Combinatorial Long-Short Term Memory– Group Interaction in Contribution Method for Estimation of Lethal Dose of Pesticides

Abstract

     Humans  are  continuously  exposed  to a variety  of  chemicals , including  pesticides , many of which  are  potentially  toxic and  have carcinogenic  effects . Determining  the  human  toxicity  of  chemicals  remained  a  challenge  due  to  the large resources required  to  evaluate  a  chemical  in  vivo. In  this  study , a  hybrid  model  that combined  the  long-short  term memory  (LSTM)  and  Group  Interaction  Contribution  (GIC) was  used  to  predict  the  LD₅₀  toxicity  in  rats  with chemical  pesticides , which  was  developed  using  a database  relatively  large  consisting  of  303  pesticides  belonging  to different  chemical  groups . The  architecture  of  LSTM-MLR  hybrid  model  was  carefully  selected  by  testing  different number  of  hidden  neurons  and  different  number of  training  iterations  in  order to  avoid  the  model  over  fitting  . The parameters  selection  process  revealed  that  the  model  should  be  developed  using  300  hidden  neurons  and  trained with 1000  iterations . The developed  model  with  best  selected  parameters  achieved  a very  interesting  results  and  a high  accuracy  in  the  prediction  of  LD₅₀  in  testing  phases  with  0.8888 , 0.1312 , 0.3622 , 0.2926 , 20.64 , and  39.99, for  R² , MSE , RMSE , MAE , RAE,  and  RRSE , respectively.