Machine learning and statistics shape a novel path in archaeal promoter annotation

Abstract

Background Archaea are a vast and unexplored domain. Bioinformatic techniques might enlighten the path to a higher quality genome annotation in varied organisms. Promoter sequences of archaea have the action of a plethora of proteins upon it. The conservation found in a structural level of the binding site of proteins such as TBP, TFB, and TFE aids RNAP-DNA stabilization and makes the archaeal promoter prone to be explored by statistical and machine learning techniques. Results and discussions In this study, experimentally verified promoter sequences of the organisms Haloferax volcanii, Sulfolobus solfataricus, and Thermococcus kodakarensis were converted into DNA duplex stability attributes (i.e. numerical variables) and were classified through Artificial Neural Networks and an in-house statistical method of classification, being tested with three forms of controls. The recognition of these promoters enabled its use to validate unannotated promoter sequences in other organisms. As a result, the binding site of basal transcription factors was located through a DNA duplex stability codification. Additionally, the classification presented satisfactory results (above 90%) among varied levels of control. Concluding remarks The classification models were employed to perform genomic annotation into the archaea Aciduliprofundum boonei and Thermofilum pendens, from which potential promoters have been identified and uploaded into public repositories.

Explainable Artificial Intelligence as a reliable annotator of archaeal promoter regions

“…First, we posit for the balance of the classi cation achieved by our train/test (84.55% precision) and the external sequences (81%), which is further con rmed by the likeliness of the DDS pro le of these sequences. Next, we argue upon the fact that better performance has been obtained classifying archaea (Martinez et al, 2022), eukarya (Oubounyt et al, 2019), and bacteria (de Avila e Silva et al, 2011). The authors mainly employed Arti cial Neural Networks for classi cation, which poses as a classi cation rationale that is mathematically more complex than the SVMs we employed (Pisner and Schnyder, 2020).…”

“…Promoters were derived from the transcripts of the organisms. In addition, 405 promoters in-silico predicted (Martinez et al, 2022) of A. boonei and T. pendens were considered. In brief, Martinez et al (2002) employed machine learning and statistics to validate promoter sequences of unannotated archaea.…”

“…A total of 3935 sequences (length = 100) were used as the negative sequences in the classi cation model. For that, we shu ed versions of the sequence depicted in the train/test dataset, as it has been previously suggested by Martinez et al (2022) because the shu ed version of the original sequence is the underperformer among others. Hence, to stress our model to the most, we opted to use this form of acquiring control sequences.…”

Explainable artificial intelligence as a reliable annotator of archaeal promoter regions

“…It has been previously benchmarked by Ref. 19 that among non-coding and shuffled promoter sequences, the shuffled version of a control dataset yielded in the worst classification. Thus, we opted to use this method of obtaining negative sequences to stress our classification model.…”

“…Many promoter predictors use ML as their form of classification. In this task, different algorithms have been proposed, such as Artificial Neural Networks 13 , 15 , 19 , Support Vector Machines (SVM) 11 , Recurrent Neural Networks 16 , among others. However, due to the mathematical complexity, many of these tools will function as black-box classifiers, where one just knows the output associated with given input features.…”