ABSTRACT
In long-term individual-based field studies, several parameters need to be assessed repeatedly to fully understand the potential fitness effects on individuals. Often studies only evaluate capture stress that appears in the immediate weeks or breeding season and even long-term studies fail to evaluate the long-term effects of their capture procedures. We investigated effects of long-term repeated capture and handling of individuals in a large semi-aquatic rodent using more than 20 years of monitoring data from a beaver population in Norway. To investigate the effects, we corrected for ecological factors and analysed the importance of total capture and handling events, years of monitoring and deployment of telemetry devices on measures related to body condition, reproduction and survival of individual beavers. Body mass of dominant individuals decreased considerably with number of capture events (107 g per capture), but we found no statistically clear short or long-term effects of capture and handling on survival or other body condition indices. Annual litter size decreased with increasing number of captures among older individuals. Number of captures furthermore negatively affected reproduction in the beginning of the monitoring, but the effect decreased over the years, indicating habituation to repeated capture and handling. By assessing potential impacts on several fitness-related parameters at multiple times, we can secure the welfare of wild animal populations when planning and executing future conservation studies as well as ensure ecologically reliable research data.
Subject(s)
Body Weight/physiology , Reproduction/physiology , Rodentia/physiology , Animals , Animals, Wild , Norway , SeasonsABSTRACT
A study is presented on how well possible drug-molecules can be predicted with respect to their function and binding to a selection of neuro-receptors by the use of artificial neural networks. The ligands investigated in this study are chosen to be corresponding to the G protein-coupled receptors µ-opioid, serotonin 2B (5-HT2B) and metabotropic glutamate D5. They are selected due to the availability of pharmacological drug-molecule binding data for these receptors. Feedback and deep belief artificial neural network architectures (NNs) were chosen to perform the task of aiding drugdesign. This is done by training on structural features, selected using a "minimum redundancy, maximum relevance"-test, and testing for successful prediction of categorized binding strength. An extensive comparison of the neural network performances was made in order to select the optimal architecture. Deep belief networks, trained with greedy learning algorithms, showed superior performance in prediction over the simple feedback NNs. The best networks obtained scores of more than 90 % accuracy in predicting the degree of binding drug molecules to the mentioned receptors and with a maximal Matthew`s coefficient of 0.925. The performance of 8 category networks (8 output classes for binding strength) obtained a prediction accuracy of above 60 %. After training the networks, tests were done on how well the systems could be used as an aid in designing candidate drug molecules. Specifically, it was shown how a selection of chemical characteristics could give the lowest observed IC50 values, meaning largest bio-effect pr. nM substance, around 0.03-0.06 nM. These ligand characteristics could be total number of atoms, their types etc. In conclusion, deep belief networks trained on drug-molecule structures were demonstrated as powerful computational tools, able to aid in drug-design in a fast and cheap fashion, compared to conventional pharmacological techniques.