UWash-n-Fold: Part 2 of the unsolved problem of protein folding.

The 2024 Nobel Prize in chemistry is awarded to the developers of AlphaFold for FAILING to solve the protein folding problem.

In 2024, the prestigious Nobel Prize in Chemistry has been awarded to David Baker, the director of the Institute for Protein Design at the University of Washington (UWash). He is sharing it with - but of course! - Google DeepMind CEO Demis Hassabis and his colleague John Jumper, a scientist at the Alphabet/Google unit.

Baker received one half of the Nobel for his lab’s work on computational protein design. Hassabis and Jumper received their half for work on protein structure prediction, including the AlphaFold models.

Baker is positively bubbling in this interview for Endpoints News:

We know proteins can carry out an amazing array of functions, evolved over millions or billions of years to solve the problems. You think about photosynthesis or all the ion channels in our brains that mediate cognition. We work just amazingly well because we have amazing proteins.

There are new problems today. In medicine, we live longer, so there are new diseases. There’s always potential for new pandemic viruses. And outside of medicine, we’re heating up the planet and polluting it.

Some of these probably would be solved if there was evolutionary pressure and we had another 100 million years to wait. The promise of protein design is to be able to design new proteins that solve current problems as well as proteins in nature solve problems that were relevant during natural selection.

Note that this is not from a live video interview, this is a written quote submitted for an article in trade press. If you are confused by the neo-Darwinian word salad, so am I. Amazing proteins work amazingly because of amazing evolution. Pandemic viruses jump out suddenly from everywhere. In the previous sentence billions of years of evolution was good. But in the next sentence it’s bad and we can’t rely on it because climate change and too many people…. yada-yada… goes the Alphabet mafia errand boy, claiming he is going to OUT-DESIGN billions of years of natural evolution (or God’s design). Phew, nature is stupid and slow. We can’t wait… must accelerate…

OK. Maybe written communications is not the strongest side of the Nobel laureates in chemistry.

Baker is credited with co-founding 21 (!) biotech companies. However, it does not appear that any of them generated substantial monetary value, as he still seems to need the full time employment in academia. Baker’s most recent venture - Xaira Therapeutics, which received $1B from 13 venture investment funds on April 23, 2024, is one of the most richly funded new companies, not only in biotech but across the startup world. Yes, $1B is a heck of a lot of money for the initial round for any company! Typical first VC round is $1-10M. It may seem that way, until you start considering that in AI-dollars, this is like $1.95, and the start-up (now counting about 50 employees) will blow through this pile of cash in no time. Here is a very sane review of the current AI economics and why they don’t make any business sense.

And - you are not going to believe this! - the group has tapped Marc Tessier-Lavigne, formerly the president of Stanford University and chief scientific officer of Genentech, as CEO to turn a cash-flush vision into reality. Remember this story? This is the guy who may or may not have falsified his scientific research and was ousted by Stanford as a result:

Tessier-Lavigne is aglow about Xaira and AI, too. Quote from an interview:

“AI is going to transform every step of the drug discovery process,” Tessier-Lavigne said in an interview with Endpoints. “At the very least, everybody would agree it’s going to improve things incrementally: 10% here, 20% there, 30%. You multiply all of that out and you could get two-, three-fold improvements in speed and success rates.”

Maybe math is not a strong suit of former deans of major universities? A 10%-30% incremental improvements in some steps of the highly complex and multi-step process of drug discovery and development cannot possibly generate a 2-3X improvement in the overall speed and success rates.

You are not going to believe this even more! - Board members include former FDA head, CIA-DOD-Resilience guy, Scott Gottlieb. Scott Gottlieb is reliably found in every corner of the military-industrial globocabal flush with deep state cash and asinine ideas that lead to mass poisoning (as this one surely will, unless, hopefully, they implode first).

The business plan of this venture is pure genius. They are claiming that they will not only design proteins that do not exist in nature, but also “predict” how people’s bodies will respond to them, thus eliminating the need for clinical trials!

It goes without saying that I advise all my readers to stay as far away as possible from any “biologics”, injectable proteins, antibodies and vaccines. This will only get worse, until these morons win all the Darwin awards. Sit this one out, folks, and we will end up running planet Earth.

Review of AlphaFold

I have previously written about the unsolved problem in biological sciences - protein folding. This problem remains unsolved because at least two failed theories have been hard-coded into all protein science: 1) the “dogma” that protein structure and function are determined ONLY by its molecular sequence; 2) the “dogma” that protein folds from the amino-acid chains based on the Newtonian principle of the minimum energy state. Both of these dogmas are hogwash, yet no research grant in this area of science will get funding if it does not affirm these failed concepts. If you are forced to build a house only with plastic cups, forks and spoons you may construct something, but it ain’t going to be a Taj Mahal.

Please note that this section is based on summarizing available information about AlphaFold, that is heavily reliant on the mainstream-acceptable views on the DNA/protein science.

These methods are at best guesses. The numerous assumptions and simplifications that most scientists don’t know they are making, and the inability to distinguish causes from effects renders the whole thing into a complex, money and energy-intensive, but largely useless and dangerous “science” ritual.

Training of the AlphaFold model, using available data:

AlphaFold’s training relied heavily on data from existing databases of known protein structures and sequences. The data in these databases has been assembled from historical experiments on protein folding using methods such as X-Ray crystallography, which I will cover in a future article. The point is, data exists only for a small portion of proteins that could be crystallized by chemists into solids (not at all their natural state, unless you were turned into a salt pillar because you were a bad Sodomite - Gomorrahn):

• Protein Data Bank (PDB): A repository of experimentally determined protein structures. The PDB provides a wealth of data that AlphaFold used to learn the relationships between amino acid sequences and 3D structures.

• Sequence Databases (e.g., UniProt): Extensive databases containing millions of protein sequences. While these sequences don’t have experimentally determined structures, they provide valuable information about evolutionary relationships and conserved patterns.

• Multiple Sequence Alignments (MSAs): These alignments are created by comparing sequences across different species to identify conserved regions. Evolutionarily conserved regions are often critical for the protein’s function and stability, and they can provide clues about the protein’s structure.

How AlphaFold makes its predictions:

AlphaFold takes as input a multiple sequence alignment (MSA) and processes them using an architecture inspired by transformers (a type of deep learning model).

In the first step of the process it’s already evident that it’s full of assumptions, averaging and reliance on guesses “suggestions” of “likelihood” of the interactions of parts of amino acids.

One of the features of AlphaFold, which is hailed as innovative, but in reality is a compounding of assumptions and thus compounding of errors - the use of distance predictions between pairs of residues. Instead of directly predicting the 3D coordinates of atoms, AlphaFold first predicts a distogram, which is a probability distribution of distances between pairs of amino acids. The distogram provides insights into which parts of the protein are likely to be close together, allowing the model to infer how the protein might fold.

The structure module takes the distance and orientation predictions and converts them into 3D atomic coordinates. This module refines the predicted structure by iteratively adjusting the coordinates to minimize physical constraints, such as bond lengths and angles. It "folds" the protein based on the learned patterns and generates a final 3D model.

As an example of how probabilities compound errors, if you have a two-step process, each of the steps with 90% probability of being correct, then the probability of success of this process is 90% x 90% = 81%. If the probability at each step is 50%, then the total is = 25%. In a system with 7 steps, even if each is 90% correct, you get 47% correct as the system output. AlphaFold contains many more steps than that. So you can appreciate how bad it is at predictions.

After the initial prediction, AlphaFold refines the 3D model to ensure it obeys the physical constraints of proteins, such as bond lengths, angles, and torsion angles using energy minimization (i.e. the false dogma of current protein science).

To evaluate and benchmark its performance, AlphaFold participated in the CASP (Critical Assessment of Structure Prediction) competition:

• In CASP13 (2018), the first version of AlphaFold won the competition, demonstrating that its deep learning approach could outperform traditional methods. However, there was still room for improvement. Translation - all models in this competition failed.

• By CASP14 (2020), AlphaFold’s next version predicted structures that were comparable to, and sometimes indistinguishable from, experimental results. Translation - AlphaFold improved, but it still falls short of the experimentally obtained protein folding maps.

Just to be clear, all this modeling is simply trying to train a computer program to regurgitate the existing experimental data! The experimentally obtained maps are not entirely correct either, and, importantly, only possible to obtain for some relatively small % of proteins in nature, and only those that can be chemically coaxed into crystallization.

What can possibly go wrong?

There are many scenarios where AlphaFold can produce errors or unreliable predictions:

Intrinsically Disordered Proteins (IDPs) or disordered regions within proteins do not adopt a fixed, well-defined 3D structure under normal physiological conditions. Instead, they exist as flexible, dynamic chains that can adopt multiple conformations. In fact, the “disordered” is a misnomer. The dynamic state maybe the true state for proteins, because they are part of one-directional, irreversible, asymmetric, cause-effect chain called the process of living, which the current “science” fails to recognize as anything worth studying.

AlphaFold, like many structure prediction tools, is trained primarily on static dead structures. This can result in:

• Overconfidence in bad predictions: AlphaFold may try to force disordered regions into a fixed structure, even though these regions are naturally flexible. This can lead to incorrect or misleading models, especially if users are unaware that the region is supposed to be disordered.

• Forcing false order: AlphaFold might not always recognize when a segment should remain unstructured and may predict it as structured, even when that contradicts experimental evidence.

AlphaFold, particularly in its early versions, was designed to predict the structure of single-chain proteins. However, many proteins in nature do not function alone; they operate as parts of larger complexes consisting of multiple protein chains (homomers or heteromers). This can produce:

• Errors in predicting interfaces: AlphaFold can struggle to accurately predict how two or more protein chains interact, leading to errors in the binding interfaces.

• Incorrect Domain Packing: AlphaFold might predict individual domains correctly but fail to model the way these domains are oriented and packed against each other.

• Fragmented Predictions: In some cases, AlphaFold can predict a structure that appears to be correct when looked at in parts, but when analyzed as a whole, the domains do not interact as they should, leading to a fragmented or unrealistic global structure.

Proteins often interact with ligands, cofactors, or metal ions as part of their biological function.

AlphaFold does not explicitly model these interactions. This will result in:

• Incorrect or Missing Binding Sites: AlphaFold may predict the overall structure of a protein correctly but fail to identify the correct binding site for a ligand or ion, leading to an incomplete understanding of the protein’s function.

• Inaccurate Side Chain Conformations: Since ligands can induce changes in nearby side chains, AlphaFold may not accurately predict these adjustments, leading to errors in modeling enzyme active sites or receptor binding pockets.

AlphaFold’s uses multiple sequence alignments (MSAs) and co-evolutionary signals to predict how amino acids might interact based on evolutionary conservation. This feature was made for designing fantasies about “virus evolution” and “variants”. Of course this bullshit-making feature is crucial for Faking Fakery Faster! (TM) and claims of evolving pandemic viruses, “variants” and zoonotic jumps of, for example, the “avian flu” from an earthworm to dolphin to a nursing home resident that has never been in contact with either of them.

AlphaFold is designed to suggest interactions that are not actually present in the native structure, particularly if the sequences are not closely related (earth worms and dolphins), or if there are errors in the alignment (which are guaranteed to be present).

• False Positives in Contacts: AlphaFold might predict interactions between residues based on co-evolutionary data that do not actually occur in the native protein. This can lead to errors in the predicted structure, particularly in regions where the evolutionary signal is weak or misleading.

• Overfitting to Evolutionary Data: In some cases, the reliance on MSAs might cause AlphaFold to overfit the evolutionary data, leading again to false predictions.

This results in “errors” which are, in fact, the intentional outcomes that pandemic mongers call “pandemic potential variants”, and based on which the “global emergencies of novel viruses” are declared and new mRNA “vaccines” (chemical poison) are produced in a few hours!

Conclusion:

Does the Nobel Prize validate that AlphaFold solved the protein folding problem? NO. The prize, in this case, is false validation that at best says that AlphaFold may one day solve it. There is no evidence today that it will solve it, however. Even the fawning industry press Endpoints News describes it as a version of Midjourney for making up 3D pictures of imaginary proteins:

These computer programs are similar to the diffusion models that power image generators like OpenAI’s DALL-E. But instead of concocting art or a photo based on a text prompt, Baker’s models can create molecular structures of built-to-purpose proteins, such as antibodies that neutralize a virus or kill cancer cells.

You can design your virus or an antibody or whatnot:

Make them scary looking:

Then design the antibody! Kapow!

I am not kidding as much as you think I am. AlphaFold, like all AI models, is not sentient and can’t solve a problem for which there isn’t even a valid theory today, and can’t predict anything. It can only scrape the existing data and provide data parsing/manipulation features. This can have some positive utility for speed and convenience of manipulating existing data, and can be a good tool for some experienced, knowledgeable designers to quickly iterate their own hypotheses, but it does not make anything new.

Deep learning models, including AlphaFold, operate as "black boxes". Tools like this do not improve human knowledge overall. They are useful for speed and computing power IF YOU KNOW WHAT YOU ARE DOING. While they can at times make accurate predictions, it is not always clear why. They also produce “hallucinations” routinely, confidently, and lie quite effectively. Most operators of these models do not understand this, nor can catch all instances of the model hallucinating or being programmed to lie, because some politically motivated behaviors were forced into it at the master level.

Thus, AlphaFold might predict a high-confidence structure for a protein with no known homologs, but the result will be incorrect and largely imaginary.

I may go as far as suggesting that the Nobel Prize was awarded for creating a protein video game so that the new generation of chemists and biologists are zombified into playing it, instead of running valid physical experiments. Thus, it is easier to hypnotize them into rubber stamping the nonsense of “global pandemics” and forget that they are just moving characters on screen that have no relationship to anything real. Chemistry is one of the most precise physical disciplines, and experienced chemical engineers are some of the last remaining bastions of real scientific rigor. That’s about to be finished. Those few remaining people who insist on physical reality and controlled experiments in science will be screened out by yet another method: if you are not using UWashnFold video game engines and the hard-coded conclusions about “pandemic potential viruses” therein, you won’t get any funding for your work.

University of Washington is a recipient of massive amounts of money from Bill and Melinda Gates Foundation.

Do you understand now why Trump made himself available for a 3-hour intimate bromance session with Bill Gates on “global health and vaccines”, and RFK Jr wasn’t invited to that meeting? Happy inauguration day!

Art for today: Still life with daisies, oil on panel, 11x14 in.