|
See: "Prediction using deep learning techniques" "Currently it is a bit slow due to the way it is integrated, this will hopefully be improved soon" This statement was written in September 2019 Computing time for a small organic molecule (C10 to C15, molecular weight somewhere between 150 and 200 amu) was around 2 minutes |
A few calculations performed in January 2022 - more than 2 years (!) later Toluene: 2 min 20 sec Steroid-analogue: 1 min 50 sec Heptene-derivative: 1 min 40 sec My question therefore: When is "soon" according to the NMRShiftDB-understanding ? |
| It should offer improved results - see the paper for details Improved results - sounds good - lets run trivial examples |
Alkene-Derivative (Screenshots from Jan 10th, 2022) Both methylgroups (1 and 3) are predicted at 21.65 ppm; it is well known that the cis-Me (#3) is around 18ppm and the trans-Me (#1) around 25ppm. I strongly recommend to have a look into a NMR-textbook to avoid such 'fake-news'. This impression of spoofing the community is intensified when clicking on the line "Show structure with shifts". Now the following picture appears: 
Do you see shift values ? |
| Another "improved" result Screenshots from Jan 10th, 2022 |
 The diastereotopic Me-groups (19 and 20 according to table, 23 and 24 according to picture) show identical shift values at 27.63ppm ! The correct values should be somewhere around 16 and 28 ppm, respectively. Furthermore C25 can't be found in the table and H21 has a shiftvalue of 14.81ppm - a "funny" mixup of numbering only the carbons and all atoms independent of their type ! Clicking "Show structure with shifts" gives again only the structure without shift values.  My summary: Programming as a hobbyhorse ..... poor/no testing ..... please dont bother the community which such poor quality of algorithms and presentation ! |
It makes the already bad situation with NMR-signal assignments worse, because trivial facts at basic text-book level are ignored. The effect of cis/trans-isomerism was already correctly implemented in Wolfgang Bremsers "SPECINFO" at BASF in the late 70's of the last century ! More than 40 years later the same effect is not implemented in the NN-prediction of NMRShiftDB. More to come - stay tuned Dont forget: NMRShiftDB is supported by the German tax-payer ! | |