Hello Dr. Tainer

Thanks for providing this question and discussion platform regarding the potential advances of AlphaFold

I think the possible advances of using the AlphaFold in research and medicine is primarily for structure-based drug discovery and development. As well as, I think the accuracy of using this algorithm would aid in rational design of drugs based on their human protein targets in evaluating their biophysicochemical properties in silico.

Deviating from the human protein folds, I hope that using AlphaFold too would allow accurate prediction of peptide sequences from nature (e.g. venom peptides) that has no experimental structural characterizations but has promising bioactivities for pain and neurodegeneration to aid in drug development.

Best regards.

I have been really shook up by AlphaFold given that I did not see this advance coming. I feel like it really does change the way we do structural biology or even if we need to do structural biology. Since like 2 days ago, DeepMind and EMBL-EBI released AlphaFold Protein Structure Database in which they applied AlphaFold to the human proteome of sequences in the UniProt database and they intend to use it on all proteins. It took them 2 actual days to AlphaFold 350,000+ proteins in multiple organisms from their sequences given that they have amazing supercomputers.

I was really shook up by this research article:

Article Crystallographic molecular replacement using an in silico‐ge...

Basically, AlphaFold predicted the structure of ORF-8, a coronovirus protein, and the predicted solution was sufficient to model into actual electron density. Existing homologous structures were not sufficient. They tested other protein prediction models from the other groups in the CASP-14 competition and none of those solutions were sufficient.

I am almost somewhat skeptical about drug discovery, though because even for proteins that already have known structures, in silico-modeling of a huge drug databases gives hits for drugs. But some the best predicted drugs don't actually bind with high affinity to the proteins, with known structures, as shown from experimental studies. It can be about as efficient as just screening a bunch of drugs that are known to inhibit enzymes in the same family.

I don't really know how well AlphaFold can predict protein structures when they are in the multiple states. There are multiple possible states:

a protein in solution, a protein bound to the ligand, a protein bound to its product. For big substrates like DNA or hemes, the binding of the ligand can induce a large conformation change that can fold large region of the protein that is otherwise disordered according to optical techniques like circular dichroism.

I wonder if AlpaFold predicts protein structures as if they are free in the solution or bound to their ligand. Obviously there is no ligand in the AlphaFold predictions, but the Artifical Intelligence (AI) was trained on protein structures in the protein data bank. If almost all the structures in the PDB for given enzyme are of the complex because the free enzyme doesn't crystallize, then the known fold that AlphaFold should predict should be the enzyme in a closed state as if it was binding its ligand. There are intrinsically disordered proteins which have no secondary structure according to circular dichroism or nuclear magnetic resonsance. However, when the enzyme binds multiple of its ligand, lots of evidence of secondary structure is observed. Which state does AlphaFold predict?

Somebody posted on twitter that they were trying to use AlfaFold to predict structures of complexes by connected the sequences of two interacting proteins with a long glycine linker connecting them. I wonder, if you connected two proteins that are known to not interact at a all with a long linker if AlfaFold would predict an interaction when there is none.

According to the CASP14 press release, the topic of CASP15 is to predict the structures of protein complexes: https://www.predictioncenter.org/casp14/doc/CASP14_press_release.html

I think AlphaFold is going to dominate that again.

There has been many advances recently but I am not sure if it is by this

The credibility of structural models predicted using AlphaFold depends heavily on existing experimental structures, and as mentioned in many articles, DeepMind’s code will provide no information about how polypeptides are assembled into the complex structures needed for their function. It can only provide a snapshot of the protein conformation. In my opinion, AlphaFold facilitates protein structure prediction, but it is not enough to be used as a reliable basis for protein structure determination and drug design. We need to verify the predicted model experimentally in the end.

Roger Davis makes a good point. That said, having this folded proteome teaches us some things about the information in both evolutionary sequence comparisons and structures in the Protein Structure Database. These testable fold models were based largely upon X-ray crystal structures. So it will be important to know for example if we can detect solution structure differences of functional relevance.

Bhaskar Ganguly rightly complains about AlphaFold swooping in to steal the fun of folding with logical inferences. However I suggest that we can now more readily take this to the next level by having relatively accurate folds. This next level is all about using accurate folds to define both conformations (allostery) and complexes (interfaces and communication/regulation). In a sense there is more juice here where the complexes and conformations are more likely to speak directly to activity, regulation, and function. Why not jump this fold level and go bigger for more juice?

Useful blog on AlphaFold2 - https://www.blopig.com/blog/2021/07/alphafold-2-is-here-whats-behind-the-structure-prediction-miracle/

For long years we have desired a tool that would quickly give us accurate predictions of protein structures. X-rays, then NMR are valid tools but it takes time, expertise, and adequate instrumentation. To all this was added molecular dynamics, strongly complementary to the other two technologies, which also gave us information that is not easily accessible. In all cases it takes time.

Now we have AlphaFold that with a click gives us the result. Certainly not all that glitters will be gold, but it is a huge step forward. Some time ago I was considering a receptor of about 1500 aa with partial reconstructions made with X-rays and NMR. These reconstructions revealed two large, well-structured domains on the extracellular and cytoplasmic side. I found the structure of the whole protein on AlphaFold. To my surprise, I found that the two external domains were mixed structures (more or less long disordered segments mixed with well-organized domains). I found long disordered segments but some were also quite short (about 10 aa), not always easily visible, also with predictive sequence techniques, because sequence predictions themselves were dubious. This changed my ideas on the ability of the two domains to interact with numerous molecular partners and on the localization of the active site. The advantage was exactly the fact that it quickly showed me aspects that hadn't been seen before, but with a simple click. While I understand what alpha fold does, this is irrelevant, as it will allow a great many biomedical researchers to quickly access previously forbidden or hard-to-reach levels of knowledge. Certainly we have lost the aura of respect that those who dedicated themselves to predictions had among clinical colleagues, but now we can access higher levels of biological knowledge more easily and with greater confidence. One for all, what John Tainer suggested, the allostery and its regulation. For example, how does biological information travel through a transmembrane receptor? With what energies at stake? With quantum mechanical effects? Precise knowledge of the structure and a little molecular dynamics in water could give answers that we still don't have clear today. The future of biology strongly depends once again on the knowledge of structure-function relationships. Otherwise, high-throughput data biology will proceed undaunted to predict the functions of biomolecules in silico with comprehensive bioinformatics analysis on data from omics technologies, not caring about the fundamental paradigm of structure-function relationships and data experimentally proven in lab.

Greetings

Much of the actual experimental structure determination concentrates not so much on fold determination, but on the subtle conformational changes involved in the regulation of protein function, and in protein-protein and protein-ligand interactions, with important lessons learned not from one single structure, but from a large number of structures determined in different contexts and by different methods. As to natively unfolded or conditionally folded regions of proteins, I am not impressed by the alphaFold models shown e.g. in UniProt. This is not surprising - since the conformational range of these regions is not easily accessible to direct experimental observation, there is no training set for artificial intelligence yet to learn from.

In my own career (modelling in intense interaction with experimentalist to explain and adapt sequence-structure-function relationships), I usually have learned much more from the actual modelling process than from the final model itself.