In defence of experimental science, may I also suggest that the models AlphaFold2 and other predictors worked on were derived from sequences for which they knew a well-defined structure is possible. Much of the work we do is exactly on taking unwieldy sequences with disordered elements, multiple domains etc, and optimising these into constructs that produce well-diffracting crystals. Though NMR and CryoEM do not need crystals, construct optimization is often just as important for them too. Perhaps AlphaFold3 will be able to predict structures from gene sequences without that implicit optimisation, but that's not the case just yet.
Best John ________________________________ From: CCP4 bulletin board <CCP4BB@JISCMAIL.AC.UK> on behalf of Jan Löwe <j...@mrc-lmb.cam.ac.uk> Sent: 04 December 2020 10:33 To: CCP4BB@JISCMAIL.AC.UK <CCP4BB@JISCMAIL.AC.UK> Subject: Re: [ccp4bb] AlphaFold: more thinking and less pipetting (?) AlphaFold2 and its performance are indeed a true breakthrough of potentially seismic proportion - and will accelerate much of what we are trying to achieve. But I feel it is also important to point out that the method principally relies on experimental data: lots of protein sequences that through evolution have evolved to generate very similar folds and structures. Machine learning is used to transform the resulting (experiment-derived) co-evolution matrices (can be thought of as images for ML training) into distances between atoms. The distances are then used to generate models (minimiser in Alpha Fold 1, something ML in 2, as far as I could figure out). Note that contacts between proteins can also generate evolutionary couplings (as already used by the pioneers of the field, such as Debora Marks and Chris Sander), so something like AlphaFold will be able to make inroads there as well and that application might well have a greater impact. This leaves three important goals remaining if I may add: 1) a way to do this for any single sequence, without alignment, or indeed any folding polymer (for when we will be able to make coded polymers that are not made from amino acids). 2) a method to obtain accurate numbers, such as binding energies and rates. 3) the inverse: predicting sequences that have a particular function/fold. I would like to suggest that all three will require looking at how we can use more of the physics of the problem, but might well involve more machine learning. Jan On 04/12/2020 00:49, Paul Adams wrote: I agree completely Tom. Having been recently involved in some efforts to identify interesting compounds against SARS-CoV-2, I can say that the current AI/ML methods for docking/predicting small molecule binding have very very low success rates (I’m being generous here), even when you are working with the experimental protein structure! Maybe this is the next frontier for the prediction methods (after they’ve solved the protein/protein complex problem of course), but it seems there is a long way to go. Given that many structures are solved to look at their interaction with other proteins or small molecules I think that experimental structural biology is here to stay for a while - past Tom’s retirement even! However, will these fairly accurate protein predictions make experimental phasing a thing of the past? On Dec 3, 2020, at 4:16 PM, Peat, Tom (Manufacturing, Parkville) <tom.p...@csiro.au<mailto:tom.p...@csiro.au>> wrote: Although they can now get the fold correct, I don't think they have all the side chain placement so perfect as to be able to predict the fold and how a compound or another protein binds, so we can still do complexes. I don't know what others end up spending their time doing, but much of my work has been trying to fit ligands into density, which may take another few years of algorithm development, which is fine for me as I can retire! cheers, tom Tom Peat, PhD Proteins Group Biomedical Program, CSIRO 343 Royal Parade Parkville, VIC, 3052 +613 9662 7304 +614 57 539 419 tom.p...@csiro.au<mailto:tom.p...@csiro.au> ________________________________ From: CCP4 bulletin board <CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK>> on behalf of Jon Cooper <0000488a26d62010-dmarc-requ...@jiscmail.ac.uk<mailto:0000488a26d62010-dmarc-requ...@jiscmail.ac.uk>> Sent: Friday, December 4, 2020 9:55 AM To: CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK> <CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK>> Subject: Re: [ccp4bb] AlphaFold: more thinking and less pipetting (?) Thanks all, very interesting, so our methods are just needed to identify the crystallization impurities, when the trays have been thrown away ;- Cheers, Jon.C. Sent from ProtonMail mobile -------- Original Message -------- On 3 Dec 2020, 22:31, Anastassis Perrakis < a.perra...@nki.nl<mailto:a.perra...@nki.nl>> wrote: AlphaFold - or similar ideas that will surface up sooner or later - will beyond doubt have major impact. The accuracy it demonstrated compared to others is excellent. “Our” target (T1068) that was not solvable by MR with the homologous search structure or a homology model (it was phased with Archimboldo, rather easily), is easily solvable with the AlphaFold model as a search model. In PHASER I get Rotation Z-score 17.9, translation Z-score 26.0, using defaults. imho what remains to be seen is: a. how and when will a prediction server be available? b. even if training needs computing that will surely unaccessible to most, will there be code that can be installed in a “reasonable” number of GPUs and how fast will it be? c. how do model quality metrics (that do not compared with the known answer) correlate with the expected RMSD? AlphaFold, no matter how impressive, still gets things wrong. c. will the AI efforts now gear to ligand (fragment?) prediction with similarly impressive performance? Exciting times. A. On 3 Dec 2020, at 21:55, Jon Cooper <0000488a26d62010-dmarc-requ...@jiscmail.ac.uk<mailto:0000488a26d62010-dmarc-requ...@jiscmail.ac.uk>> wrote: Hello. A quick look suggests that a lot of the test structures were solved by phaser or molrep, suggesting it is a very welcome improvement on homology modelling. It would be interesting to know how it performs with structures of new or uncertain fold, if there are any left these days. Without resorting to jokes about artificial intelligence, I couldn't make that out from the CASP14 website or the many excellent articles that have appeared. Best wishes, Jon Cooper. Sent from ProtonMail mobile -------- Original Message -------- On 3 Dec 2020, 11:17, Isabel Garcia-Saez < isabel.gar...@ibs.fr<mailto:isabel.gar...@ibs.fr>> wrote: Dear all, Just commenting that after the stunning performance of AlphaFold that uses AI from Google maybe some of us we could dedicate ourselves to the noble art of gardening, baking, doing Chinese Calligraphy, enjoying the clouds pass or everything together (just in case I have already prepared my subscription to Netflix). https://www.nature.com/articles/d41586-020-03348-4 Well, I suppose that we still have the structures of complexes (at the moment). I am wondering how the labs will have access to this technology in the future (would it be for free coming from the company DeepMind - Google?). It seems that they have already published some code. Well, exciting times. Cheers, Isabel Isabel Garcia-Saez PhD Institut de Biologie Structurale Viral Infection and Cancer Group (VIC)-Cell Division Team 71, Avenue des Martyrs CS 10090 38044 Grenoble Cedex 9 France Tel.: 00 33 (0) 457 42 86 15 e-mail: isabel.gar...@ibs.fr<mailto:isabel.gar...@ibs.fr> FAX: 00 33 (0) 476 50 18 90 http://www.ibs.fr/ ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 -- Paul Adams Division Director, Molecular Biophysics & Integrated Bioimaging, LBL (http://biosciences.lbl.gov/divisions/mbib) Principal Investigator, Computational Crystallography Initiative, LBL (http://cci.lbl.gov) Vice President for Technology, the Joint BioEnergy Institute (http://www.jbei.org) Principal Investigator, ALS-ENABLE, Advanced Light Source (http://als-enable.lbl.gov) Division Deputy for Biosciences, Advanced Light Source (https://als.lbl.gov) Laboratory Research Manager, ENIGMA Science Focus Area (http://enigma.lbl.gov) Adjunct Professor, Department of Bioengineering, U.C. Berkeley (http://bioeng.berkeley.edu) Adjunct Professor, Comparative Biochemistry, U.C. Berkeley (http://compbiochem.berkeley.edu) Building 33, Room 250 Building 978, Room 4126 Building 977, Room 180C Tel: 1-510-486-4225 http://cci.lbl.gov/paul ORCID: 0000-0001-9333-8219 Lawrence Berkeley Laboratory 1 Cyclotron Road BLDG 33R0345 Berkeley, CA 94720, USA. Executive Assistant: Ashley Dawn [ ashleyd...@lbl.gov<mailto:ashleyd...@lbl.gov> ][ 1-510-486-5455 ] -- ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ######################################################################## To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 This message was issued to members of www.jiscmail.ac.uk/CCP4BB, a mailing list hosted by www.jiscmail.ac.uk, terms & conditions are available at https://www.jiscmail.ac.uk/policyandsecurity/