Lerna: transformer architectures for configuring error correction tools for short- and long-read genome sequencing

Table 3 Lerna evaluated on 7 Illumina (Table 2) short read datasets

Dataset	Read length	Exhaustive search		Lerna (Char)		Lerna (word)		Athena (RNN)
	Base pairs bp	Selected k	Alignment rate (%)	Selected k	Alignment rate (%)	Selected k	Alignment rate (%)	Selected k	Alignment rate (%)
D1	36	17	98.95	Same as exhaustive search		Same as exhaustive search		Same as exhaustive search
D2	47	15	61.42	Same as exhaustive search		19	61.27%	k=17	61.15%
D3	36	15	80.44	k=17	80.39%	Same as exhaustive search		k=17	80.39%
D4	75	17	93.95	k=15	93.72%	Same as exhaustive search		Same as exhaustive search
D5	100	17	92.15	Same as Exhaustive Search		Same as exhaustive search		k=25	92.09%
D6	250	25	86.16	Same as Exhaustive Search		Same as exhaustive search		k=17	85.63%
D7	101	17	40.53	k=15	40.24%	k=15	40.24%	Same as exhaustive search

We test both the Transformer character- and word-level LMs. We observed that the best performance was attained by using the Transformer word-level LM with word length 4 (|w|=4). For the Transformer word-level LM (|w| = 4), Lerna finds either the optimal value or a value with an alignment rate within 0.31% of the theoretical best, consistent with the reported results by Lighter (Figure 5 in [18]). These slightly sub-optimal configurations are an artefact of sub-sampling

ISSN: 1471-2105