Genome analysis tools and the future of AI

Analysing a genome follows a series of steps. Sequencers give either short reads (about 100-150 bases) or long reads (thousands of bases).

• Check and clean: reads are first quality-checked and trimmed of poor-quality ends, with tools like FastQC and Trimmomatic.
• Align or assemble: if a reference genome already exists, reads are aligned (matched) to it (re-sequencing). If not, they are assembled from scratch (de novo) by finding where they overlap, joining them into continuous stretches called contigs. Tools: BLAST and BLAT (searching), CLUSTAL, MUSCLE and T-Coffee (alignment), SAMtools and Picard (handling files).

• Annotate: find every feature in the assembled genome: genes, the junctions between exons and introns, repeats (transposons, satellites, low-complexity regions) and pseudogenes (broken copies of genes). Genes can be predicted from the sequence alone (ab initio), by similarity to known genes, or by combining both. Gene Ontology describes each gene product by the biological process it takes part in, the part of the cell where it works, and its molecular function. Cufflinks predicts transcripts.
• Interpret: for example, phylogenetic analysis (with PHYLIP or PAUP) works out how organisms are related through evolution and draws it as a tree (a cladogram).

Artificial intelligence is already powerful: programs have beaten experts at games (Libratus at poker, AlphaGo at Go) and help diagnose eye diseases; machine learning is used to identify genetic variants. But making AI truly useful for biology still needs time and teamwork among biologists, computer scientists and statisticians. It has not replaced the experts.