Obtaining genetics insights from deep learning via explainable artificial intelligence.

Artificial intelligence (AI) models based on deep learning now represent the state of the art for making functional predictions in genomics research. However, the underlying basis on which predictive models make such predictions is often unknown. For genomics researchers, this missing explanatory information would frequently be of greater value than the predictions themselves, as it can enable new insights into genetic processes. We review progress in the emerging area of explainable AI (xAI), a field with the potential to empower life science researchers to gain mechanistic insights into complex deep learning models. We discuss and categorize approaches for model interpretation, including an intuitive understanding of how each approach works and their underlying assumptions and limitations in the context of typical high-throughput biological datasets.

A novel approach to assessing the ecosystem-wide impacts of reintroductions.

Reintroducing a species to an ecosystem can have significant impacts on the recipient ecological community. Although reintroductions can have striking and positive outcomes, they also carry risks; many well-intentioned conservation actions have had surprising and unsatisfactory outcomes. A range of network-based mathematical methods has been developed to make quantitative predictions of how communities will respond to management interventions. These methods are based on the limited knowledge of which species interact with each other and in what way. However, expert knowledge isn't perfect and can only take models so far. Fortunately, other types of data, such as abundance time series, is often available, but, to date, no quantitative method exists to integrate these various data types into these models, allowing more precise ecosystem-wide predictions. In this paper, we develop mathematical methods that combine time-series data of multiple species with knowledge of species interactions and we apply it to proposed reintroductions at Booderee National Park in Australia. There have been large fluctuations in species abundances at Booderee National Park in recent history, following intense feral fox (Vulpes vulpes) control, including the local extinction of the greater glider (Petauroides volans). These fluctuations can provide information about the system isn't readily obtained from a stable system, and we use them to inform models that we then use to predict potential outcomes of eastern quoll (Dasyurus viverrinus) and long-nosed potoroo (Potorous tridactylus) reintroductions. One of the key species of conservation concern in the park is the Eastern Bristlebird (Dasyornis brachypterus), and we find that long-nosed potoroo introduction would have very little impact on the Eastern Bristlebird population, while the eastern quoll introduction increased the likelihood of Eastern Bristlebird decline, although that depends on the strength and form of any possible interaction.

Environmental variables influence patterns of mammal co-occurrence following introduced predator control.

Co-occurring species often overlap in their use of resources and can interact in complex ways. However, shifts in environmental conditions or resource availability can lead to changes in patterns of species co-occurrence, which may be exacerbated by global escalation of human disturbances to ecosystems, including conservation-directed interventions. We investigated the relative abundance and co-occurrence of two naturally sympatric mammal species following two forms of environmental disturbance: wildfire and introduced predator control. Using 14 years of abundance data from repeat surveys at long-term monitoring sites in south-eastern Australia, we examined the association between a marsupial, the common brushtail possum Trichosurus vulpecula, and a co-occurring native rodent, the bush rat Rattus fuscipes. We asked: In a fox-controlled environment, are the abundances of common brushtail possums and bush rats affected by environmental disturbance and each other's presence? Using Bayesian regression models, we tested hypotheses that the abundance of each species would vary with changes in environmental and disturbance variables, and that the negative association between bush rats and common brushtail possums was stronger than the association between bush rats and disturbance. Our analyses revealed that bush rat abundance varied greatly in relation to environmental and disturbance variables, whereas common brushtail possums showed relatively limited variation in response to the same variables. There was a negative association between common brushtail possums and bush rats, but this association was weaker than the initial decline and subsequent recovery of bush rats in response to wildfires. Using co-occurrence analysis, we can infer negative relationships in abundance between co-occurring species, but to understand the impacts of such associations, and plan appropriate conservation measures, we require more information on interactions between the species and environmental variables. Co-occurrence can be a powerful and novel method to diagnose threats to communities and understand changes in ecosystem dynamics.

INGOT-DR: an interpretable classifier for predicting drug resistance in M. tuberculosis.

MOTIVATION: Prediction of drug resistance and identification of its mechanisms in bacteria such as Mycobacterium tuberculosis, the etiological agent of tuberculosis, is a challenging problem. Solving this problem requires a transparent, accurate, and flexible predictive model. The methods currently used for this purpose rarely satisfy all of these criteria. On the one hand, approaches based on testing strains against a catalogue of previously identified mutations often yield poor predictive performance; on the other hand, machine learning techniques typically have higher predictive accuracy, but often lack interpretability and may learn patterns that produce accurate predictions for the wrong reasons. Current interpretable methods may either exhibit a lower accuracy or lack the flexibility needed to generalize them to previously unseen data. CONTRIBUTION: In this paper we propose a novel technique, inspired by group testing and Boolean compressed sensing, which yields highly accurate predictions, interpretable results, and is flexible enough to be optimized for various evaluation metrics at the same time. RESULTS: We test the predictive accuracy of our approach on five first-line and seven second-line antibiotics used for treating tuberculosis. We find that it has a higher or comparable accuracy to that of commonly used machine learning models, and is able to identify variants in genes with previously reported association to drug resistance. Our method is intrinsically interpretable, and can be customized for different evaluation metrics. Our implementation is available at github.com/hoomanzabeti/INGOT_DR and can be installed via The Python Package Index (Pypi) under ingotdr. This package is also compatible with most of the tools in the Scikit-learn machine learning library.

Can evolutionary theories of dispersal and senescence predict postrelease survival, dispersal, and body condition of a reintroduced threatened mammal?

Theories of dispersal and senescence (or aging) predict that dispersal, and ongoing survival and body condition, are influenced by evolutionary drivers, along with intrinsic and extrinsic factors. Such theories are relevant to translocations of animals where high mortality, loss of body condition, and dispersal beyond the area of release are commonly reported. However, these theories have rarely been tested using data from translocations.We explore whether theories of dispersal and senescence, together with biological knowledge and management interventions, can predict rates of postrelease dispersal, survival and change in body condition of a translocated endangered meso-predator, the eastern quoll Dasyurus viverrinus.Captive-bred quolls (n = 60) from three sanctuaries were translocated to an unfenced, predator-managed reserve (Booderee National Park) over 2 years (2018, 2019). Survival, dispersal and body mass were monitored via GPS/VHF tracking and targeted trapping for 45 days postrelease.We found support for the "social subordinate" hypothesis, with smaller quolls dispersing further. Consistent with theories of senescence and the biology of our species, survival was marginally greater for females, and females regained losses in body mass in both years following release. In contrast, males recovered body condition in the first but not the second release as this coincided with breeding. Quolls that originated from the mainland sanctuary were on average heavier at release and, after accounting for weight, dispersed further. Synthesis and applications. Using theory to test outcomes of wildlife translocations can provide insights into patterns across taxa and under different conditions, enabling useful improvements to future fauna translocations. This allows for better predictions to be made about the likelihood of success from proposed translocations, changes to planning to improve outcomes (e.g., modifying sex ratios, individual selection and release cohort), and improved animal welfare as fewer animals are subjected to trials.

Eruptive dynamics are common in managed mammal populations.

Successful conservation management is often based on the principle that small or declining populations can recover if we identify and remove the factors that caused them to decline in the first place. But what form will that recovery take? Theory tells us that when a strong limiting factor is removed, a population should increase in size to where it becomes limited by some other factor. However, if the subsequent limitation involves feedbacks between the density of a consumer and its resource, there is potential for the consumer population to undergo substantial fluctuations in size that we would characterize as boom-bust or eruptive dynamics. We analysed long-term (7.6-29 yr) data documenting changes in the abundance of 169 populations of 20 mammal species released from a strong limiting factor (fox predation) in Australia. We show that many populations (44) exhibited eruptive dynamics, with exponential increase to a peak and subsequent population decline. Of 51 populations showing eruptive dynamics (the Australian populations plus seven translocated ungulate populations), the time taken for erupting populations to reach a peak before declining was related negatively to the intrinsic rate of population growth and positively to body mass, such that larger-bodied species with slow rates of population growth had a longer period of population increase before declining. Our results suggest that a substantial proportion of populations recovering after removal of a threatening process are likely to exhibit eruptive dynamics, and that managers of recovering or translocated populations should anticipate this outcome in conservation planning.

Finding food in a novel environment: The diet of a reintroduced endangered meso-predator to mainland Australia, with notes on foraging behaviour.

Translocated captive-bred predators are less skilled at hunting than wild-born predators and more prone to starvation post-release. Foraging in an unfamiliar environment presents many further risks to translocated animals. Knowledge of the diet and foraging behaviour of translocated animals is therefore an important consideration of reintroductions. We investigated the diet of the endangered meso-predator, the eastern quoll Dasyurus viverrinus. We also opportunistically observed foraging behaviour, enabling us to examine risks associated with foraging. Sixty captive-bred eastern quolls were reintroduced to an unfenced reserve on mainland Australia (where introduced predators are managed) over a two year period (2018, 2019). Quolls were supplementary fed macropod meat but were also able to forage freely. Dietary analysis of scats (n = 56) revealed that quolls ate macropods, small mammals, birds, invertebrates, fish, reptiles and frogs, with some between-year differences in the frequency of different diet categories. We also observed quolls hunting live prey. Quolls utilised supplementary feeding stations, indicating that this may be an important strategy during the establishment phase. Our study demonstrated that, in a novel environment, captive-bred quolls were able to locate food and hunt live prey. However, foraging was not without risks; with the ingestion of toxic substances and foraging in dangerous environments found to be potentially harmful. Knowledge of the diet of reintroduced fauna in natural landscapes is important for understanding foraging behaviour and evaluating habitat suitability for future translocations and management.

Modeling ecological traps for the control of feral pigs.

Ecological traps are habitat sinks that are preferred by dispersing animals but have higher mortality or reduced fecundity compared to source habitats. Theory suggests that if mortality rates are sufficiently high, then ecological traps can result in extinction. An ecological trap may be created when pest animals are controlled in one area, but not in another area of equal habitat quality, and when there is density-dependent immigration from the high-density uncontrolled area to the low-density controlled area. We used a logistic population model to explore how varying the proportion of habitat controlled, control mortality rate, and strength of density-dependent immigration for feral pigs could affect the long-term population abundance and time to extinction. Increasing control mortality, the proportion of habitat controlled and the strength of density-dependent immigration decreased abundance both within and outside the area controlled. At higher levels of these parameters, extinction was achieved for feral pigs. We extended the analysis with a more complex stochastic, interactive model of feral pig dynamics in the Australian rangelands to examine how the same variables as the logistic model affected long-term abundance in the controlled and uncontrolled area and time to extinction. Compared to the logistic model of feral pig dynamics, the stochastic interactive model predicted lower abundances and extinction at lower control mortalities and proportions of habitat controlled. To improve the realism of the stochastic interactive model, we substituted fixed mortality rates with a density-dependent control mortality function, empirically derived from helicopter shooting exercises in Australia. Compared to the stochastic interactive model with fixed mortality rates, the model with the density-dependent control mortality function did not predict as substantial decline in abundance in controlled or uncontrolled areas or extinction for any combination of variables. These models demonstrate that pest eradication is theoretically possible without the pest being controlled throughout its range because of density-dependent immigration into the area controlled. The stronger the density-dependent immigration, the better the overall control in controlled and uncontrolled habitat combined. However, the stronger the density-dependent immigration, the poorer the control in the area controlled. For feral pigs, incorporating environmental stochasticity improves the prospects for eradication, but adding a realistic density-dependent control function eliminates these prospects.

Unintended consequences of invasive predator control in an Australian forest: overabundant wallabies and vegetation change.

Over-abundance of native herbivores is a problem in many forests worldwide. The abundance of native macropod wallabies is extremely high at Booderee National Park (BNP) in south-eastern Australia. This has occurred because of the reduction of exotic predators through an intensive baiting program, coupled with the absence of other predators. The high density of wallabies at BNP may be inhibiting the recruitment of many plant species following fire-induced recruitment events. We experimentally examined the post-fire response of a range of plant species to browsing by wallabies in a forest heavily infested with the invasive species, bitou bush Chrysanthemoides monilifera. We recorded the abundance and size of a range of plant species in 18 unfenced (browsed) and 16 fenced (unbrowsed) plots. We found the abundance and size of bitou bush was suppressed in browsed plots compared to unbrowsed plots. Regenerating seedlings of the canopy or middle storey tree species Eucalyptus pilularis, Acacia implexa, Allocasuarina littoralis, Breynia oblongifolia and Banksia integrifolia were either smaller or fewer in number in grazed plots than treatment plots as were the vines Kennedia rubicunda, Glycine tabacina and Glycine clandestina. In contrast, the understorey fern, Pteridium esculentum increased in abundance in the browsed plots relative to unbrowsed plots probably because of reduced competition with more palatable angiosperms. Twelve months after plots were installed the community structure of the browsed and unbrowsed plots was significantly different (P = 0.023, Global R = 0.091). The relative abundance of C. monilifera and P. esculentum contributed most to the differences. We discuss the possible development of a low diversity bracken fern parkland in Booderee National Park through a trophic cascade, similar to that caused by overabundant deer in the northern hemisphere. We also discuss its implications for broad scale fox control in southern Australian forests.