The scientists claim that this configuration allows Evo to “connect nucleotide-level patterns to kilobase-scale genomic context.” In simpler terms, when provided with a substantial segment of genomic DNA, Evo can understand it as an LLM would handle a query and generate an output that is biologically relevant to that understanding.
The scientists hypothesized that, due to its training on bacterial genomes, they could use an identified gene as a prompt, expecting Evo to generate an output that encompasses regions encoding proteins with similar functions. The central inquiry is whether it would merely reproduce the sequences of already known proteins or produce more unexpected outputs.
New proteins
To initiate the evaluation of the system, the researchers provided it with gene fragments of known proteins and assessed whether Evo could complete them. For instance, when supplied with 30 percent of the sequence of a known protein’s gene, Evo successfully generated 85 percent of the remaining sequence. When given 80 percent of the sequence, it could reconstruct the entire missing sequence. Furthermore, when one gene was removed from a functional cluster, Evo was also able to accurately identify and restore the absent gene.
The extensive training data also guaranteed that Evo accurately recognized the most crucial areas of the protein. If alterations were made to the sequence, they generally occurred in regions where variability is acceptable. This means that its training allowed the system to assimilate the principles of evolutionary constraints on modifications in known genes.
Consequently, the researchers aimed to explore what transpired when Evo was instructed to generate something novel. They utilized bacterial toxins, which are usually encoded alongside an anti-toxin that prevents the cell from self-destruction upon activation of the genes. Many such examples exist, and they tend to undergo rapid evolution due to the ongoing arms race between bacteria and their adversaries. Thus, the team created a toxin that was only slightly related to existing ones and lacked a known antitoxin, then provided its sequence to Evo as a prompt. This time, they filtered out any responses resembling known antitoxin genes.
