Deepmind AI, Alphafold-3 is useless, unable to simulate actual proteins.

Home page

• Alphafold cannot predict proteins dissimilar to its training dataset
• Alphafold cannot explain dynamical proteins' change or reaction.
• Alphafold has to rely on impractical MD to simulate proteins.
• Alphafold relies on useless static protein structures in PDB.
• Hyped Alphafold is useless for drug discovery.

AI, Alphafold just giving (useless) static protein structures from experimentally-obtained protein database cannot predict unknown proteins, drug interaction nor biological reactions.

(Fig.1) AI-Alphafold based on experimentally-obtained static protein structure database (= Not based on useless quantum mechanics ) is unable to clarify true protein mechanism or interaction.  → No drug discovery

AI, Alphafold cannot clarify true protein mechanism nor find new proteins or drugs not included in its training dataset.

Recent Alphafold-2, Alphafold-multimer, Alphafold-3 developed by Google DeepMind published in Nature (+ getting Nobel prize chemistry ) represent today's AI or machine-leaning methods for potential drug discovery.

But the overhyped artificial intelligence (= AI ) and Alphafold are completely useless for finding effective drugs or treatments ( this-8th-paragraph ), far from solving 50-year-old protein folding problems, contrary to many hypes.

AI and Alphafold relying only on already-existing data set of the whole protein structures ( this 3rd-paragraph ) without understanding underlying atomic mechanisms are unable to discover new drugs or treatments that are not included in their training database.

Alphafold has Not clarified atomic mechanism of protein folding at all.

This-last-paragraph says
"but they (= Alphafold-3 ) cannot say why they are folded that way (= protein's mechanism is still unknown ); it will not magically improve drugs' chances in human clinical trials."

This p.1-abstract says
"AlphaFold does Not resolve the decades-long protein folding challenge, nor does it identify the folding pathways"

All these AI and Alphafold rely on experimentally-obtained static protein structure database (= PDB = protein data bank,  this p.8 ), Not on (useless) quantum mechanics.

These AI or Alphafold methods have Not clarified any true molecular mechanisms of protein folding or biological reactions.

Alphafold cannot predict disordered proteins, molecules not included in its training database.

Alphafold's inability to clarify the underlying physical mechanism is why Alphafold cannot predict structures of mutated disordered proteins ( this p.6-last-paragraph,  this-7th-paragraph ), new proteins (or drug molecules ) dissimilar from the protein PDB database on which Alphafold-2,3 (= AF-2,3 ) was trained on ( this p.2-1st-paragraph ).

This-p.2-right says
"It has been trained on hundreds of thousands of experimentally determined protein structures and sequences in the PDB and other databases... AlphaFold has some limitations. It is not designed to predict the effect of mutations"

This-p.4-4th-paragraph says
"AF3 (= Alphafold3 ) failed to produce the correct pose even in the unmutated case,"

Alphafold also cannot explain posttranslational modification of proteins, molecules (= lipids, saccharides,  this-p.5-right-lower,  this-middle-challenge,  this p.10-left-lower ).

The 3rd~4th paragraphs of this site says
"The tool (= Alphafold ) is not capable of predicting metal ions, cofactors or ligands,... PTMs (= post-translational modification ) are not considered in AlphaFold’s structure predictions."

Even the latest Alphafold-3's prediction rate of RNA, DNA structures is much worse than proteins ( this-middle-Limitation,  this p.5-left-1st-paragraph ).

(Fig.2) Alphafold-3 cannot predict protein or enzymatic conformational change.

Alphafold-2,3 cannot explain protein conformational change or enzymatic reactions.

Alphafold-2,3 giving only static protein structures cannot explain protein conformational change (= allosteric enzymatic reaction ) nor dynamical biological reactions ( this-2nd-paragraph,  this 6~7th-paragraphs,  this-middle-alphafold3-limitation ).

So AI-Alphafold cannot be used for drug discovery by clarifying drug's action mechanism.

This-29~35th-paragraphs say
"AlphaFold and other similar self-didactic programmes can predict only the static 3D structure of a protein."

"none of the tools scientists can currently access are capable of generating a clear picture of how the protein structures change in time, and in response to chemical changes."

This-Alphafold-3-Nature paper p.6 says
"A key limitation of protein structure prediction models is that they typically predict static structures as seen in the PDB, not the dynamical behaviour of biomolecular systems in solution. This limitation persists for AF3 (= Alphafold-3 )"

"in AF3, there are still many targets for which accurate modelling can be challenging"

Alphafold cannot predict interactions (= docking ) among proteins and drugs.

Protein conformational change plays an important role in protein (and drug ) interactions, which is why Alphafold cannot predict protein (or drug ) interaction or docking ( this-abstract-lower ).

This or this 5~6th-paragraphs say
"using the AlphaFold structures gives significantly worse results... AlphaFold does not give you any advantages in compound docking"

Even the latest Alphafold-3 (= AF-3 ) failure rate of predicting protein interaction is 60% (= this-abstract-lower,  Alphafold-3 seemingly tries to raise prediction rate by randomly changing initial structural free parameters called "seeds" and choosing the best result, which is Not a legitimate prediction ).

Alphafold cannot contribute to curing cancers.

To cure cancers, researchers must find drugs that bind to only cancer cells, and destroy them by designing and causing the target cell's proteins' conformational change or artificial apoptosis, which is impossible in today's AI or Alphafold giving only (useless) static protein structures.

Cells of cancers, Alzheimer, auto-immune diseases using ordinary humans' (= endogenous ) enzymes cannot be treated like HIV drugs targeting enzymes specific to the exogenous viruses.

↑ Just finding some molecules binding to cancer cells or target proteins (= which is possible in ordinary experimental screening methods that do Not need AI nor Alphafold ) is Not enough to treat cancers.  Designing precise protein conformational change or biological reactions of killing only cancer cells, which Alphafold-2,3 cannot do, is necessary.

Alphafold giving only static proteins can tell us nothing about actual enzymatic reactions, biological functions occurring inside human bodies, so curing many intractable diseases such as Alzheimer and autoimmune diseases by AI Alphafold is impossible forever.

(Fig.3) Alphafold-3 unable to predict protein motion has to rely on very old impractical molecular dynamics (= MD ) that is too time-consuming to simulate biological reactions.  → drug discovery is impossible.

AI, Alphafold giving only useless static protein structure has to rely on unrealistic time-consuming old MD for explaining protein's change.

Even in today's alleged AI or Alphafold era, researchers have to rely on very old impractically-time-consuming molecular dynamics (= MD ) for simulating protein conformational change based on experimentally-estimated (pseudo-)potential called force field ( this p.1-left ).

This current only method (= MD ) of simulating dynamical protein conformational change unable to give actual shapes to individual atoms have to repeat updating each atomic position little by little, many, many times by differentiating pseudo-potential or force fields, which is too time-consuming (= only less than microsecond-MD simulation is possible ) and unable to simulate much-longer important biological reactions or protein conformational change (= protein folding, synthesis ) of second ~ hour time scale ( this 3~5th-paragraphs,  this p.14-left-2nd-paragraph ).

So today's science stops progressing in this unrealistically-time-consuming molecular dynamics (= MD,  this p.8-Limitation ) due to the unphysical shapeless quantum mechanical atomic model.

For example, this recent research ( this p.6 ) used Alphafold2 for estimating only an initial protein's static structure without modification, and relied on the extremely-time-consuming MD to simulate this protein's slight motion of only 500ns.

Another recent research used only the impractically-time-consuming MD simulation of only 500ns-molecular motion without using (useless) AI or Alphafold ( this p.10-MD-simulation ).

This-p.11 and this-p.11 also used only time-consuming MD (= molecular dynamics ) without using useless Alphafold-AI.

Old impractical time-consuming MD instead of Alphafold, AI is still used even in the latest research.

The 5th, 16th paragraphs of this recent (hyped) news (11/1/2024) says

"Deep learning tools like AlphaFold,.. can predict (static) protein structures but Not their folding dynamics (= Alphafold is useless for simulating dynamical protein function ). Traditional (time-consuming) molecular dynamics (= MD ) methods are limited to short timescales"

"It has also been employed to model chemical reactions, such as fullerene synthesis,. (= only impractically-short time scale )"

↑ This research paper's p.2-C60 system ~ used LAMMPS software for the impractically-time-consuming molecular dynamics (= MD ) simulating only 179.44 ns, which is too short to explain important biological reactions such as protein folding ( this p.2-1 introduction-1st-paragraph ).

Medical researches use only macroscopic biological tools without looking into atomic mechanism due to useless quantum mechanics, MD or AI-Alphafold.

(Fig.3')  Overhyped AI, Alphafold, quantum mechanics are useless, Not utilized in actual medical researches.

Medical researches don't use AI or Alphafold.

Most medical researches rely only on traditional macroscopic biological tools without using the useless quantum mechanics (= Schrodinger equation, DFT ), MD nor AI-Alphafold ( this research's this p.8-method~p.11 ), as shown in this.

↑ So today's science, medicine and nano-technology (= cannot utilize detailed atomic interactions ) stop progressing due to useless quantum mechanical shapeless atomic models.

AI, machine-learning methods predict wrong results in proteins that are not included in training data sets.

There are a few AI or machine-learning methods such as DynamicBind and NeuralPlexer2 which are said to change protein structure slightly based on experimentally-obtained protein (complex) structure database ( this p.11-left-Dataset ).

↑ These AI methods based only on the whole protein (static) structure database not understanding true individual atomic interactions tend to give wrong results of protein structures that are dissimilar to the training data sets, so AI is useless for discovering new drugs (= Not included in dataset ) effective for cancers or Alzheimer.

This p.7-Figure.2 shows these latest AI methods such as DynamicBind and NeuralPlexer allegedly dealing with slight protein conformational change gave bad prediction results (= prediction rates were less than 20% ) of protein complexes, when they were tested against new protein structure data (= from new PDB data that is not included in training data sets ) called PoseBusters ( this p.5-2.6 ).

This recent paper's p.1-abstract-lower says
"No deep learning-based (= AI ) method yet outperforms classical docking tools."

Alphafold-3 often gives illusory protein structures called hallucination.

Even the latest Alphafold-3 using the generative models (= Not understanding true atomic interaction ) tends to give illusory protein structures called hallucination.

This-2.risk of hallucination says "The use of diffusion techniques (= like Alphafold-3 ),.. introduces the risk of the model hallucinating — generating plausible but non-existent structures"

This 1. and 6. say
"AF3 (= Alphafold3 ) is a generative model. This means it can give different outputs for the same input (= random outputs with No physical principle ) and that it can hallucinate (= giving illusory protein structure )"

"We're still a long way from replacing experimental methods with ML (= machine-learning ) in the life sciences"

Medical researches have to rely on traditional macroscopic biological tools originating from natural organisms without utilizing nor clarifying atomic mechanism due to the useless quantum mechanics, impractically-time-consuming MD, AI-Alphafold that just gives static protein structure.

We need multi-probe atomic force microscopes to know and manipulate precise protein behavior at atomic level.

(Fig.4) Today's experimental methods (= X-ray crystallography ), AI, Alphafold cannot clarify real protein structures or biological reactions.

Today's experimental methods such as X-ray crystallography, cryo-EM, NMR cannot tell us real dynamic protein's motion and reaction inside human bodies.

It is impossible to know true protein, enzymatic conformational change and reaction only from the static protein structures (= PDB ) obtained from X-ray crystallography (= using crystallized static proteins ), NMR (= measuring only small proteins ), cryo-EM (= electron microscopy insufficient for drug development ).

Actual dynamical protein structures inside human bodies are often different from static protein structures (= such as X-ray crystallography ) registered in PDB database on which today's AI and Alphafold rely.

This is why all the current AI and Alphafold are unable to predict actual dynamical protein or enzymatic structures and functions.

Developing atomic force microscopes with multiple probes to manipulate individual atoms is necessary.

The only possible way to know precise molecular or protein conformational change at an atomic level is atomic force (= AFM ) or scanning tunnel microscopes (= STM ).

The problem is today's atomic force or scanning tunnel microscope has only one probe tip that cannot know the target atom or grab individual atoms freely.

There is already the technology of atomic force or scanning tunnel microscopes with multiple probes, but these are used only for measuring electric conductance between two probes, Not for grabbing or identifying target atoms with multiple probes.

It is surprising that today's atomic force (or scanning tunneling ) microscopes still have only one probe tip, though more than 30 years have passed since they first succeeded in manipulating a single atom.

↑ This is due to the unrealistic quantum mechanical shapeless atomic model prohibiting researchers from dealing with atoms as real objects by developing multi-probe atomic force microscopes for a long time.

We should develop atomic force or scanning tunneling microscopes with multiple probe tips manipulating individual atoms freely as soon as possible to clarify precise proteins' behavior at atomic level and develop effective drugs for the current intractable diseases.

Bottom-up and top-down approaches to precisely knowing target proteins' behavior is necessary.

By using multiple probe tips (= to which single atoms are attached ), we can design and construct artificial molecular or protein devices in the bottom-up approach (= microscopes can manipulate molecular bond formation ).

In the same way, we can disassemble and analyze isolated natural proteins little by little by using multiple probes of atomic force or scanning tunneling microscopes (= by causing artificial hydrolysis ) to know the precise protein structures inside human bodies in the top-down approach.

↑ By combining these top-down and bottom-up approaches using multi-probe atomic force microscopes, we can know and confirm the actual protein interaction, behavior and conformational change (= induced by approaching molecules or ions carried by microscope tips ).

Alphafold could predict new protein complex (= Tmem-81 ) in fish's sperm ?  ← But Alphafold tells us nothing about its biological mechanism.

(Fig.5)  Alphafold, AI, today's medical research cannot clarify true molecular mechanism nor find effective treatments due to useless quantum mechanics.

AI-Alphafold tells us nothing about true physical or biological mechanism.

Today's overhyped AI, Alphafold and medical researches are unable to consider actual atomic interactions due to useless quantum mechanics.

This is why finding effective drugs for cancers and Alzheimer is still impossible despite long time researches (of AI or medicine ).

For example, the 5th, 7th paragraphs of this hyped news (10/27/2024) say

"Google DeepMind's artificial intelligence tool AlphaFold... to help them identify a new protein (= false, already-known protein ) that allows the first molecular connection between sperm and egg."

"Scientists still don't know how the sperm actually gets inside the egg after it attaches and hope to delve into that next."  ← AI-Alphafold and biological researches are still unable to explain true mechanism of fertilization after all.

Alphafold cannot clarify true mechanism → Biology also cannot explain atomic mechanism.  → No cure for intractable diseases.

↑ This research paper ↓

p.2-1st-paragraph says "our insights into vertebrate sperm-egg binding and fusion lack mechanistic understanding"  ← (useless) Quantum mechanics failed to clarify fertilization mechanism even 100 years after its foundation.

p.3-1st-paragraph says " we performed an AlphaFold-Multimer in silico screen against ~1400 zebrafish testis-expressed secreted/transmembrane proteins (= these are already-known proteins, Not new )"

"Intriguingly, Izumo1-Spaca6 was one of the top-scoring predicted interacting pairs, which was unexpected given that recombinant human IZUMO1 and SPACA6 ectodomains do not appear to interact in vitro"  ← Alphafold's prediction failed.

p.3-2nd-paragraph says "AlphaFoldMultimer predicted with high confidence the formation of a trimer between zebrafish Izumo1, Spaca6, and Tmem81 (= new finding )"

p.5-2nd-paragraph says "we generated knockout lines in zebrafish and mice using CRISPR/Cas9.. Tmem81-/- male fish (= Tmem81 gene deleted ) and mice were sterile,"  ← Tmem81 protein seemed to be necessary for fertilization (= which was confirmed by biological experiments. Not by Alphafold ).

p.14-last-paragraph says "Tmem81 was independently discovered in a search for remote homologs of Spaca6 proteins (= so this Tmem81 protein was already known to have homology with fertilization-related proteins before Alphafold )"

p.15~p.20 used conventional macroscopic biological experiments such as zebrafish/mouse fertility tests (using bacterial-CRISPR gene knock out ), immunoprecipitation based on antibodies (= obtained from immunized animals ), GFP fluorescent protein from jellyfish (= No quantum mechanics nor Schrodinger equation was used in this biological research ), without looking into detailed atomic interaction

Alphafold was useless for elucidating the mechanism of fertility after all.  → No infertility treatment nor drug discovery.

This or this-p.10-Limitation of the study (= Discussion-lower ) admits
"we lack details on stoichiometry and structure... Furthermore, our study does not provide in vivo evidence for complex formation... it is unclear whether Tmem81's function extends beyond"

↑ Today's medical researches and AI-Alphafold had No ability to clarify atomic mechanism ( this 11th-paragraph ), so finding effective treatments for diseases is impossible, contrary to hypes.

Feel free to link to this site.