3 ms·

They mention AI used to weed out molecules but don't mention how it was used. I'd be interested in know did they use pre-trained models or train one from the g

They mention AI used to weed out molecules but don't mention how it was used. I'd be interested in know did they use pre-trained models or train one from the ground up.

lithos 2y ago
https://www.sciencealert.com/ai-experiment-generated-40-000-hypothetical-bioweapons-in-6-hours-scientists-warn https://www.sciencealert.com/ai-experiment-generated-40-000-... For another example, and it actually calls out the name of the program.
jqpabc123 2y ago
I would expect some custom training was needed in order to address this highly specialized use case. There are 2 primary uses cases emerging where LLMs are proving advantageous. 1) Summarization --- where the results can be nebulous and it don't really much matter. For example, casual web search. 2) Research based on trial and error where the nebulous results will be subjected to thorough verification --- i.e. this case. In any case, the statistical results can be nebulous and using them without verification is a recipe for disaster. For example: https://www.lawnext.com/2025/02/federal-judge-sanctions-morgan-morgan-attorneys-for-ai-generated-fake-cases-in-court-filing.html https://www.lawnext.com/2025/02/federal-judge-sanctions-morg...
- bitwize 2y ago
  I don't think it's a matter of firing up an LLM and saying "Hey LLM, please find me a molecule that acts on human GLP-1 receptors but does not have the negative side effects of semaglutide." And it's all like "Certainly! Here are some candidate GLP-1 agonists that blah blah blah..." It's probably more like AlphaFold -- a statistical model of molecular structure and action, not a language model.
xhevahir 2y ago
It looks like it's this: https://github.com/Svensson-Lab/pro-hormone-predictor https://github.com/Svensson-Lab/pro-hormone-predictor Annoying that journalists call everything "AI" these days.

They mention AI used to weed out molecules but don't mention how it was used. I'd be interested in know did they use pre-trained models or train one from the ground up.

lithos 2y ago

https://www.sciencealert.com/ai-experiment-generated-40-000-hypothetical-bioweapons-in-6-hours-scientists-warn https://www.sciencealert.com/ai-experiment-generated-40-000-... For another example, and it actually calls out the name of the program.

jqpabc123 2y ago

I would expect some custom training was needed in order to address this highly specialized use case. There are 2 primary uses cases emerging where LLMs are proving advantageous. 1) Summarization --- where the results can be nebulous and it don't really much matter. For example, casual web search. 2) Research based on trial and error where the nebulous results will be subjected to thorough verification --- i.e. this case. In any case, the statistical results can be nebulous and using them without verification is a recipe for disaster. For example: https://www.lawnext.com/2025/02/federal-judge-sanctions-morgan-morgan-attorneys-for-ai-generated-fake-cases-in-court-filing.html https://www.lawnext.com/2025/02/federal-judge-sanctions-morg...

bitwize 2y ago
I don't think it's a matter of firing up an LLM and saying "Hey LLM, please find me a molecule that acts on human GLP-1 receptors but does not have the negative side effects of semaglutide." And it's all like "Certainly! Here are some candidate GLP-1 agonists that blah blah blah..." It's probably more like AlphaFold -- a statistical model of molecular structure and action, not a language model.

xhevahir 2y ago

It looks like it's this: https://github.com/Svensson-Lab/pro-hormone-predictor https://github.com/Svensson-Lab/pro-hormone-predictor Annoying that journalists call everything "AI" these days.