AliFilter: a Machine Learning Approach to Alignment Filtering.

INTRODUCTION: Multiple sequence alignments are a crucial step in many bioinformatic and computational biology analyses, from protein structure and function prediction to the inference of phylogenetic trees. However, highly divergent sequence alignments often contain a significant amount of noise. Reducing noise is normally achieved by filtering the alignment to remove columns that are poorly aligned or offer minimal useful information; either automatically using various software tools, or through manual inspection. Manual approaches are labour-intensive and less reproducible, but can utilise the researcher's specialist knowledge, rather than relying on filtering criteria that might not be adequate for each alignment. MATERIALS AND METHODS: AliFilter bridges these two approaches to alignment curation, using machine learning to automate manual alignment filtering. AliFilter uses a supervised learning approach to create a model from a small number of manually annotated alignments, then applies this model to reproduce the manual annotation on different datasets. Users can employ the program with a default model or create customised models for individual datasets or filtering criteria. RESULTS AND DISCUSSION: AliFilter accurately reproduces the results of manual annotation (98% accuracy), while being resilient to mistakes in the training data. In a typical phylogenomic workflow, AliFilter reduced the runtime by 35%, whilst retaining results that were almost identical to the full alignment, unlike other filtering tools we tested. AliFilter is free and open-source software; it is written in C# and distributed under a GPLv3 licence from https://github.com/arklumpus/AliFilter, where both the source code and standalone executables for Windows, macOS and Linux are available.