the mortality events and a hypergeometric distribution to choose from which stages to remove lice.
For the sake of notation the stage has been omitted but not all stages are taken into account. For examples, since EMB affects lice attachment only the stages from Chalimus onward are affected.
For more information check slim.types.TreatmentTypes
.
Evolution
Similarly to 5 we avoid to model age dependencies explicitly to keep the model complexity simple and rather use explicit rates across stages. However, due to interesting interactions between fish population, weight and lice infection we preferred to use 2’s approach to infection.
Stage 
Approach 
Source 

Eggs 
Explicit hatching queue 

Nauplius 
Evolution rate 

Copepopids 
Explicit infection 

Preadults 
Evolution rate 

Adults 
Evolution rate 
Reproduction
During mating alleles are recombined according to a Mendelian approach. The number of reproduction events is calculated on the estimated number of matings that can happen on a single host. We assume a scenario in which one female lice can mate with only one male lice before being fecundated. As in 3 we estimate such number via a negative multinomial distribution.
Assuming a louse can attach to either fish, we reduce this problem to finding the probability that \(k\) lice land to \(n\) fish, and then find such expectation.
The expectation can be found to be: 6
To use the negative multinomial distribution we also need a VMR: variancetomean ratio. The variance in this case refers to the variance of the lice occupation \(Y \sim Multinomial(\textbf{p} = \left(\frac{1}{n}, \ldots \right))\) which we reduced to a multinomial problem. Thus the formula for the variance is the usual (with a notation abuse for X):
The probability of a matching occurring between two lice on the same host is thus:
where:
The number of matings is thus just a proportion on the number of free females (denoted with \(N_AF'\))
The number of produced eggs is defined in a similar way to 2 and follows a power law
parametrised on the (virtual) age distribution. For details see slim.simulation.cage.Cage.get_num_eggs()
Once an adult female has been breed she enters a refractory period that lasts up to 3 days.
Matings and gene inheritance
Explicit mating modelling would be pointless without addressing the genomics.
In principle, inheritance follows a Mendelian scheme: because of heterozygosity there are only a handful of valid pairings that can yield to 3 distinct traits, each of them with different probabilities depending on the parents’ genes.
To accommodate the need for fast processing, we use double counting to consider the number of possible matches, then use a multinomial distribution to generate the bespoke genotype distribution given such (normalised) pairings.
Offspring distribution
There are two mechanisms that distribute the offspring: farmtofarm movements and external pressure integration.
FarmtoFarm movements
The majority of lice offsprings is typically lost, some are transmitted to neighbouring farms and very few are reintegrated into the reservoir. For each pair of farms \(f_i\) and \(f_j\) we have data on the travel distance \(d_{i,j}\) and lice rate \(r_{i,j}\).
If a farm \(f_i\) generates \(N^{Egg}_i\) eggs with genotype distribution \(G^{Egg}_i\) these will be distributed to the farm \(f_j\) with the following formulae:
Where \(\lambda\) is the Poisson distribution.
External Pressure
Quite differently from previous authors, we only model the external pressure without modelling a reservoir as a special cage. Instead, we consider the reservoir as an infinite, _dynamic_ generator of new lice.
Indeed, we assume that a fraction of the eggs extruded during mating will be reabsorbed by the loch rather than just disappearing, for example by wild salmons. The external pressure therefore emits a dynamic number of lice depending on two things:
the average offspring throughput over the last 30 days;
the genotype distribution of such offspring.
As for the first requirement, the overall number of lice is: \(N_{EXT}^t = N_{EXT}^0 + k{N_{Eggs}^{t30...t1}}\). As for the second, we use a DirichletMultinomial Bayesian process to infer the genotype ratio of the new lice. The objective is to guarantee a _reserve_ of each genotype (even rare ones) while favouring the most prolific trait.