4 ms·

Not with recent results in Nature I believe reporting glaring mispredictions. Lots of promotion notwithstanding, AlphaFold may not be usable yet.

by gone35 4y ago

dekhn 4y ago
Do you have links to that? have not seen those.
- gone35 4y ago
  Yes. Posted here (and with some important qualifications): https://news.ycombinator.com/item?id=32948345 https://news.ycombinator.com/item?id=32948345
gone35 4y ago
Not Nature but Molecular Systems Biology [2], apologies, in the context of reverse docking for antibiotics discovery: "We confirm extensive promiscuity, but find that the average area under the receiver operating characteristic curve (auROC) is 0.48, indicating weak model performance." Derek Lowe had a post about this earlier this month [1] (which includes important qualifications I failed to omit). [1] "Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discovery", Molecular Systems Biology 2022 18:e11081 https://doi.org/10.15252/msb.202211081 https://doi.org/10.15252/msb.202211081 [2] "Not AlphaFold's Fault", September 7 2022, https://www.science.org/content/blog-post/not-alphafold-s-fault https://www.science.org/content/blog-post/not-alphafold-s-fa...
- flobosg 4y ago
  This is already known. Structural models are usually not well suited for protein-ligand docking, where proper pocket modelling and accurate side chain positioning are key for the determination of true positive hits.
dekhn 4y ago
Sorry, you don't know what you're talking about. Nobody is truly claiming these models can be used directly for drug discovery (well, some people claim that; but they're wrong). We already knew that, though- the same problem exists with high quality crystal structures. What would be more interesting is if we did a whole bunch of crystal or cryo-EM structures far from what we've previously determined and demonstrate whether alphafold could make out-of-class predictions for them.

Not with recent results in Nature I believe reporting glaring mispredictions. Lots of promotion notwithstanding, AlphaFold may not be usable yet.

by gone35 4y ago

Not with recent results in Nature I believe reporting glaring mispredictions. Lots of promotion notwithstanding, AlphaFold may not be usable yet.

dekhn 4y ago

Do you have links to that? have not seen those.

gone35 4y ago
Yes. Posted here (and with some important qualifications): https://news.ycombinator.com/item?id=32948345 https://news.ycombinator.com/item?id=32948345

gone35 4y ago

Not Nature but Molecular Systems Biology [2], apologies, in the context of reverse docking for antibiotics discovery: "We confirm extensive promiscuity, but find that the average area under the receiver operating characteristic curve (auROC) is 0.48, indicating weak model performance." Derek Lowe had a post about this earlier this month [1] (which includes important qualifications I failed to omit). [1] "Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discovery", Molecular Systems Biology 2022 18:e11081 https://doi.org/10.15252/msb.202211081 https://doi.org/10.15252/msb.202211081 [2] "Not AlphaFold's Fault", September 7 2022, https://www.science.org/content/blog-post/not-alphafold-s-fault https://www.science.org/content/blog-post/not-alphafold-s-fa...

flobosg 4y ago
This is already known. Structural models are usually not well suited for protein-ligand docking, where proper pocket modelling and accurate side chain positioning are key for the determination of true positive hits.

dekhn 4y ago

Sorry, you don't know what you're talking about. Nobody is truly claiming these models can be used directly for drug discovery (well, some people claim that; but they're wrong). We already knew that, though- the same problem exists with high quality crystal structures. What would be more interesting is if we did a whole bunch of crystal or cryo-EM structures far from what we've previously determined and demonstrate whether alphafold could make out-of-class predictions for them.