4 ms·

Pretty well done - especially the experimental validation. The main conclusions are accurate, but there is some BS around the periphery. A few notes below. -

Pretty well done - especially the experimental validation. The main conclusions are accurate, but there is some BS around the periphery. A few notes below. - The secret sauce here was the large, high quality experimental dataset they generated. - Chemprop is meh. - Is a tanimoto similarity < 0.3 really dissimilar? Depends on the chemical fingerprint parameters they are using in RDkit. For ECFP4 the expected similarity of two random molecules is about 0.1. A value of 0.3 about the cutoff you would use for a similarity based virtual screening with ECFP4 fingerprints. - The cheminformatic filtering steps that they did on their hits was well done. - The experimental validation of the target was well done. A good illustration of all the preclinical work it takes to validate a molecule and mechanism on the way to an IND. - They got very lucky. Most of the time your compound hits are not all that potent and require extensive med chem optimization. That was not the case here.

rdedev 3y ago
Thanks for the info. I'm been delving slowly into this field as part of the lab I'm working with but I come from a purely CS background. So I'm not sure how much chemistry I should be knowing to build decent ML models. For now I've been getting away with xgboost and rdkit. There are so many quirks within this field that I'm not sure how one can apply ML to it. Tautomers being one of them
- pfisherman 3y ago
  Alex Tropsha has some pretty solid articles on best practices.
- gilleain 3y ago
  Another useful resource might be this blog : https://depth-first.com/ https://depth-first.com/ Richard Apodaca has some very clear explanations of topics in the field.
RandomLensman 3y ago
Yes, and also a good counterexample to some of the "thinkism" touted in AI, i.e., this is a result based on carefully curated data not just some machine just "thinking up" something.
happytiger 3y ago
I’ll take luck when it comes to new antibiotics. Do you think we will come up with AIs that can model potency? How can we further improve leveraging AI beyond discovery?
- tomrod 3y ago
  > How can we further improve leveraging AI beyond discovery? Poorly. AI does not extrapolate. It can explore defined interpolation spaces with an outcome artifact in input areas where there is little data, but typically cannot extrapolate outside of convex hulls, provided classes, or other hard boundaries very well. In effect, you have to be crafty and clever about the space an AI is to search within.
  - happytiger 3y ago
    Thank you for this insight — it’s very enlightening for someone outside of this professional area.
- pfisherman 3y ago
  They already do. QSAR has been a thing for a long time. The limiting factor is almost always the data. The publicly available SAR data in chembl / pubchem is trash, and why would you spend time and money designing and generating a good training dataset for a model when you could just be optimizing the compound? Keep in mind that the dataset generated during optimization and the one you would prefer to train a model are quite different. Also, in practice optimizing a lead compound is about more than just potency. You are also worried about things like toxicity, bioavailability, off target effects (selectivity), synthesis routes / yields (manufacturability), etc. And functional readouts (the assays you run) are not consistent across diseases / indications / targets / compounds. That being said I am almost certain someone has used Bayesian optimization for an assist.
  - dekhn 3y ago
    IMHO comparing QSAR to modern deep learning methods is sort of like comparing linear regression to modern deep learning methods: a technique that is wholly unsuitable for the sort of problems that deep learning is good at. While I generally agree the data quality is poor, let's be clear: by exploting the data in PDB and Uniprot, Deepmind "solved protein structure prediction". There is a lot more value in the data that's out there, just mostly underutilized.
hedora 3y ago
The pharma industry has been using a combination of computational techniques and high throughput screening for decades. I really don't see what's new here (except that they found a promising compound). The part of the abstract that I have access to doesn't say how far abaucin got in the compound screening pipeline. Has it been shown effective in vitro or in vivo? What side effects have been screened for? Does it reach the infection site? Does it survive the digestive tract? Does it interact with common foods? Here's a wikipedia article about how high throughput drug screening works. The inputs to these processes are traditionally computationally generated: https://en.wikipedia.org/wiki/High-throughput_screening https://en.wikipedia.org/wiki/High-throughput_screening Choice quote from wikipedia: The term uHTS or ultra-high-throughput screening refers (circa 2008) to screening in excess of 100,000 compounds per day. edit: Also, what percentage of compounds generated with this computational technique make it to each stage of the screening pipeline, and how does that compare to existing computational techniques?
- pfisherman 3y ago
  I’ll have to double check, but I think they tested in an animal (mouse) model. Not sure how far the compound got in the pipeline got for the original indication, but since it is from the repurposing hub they might be able to jump to phase 2.

rdedev 3y ago

Thanks for the info. I'm been delving slowly into this field as part of the lab I'm working with but I come from a purely CS background. So I'm not sure how much chemistry I should be knowing to build decent ML models. For now I've been getting away with xgboost and rdkit. There are so many quirks within this field that I'm not sure how one can apply ML to it. Tautomers being one of them

pfisherman 3y ago
Alex Tropsha has some pretty solid articles on best practices.
gilleain 3y ago
Another useful resource might be this blog : https://depth-first.com/ https://depth-first.com/ Richard Apodaca has some very clear explanations of topics in the field.

RandomLensman 3y ago

Yes, and also a good counterexample to some of the "thinkism" touted in AI, i.e., this is a result based on carefully curated data not just some machine just "thinking up" something.

happytiger 3y ago

I’ll take luck when it comes to new antibiotics. Do you think we will come up with AIs that can model potency? How can we further improve leveraging AI beyond discovery?

tomrod 3y ago
> How can we further improve leveraging AI beyond discovery? Poorly. AI does not extrapolate. It can explore defined interpolation spaces with an outcome artifact in input areas where there is little data, but typically cannot extrapolate outside of convex hulls, provided classes, or other hard boundaries very well. In effect, you have to be crafty and clever about the space an AI is to search within.
- happytiger 3y ago
  Thank you for this insight — it’s very enlightening for someone outside of this professional area.
pfisherman 3y ago
They already do. QSAR has been a thing for a long time. The limiting factor is almost always the data. The publicly available SAR data in chembl / pubchem is trash, and why would you spend time and money designing and generating a good training dataset for a model when you could just be optimizing the compound? Keep in mind that the dataset generated during optimization and the one you would prefer to train a model are quite different. Also, in practice optimizing a lead compound is about more than just potency. You are also worried about things like toxicity, bioavailability, off target effects (selectivity), synthesis routes / yields (manufacturability), etc. And functional readouts (the assays you run) are not consistent across diseases / indications / targets / compounds. That being said I am almost certain someone has used Bayesian optimization for an assist.
- dekhn 3y ago
  IMHO comparing QSAR to modern deep learning methods is sort of like comparing linear regression to modern deep learning methods: a technique that is wholly unsuitable for the sort of problems that deep learning is good at. While I generally agree the data quality is poor, let's be clear: by exploting the data in PDB and Uniprot, Deepmind "solved protein structure prediction". There is a lot more value in the data that's out there, just mostly underutilized.

hedora 3y ago

The pharma industry has been using a combination of computational techniques and high throughput screening for decades. I really don't see what's new here (except that they found a promising compound). The part of the abstract that I have access to doesn't say how far abaucin got in the compound screening pipeline. Has it been shown effective in vitro or in vivo? What side effects have been screened for? Does it reach the infection site? Does it survive the digestive tract? Does it interact with common foods? Here's a wikipedia article about how high throughput drug screening works. The inputs to these processes are traditionally computationally generated: https://en.wikipedia.org/wiki/High-throughput_screening https://en.wikipedia.org/wiki/High-throughput_screening Choice quote from wikipedia: The term uHTS or ultra-high-throughput screening refers (circa 2008) to screening in excess of 100,000 compounds per day. edit: Also, what percentage of compounds generated with this computational technique make it to each stage of the screening pipeline, and how does that compare to existing computational techniques?

pfisherman 3y ago
I’ll have to double check, but I think they tested in an animal (mouse) model. Not sure how far the compound got in the pipeline got for the original indication, but since it is from the repurposing hub they might be able to jump to phase 2.