At the Western Poultry Diseases Conference, held last March in Sacramento, the audience was served a bit of a difficult subject: The use of artificial neural networks. But what does it entail and what are the benefits for the poultry industry?
By Emmy Koeleman
Molecular diagnostic methods have evolved significantly over the years and together with our increased knowledge of DNA patterns of important poultry pathogens such as E. coli we are able to predict or test the virulence and pathogenicity even better. At the same time, the computational changes have brought growth to new technologies. One of them is artificial neural networks (ANNs).
An ANN is an information processing concept which is inspired from one of the most powerful and complex things known to mankind - the human brain. Most neural networks are software simulations run on conventional computers. The neural network is simply neurons (just like in the brain) joined together, with the output from one neuron becoming input to others until the final output is reached. The use of self-organising networks such as ANNs have been widely used in many scientific areas but presently under-utilised in the poultry industry. ANNs, like people, learn by example and learn from patterns of interactions, without requiring a prior knowledge of relations between the variables under investigation.
An ANN can thus be trained to - for example - predict plasma hormones and liver enzymes in broiler chickens or the pathogenicity of E. colibacteria. But an ANN can also be committed for breeder management, broiler breeder serological interpretation and hatchery management when sufficient data is available. Not only the variety of applications in animal research make ANNs an exciting area, livestock researchers also applaud the use of ANNs because these models do not require live animals for research, and thus may help in solving the ethical problems related to this.
At the recently held Western Poultry Diseases Conference, in Sacramento, USA, a significant amount of papers were dedicated to the use of ANNs in poultry research. For example, Felipe de Oliveira Salle from the Center for Diagnosis and Research in Avian Pathology in Brazil explained recent work on using ANNs to classify antimicrobial resistance from E. coli samples isolated from broilers. E. coli is a common pathogen in the poultry industry and ANNs are increasingly being used as a tool to measure non-linear relations among variables.
Studies on E. coli using ANNs deal essentially with the genetic identification bases of DNA promoters, bacterial growth, predictions of mutagenicity and others. Virulence mechanisms of E. coli isolates potentially pathogenic for broilers have been continuously studied but the classification of genes that are pathogenic and/or associated with pathogenicity is very complex. The purpose of the study by Salle and his team was to construct an ANN to predict the results of antimicrobial resistance of 246 isolates of E. colifrom poultry production.
The inputs chosen to construct the ANN were pathogenicity indices, lesions induced in one-day old chicks, characterisation of the genes associated with pathogenicity, biochemical behaviour and the origin and motility of the isolates. The outputs were resistance or sensitivity to 14 antibiotics. The research team showed that ANNs were capable of classifying bacterial resistance to the antibiotics studied. ANNs also showed to be an excellent tool to infer the behaviour of a bacterial isolate to a certain antibiotic.
Classification of pathogenicity
Similar work on E. coli has been performed by the same Brazilian research institute. Led by Carlos Tadeu Pippi Salle, this study specifically looked into the use of ANNs in predicting the pathogenicity of E. coli isolates from broilers. The virulence mechanisms of E. coli isolates potentially pathogenic for broilers have been continuously studied and it is supposed to be of multi-factorial cause.
Certain properties are primarily associated with these strains and the most frequently mentioned include adhesion ability (pap and fel), production of colicins (cva), presence of aerobactin (iut), serum resistance (iss), temperature sensitive hemagglutin (tsh) and the presence of some capsular antigens (kps). However, what sets apart virulent and avirulent strains remains a problem in diagnosis and consequently, in the decision making by the field veterinarian. Conventional procedures for determining the pathogenicity of E. coli in which animals are inoculated are time consuming and expensive.
In this study, an ANN was set up to predict the pathogenicity of E. coli. A total of 293 E. coliisolated were analysed (76 litter samples, 159 cellulitis lesions and 58 colispeticemic broiler organs). At the conference, Salle presented three neural networks, in which the input layer included information on the presence of papC, felA, cvaC, iutA, iss, tsh and kpsII genes, isolates motility and origin (colisepticemia, cellulitis and litter). The output layer was formed by the “pathogenicity index” (PI) (Souza, 2006). In the first network, PI varies from 0 to 10. In the second one, PI’s were grouped in three classes or categories: pathogenic/low pathogenicity with PI between 0 and 3.99; intermediate pathogenicity refers to PI’s between 4.00 and 6.99 and high pathogenicity with PI ranging from 7.00 to 10.
Finally, a third network was constructed, in which daily situations experienced by field veterinarians were represented and bacteria were considered as of low pathogenicity or intermediate pathogenicity. According to the researchers, the characteristics of this model allow the classification of isolates pathogenicity in broiler houses with a good degree of reliability, considering the sensitivity and the specificity. It is important however to have a sufficient amount of isolates per category to have a high sensitivity rate. Other factors that affect the sensitivity may include high diversity of genetic profiles or the fact that isolates with the same profile presented different PIs.
Analysing lymphoid depletion
Moraes et al used the ANN technology to analyse the follicular lymphoid depletion in the bursa of Fabricius and presented some fresh results at the conference. The bursa of Fabricius is involved in many immunosuppressive diseases and has an important role in the characterisation, diagnosis and monitoring of such conditions. Currently, the evaluation of bursal lymphoid depletion depends on a subjective histological evaluation and therefore susceptible to errors. ANNs can be used to minimise errors in the conventional optical technology.
Fifty bursa of Fabricius samples were examined using conventional histological examination. The samples were scored three times by a single examiner with an interval of one day between the examinations. Ten slides of each score were randomly selected for image processing and essayed according to the following steps: change to grey scale, select the follicular area, all structures around the selected area were rubbed out, the number of pixels for grey scale of each image was estimated and the image histograms and table were constructed. These data provided the input for the ANN and used to train the ANN to obtain a digital score which is compared with the score obtained by optical classification.
This study showed that the ANN was able to make a comparable classification of digital and optical scores and was therefore – together with the image analysis – a helpful tool in the diagnosis of follicular lymphoid depletion in the bursa of Fabricius. One advantage of this new methodology is that it does not need special histological techniques, is more accurate and more reliable. In addition, it does not require sophisticated training to implement the referred protocol. However, the presence of a trained pathologist is essential for the different diagnosis of other conditions.
Objective and more accurate
In the poultry industry, the first models that used ANNs were published by Zhang et al in 1996, in the animal nutrition field. But it is only since a few years that the interest for using ANNs has grown. It is a great tool for recognising patterns in data and accurately predicting the performance and one of the great benefits is that researchers can predict pathogenicity of certain bacteria or viruses without the use of animals.
The information gained from using ANNs can provide field veterinarians with reports including the degree of pathogenicity of the isolates so that the decisions can be made more objectively and more accurately. In the near future, more research papers will probably surface, presenting more insights into how this savvy tool can be used for the benefit of the poultry industry.