Home page
AI is hyped, useless
(Fig.1) AI, Alphafold trained on static experimental proteins can Not design useful enzymes, nor clarify biological reactions.
The 1st, 3rd, 7th, last paragraphs of this hyped news (2/13/2025) say
"A new method that successfully designs serine hydrolase enzymes capable of catalyzing ester hydrolysis with high efficiency, demonstrates a computational approach for creating de novo enzymes" ← wrong ( this-last-paragraph )
" Designing functional enzymes from scratch remains a significant challenge in protein engineering. Traditional approaches often rely on inserting active sites into existing protein scaffolds, which can limit catalytic efficiency due to structural constraints. Until recently, computationally designed enzymes have had reduced efficiencies compared to natural enzymes, however, advances in machine learning and AI open new opportunities for more complex protein design options." ← hype
"The team began their work using the established generative AI framework, RFdiffusion (= trained on experimentally-obtained protein databank or PDB, this-p.68(or p.67)-4.1 ), to design proteins with complex catalytic sites."
"We hope that the concepts and methods we used in this paper will be applicable to designing new enzymes in the future (= just speculation, still useless )"
This research paper ( this ↓ )
p.2-last-paragraph says "de novo design efforts that have attempted to employ this mechanism have been largely unsuccessful" ← AI protein design is impractical
p.3-2nd-paragraph says "We trained this network, called ChemNet, on protein-small molecule complexes in the PDB (= trained on experimentally-determined protein structures = quantum mechanics was useless )"
p.4-2nd-paragraph says "designs were filtered with AF2 (= alphafold2 was used )"
"For experimental testing, we obtained synthetic genes encoding 129 and 192 designs for rounds 1
and 2, respectively, for E. coli overexpression and screening"
p.4-3rd-paragraph says "two round 1 designs (1.6% = 2/129 ) and 10 round 2 designs ( 5.2% = 10/192 ) showed catalytic activity" ← this AI success rate of predicting catalytic acivity was very bad = just 1.6 ~ 5.2%.
p.9-last-paragraph says "they are still two to three orders of magnitude less efficient than native serine hydrolases," ← This AI-designed enzyme's activity was far worse than the natural enzyme, contrary to the hyped news.
The 3rd, 5th, 6th, last paragraphs of this hyped news (7/14/2025) say
"Complete computational design of high-efficiency Kemp elimination enzymes" ← hype, this Kemp elimination enzymes (= useless for biological reactions ) have been studied for 20 years, so Not new, and computer's designed artificial enzymes were unnecessary for this reaction ( this-introduction ).
"Active-site residues were optimized using Rosetta atomistic calculations, yielding millions of designs that were filtered... A total of 73 designs were selected for experimental testing;.. 14 exhibited cooperative thermal denaturation behavior." ← bad prediction rate of only 14/74
"Introducing 5–8 active-site mutations per variant produced designs with increased catalytic efficiencies." ← Which mutations enhanced catalytic efficiencies could Not be predicted by computer AI alone.
"Future advances in modeling theoretical enzymes may (= just speculation, still useless ) enable fully programmable biocatalysis"
↑ This research paper ↓
p.1-left-1st-paragraph says enzymes designed de novo, that is, without recourse to naturally occurring enzymes that catalyse the same reaction, were orders of magnitude less active relative to comparable natural ones" ← today's computer AI prediction is useless.
p.3-left-last~right says "We selected 73 designs for experimental testing.... Three designs showed measurable KE (= Kemp elimination ) activity in an initial screen" ← So the computer AI prediction rate was very bad = only 3/73.
p.3-right-2nd-paragraph says "The catalytic rate and efficiency of these designs.. falling short by several orders of magnitude from comparable natural eliminases and from designed Kemp eliminases that were optimized through laboratory-evolution campaigns" ← So the computer AI design of enzymes is far inferior to the natural enzymes or experimental trial-and-error approach.
p.6-left-3rd-paragaph says "Replacement with Met and Leu exhibited similar Rosetta energies to the original Phe,... Strikingly, Phe113Leu (= tried to replace Phe by Leu amino acid by artificial point mutation ) led to an order of magnitude increase in catalytic efficiency... surpassing by two orders of magnitude recently designed enzymes in artificial intelligence-generated proteins" ← So the conventional experimental trial-end-error approach (= inserting random mutations ) surpassed AI-prediction, after all.
p.8-right-2nd-paragraph and p.9 used also Alphafold (= treating only useless static proteins obtained by experiments ) and the impractical time-consuming MD simulation.
Feel free to link to this site.