2 ms·

TFA goes Gehman's colleague Melanie Prentice, also with the Hakai Institute and UBC, had begun reviewing the sequencing analysis. Scrolling through data on h

TFA goes Gehman's colleague Melanie Prentice, also with the Hakai Institute and UBC, had begun reviewing the sequencing analysis. Scrolling through data on her laptop, an unmistakable pattern jumped out. Of the hundreds of microorganisms in the samples, enormous quantities of bacteria from the genus Vibrio seemed present in sick stars. The pattern was so strong, Prentice tells me, that she assumed she'd bungled the analysis. I'm not familiar with sequencing much, but would feeding these to models similar to or purpose built like AlphaFold shorten the research time drastically [0]? If not, what other recent technological advancements do you reckon that'd help here? Thanks. [0] Trying to make sense of the hype around some of these: https://fortune.com/2025/07/06/deepmind-isomorphic-labs-cure-all-diseases-ai-now-first-human-trials/ https://fortune.com/2025/07/06/deepmind-isomorphic-labs-cure...

greazy 1y ago
No. Drug discovery is a different beast. It's tricky to identify the caustive agent from sequencing data. Sometimes it sticks out like a sore thumb. Eg I recently worked on samples from a patient who died, it was very easy to identify the cause because it was in the same at massive quantities (reads). But usually, it's trawling through a hundred hits across all domains of life deciding is it consensual, contamination, pathogenic or database error. It's usually inconclusive. Oddly enough, llm bots seem to be quiet good at this task without even fine tuning.
steve_adams_86 1y ago
Do you mean it could be faster to analyze sequences using models rather than scrolling through them on a laptop? Yes, we could get some broad strokes of what's present by using analytical tools, and we do. A lot of our researchers use R and Python to write simple tools to find patterns and perform simple aggregations. We're in the process of discovering how AI can assist us, and we've had plenty of success in creating models to aid with processing and detecting things like mussel and kelp beds in aerial imagery (drone, plane, and satellite), but I suspect in this case the data can't actually tell the full story to begin with so what we're actually looking for is more like clues, not answers. When answers are present, we tend to be simply presenting findings like which organisms were present and, to some extent, their apparent population density. This is really easy to do without AI. Some models might find interesting patterns, but it's hard to tell which ones or why, or how we'd train that ourselves. It's an interesting question. I work on the genomics team so it's something I should be actively considering. The work is certainly more manual than it needs to be, so there's work to be done here for sure.

greazy 1y ago

No. Drug discovery is a different beast. It's tricky to identify the caustive agent from sequencing data. Sometimes it sticks out like a sore thumb. Eg I recently worked on samples from a patient who died, it was very easy to identify the cause because it was in the same at massive quantities (reads). But usually, it's trawling through a hundred hits across all domains of life deciding is it consensual, contamination, pathogenic or database error. It's usually inconclusive. Oddly enough, llm bots seem to be quiet good at this task without even fine tuning.

steve_adams_86 1y ago

Do you mean it could be faster to analyze sequences using models rather than scrolling through them on a laptop? Yes, we could get some broad strokes of what's present by using analytical tools, and we do. A lot of our researchers use R and Python to write simple tools to find patterns and perform simple aggregations. We're in the process of discovering how AI can assist us, and we've had plenty of success in creating models to aid with processing and detecting things like mussel and kelp beds in aerial imagery (drone, plane, and satellite), but I suspect in this case the data can't actually tell the full story to begin with so what we're actually looking for is more like clues, not answers. When answers are present, we tend to be simply presenting findings like which organisms were present and, to some extent, their apparent population density. This is really easy to do without AI. Some models might find interesting patterns, but it's hard to tell which ones or why, or how we'd train that ourselves. It's an interesting question. I work on the genomics team so it's something I should be actively considering. The work is certainly more manual than it needs to be, so there's work to be done here for sure.