Adventures in Polypeptide Space

The Cosmos is a 'mind-bogglingly' big space and it's unlikely we shall ever explore it. 'Polypeptide Space' is vast too, but there's a better chance we might get to explore those parts of it that are important to us - Michael Geisow

Polypeptides make life, but what exactly is life?

Taking the cell as the basic unit of life, let's see where proteins play their part. It's worth recalling the 8 properties currently used to define the state of 'being alive'. After all, with our push toward machine 'intelligence' and our exploration of the Cosmos, one day we may need to amend our present definition.

The accepted definition of life (and the natural roles of proteins)

• Order Proteins ensure the temporal integrity of the cell and maintain its structure, shape and motility.
• Response to stimulus All life is 'aware of' and dependent upon its environment. Proteins provide both internal and external sensors.
• Reproduction Protein machinery operates the mechanics of DNA replication and cell division.
• Growth and development No life forms are static. Their dynamic nature is operated by proteins.
• Management Proteins provide oversight of all cell processes. There's no overall boss: successful businesses that they are, cells run on 'matrix management'!
• Energy Energy is essential to life. The key energy generators of all life forms are protein complexes.
• Homeostasis 'Keeping on keeping on' through defence and repair relies absolutely upon proteins.
• Evolution Providing cells with the capacity to adapt to environmental change (and passing this capability on) is intrinsic to the way proteins are structured. Built up with domains (functional subunits) these domains can be duplicated. mutated and passed between organisms to endow them with new functions

Where did proteins come from & what kicked it all off?

It was once an RNA World We are the inheritors of biopolymers forged - perhaps accidentally - in deep time. Nucleic acids started it and (we believe) proteins were relative latecomers to the game of life. Where these nucleic acids came from (on planet earth or as those voting for an extra-terrestrial origin panspermia believe) no-one really knows. Against entropy, nucleic acid molecules grew in complexity, increasing their polymer lengths, acquiring a range of catalytic and structural functions and the capability to reproduce themselves. Eventually the information about these properties became more stably stored in DNA and the capacity 'do more interesting things' led to the invention of proteins. But, the founding member of corporate life - we assume - was RNA, which in those distant days, in the immortal words of Canadian philosopher Marshall McLuhan: 'the medium was the message'. In truth, this is a classic 'which came first, the chicken or the egg' question. The answer is buried in deep time.

Even today, RNA has retained the critical controls RNA collaborates with protein 'acolytes' in the manufacture of proteins in the molecular machines called ribosomes. RNA has retained the critical step of forming peptide bonds. And like card-driven Jacquard looms, ribosomes turn out proteins to the specifications written in messenger RNA. I like to think that early RNA 'delegated' these specifications (genes) to be stored and curated within DNA!

Genome manuals specify many more instructions than products To fulfil the 8 characteristics of life listed above, a lot of proteins are required. But (in us at least) only 1-2% of our DNA stores the information which specifies them. A much larger percentage of our DNA holds the information about which proteins will be produced (expressed) and how, when and where this will happen. The management and mechanics of this information transfer (transcription) once again is a critical collaboration between RNA and proteins themselves.

Editing and Revision Control of protein manufacture occurs at many stages. Genes (DNA) specifying proteins are subject to multiple controls, many still being uncovered. Gene transcripts (RNAs) are tightly regulated both in nature and quantity. Like sentences in the writing of a novel, RNA transcripts are subject to editing (and even redaction!) in producing the messengers (mRNAs) which will be translated into proteins. Following translation into proteins in the ribosomes, more editing occurs. Newly synthesised proteins can be subject to post-translational modification and even proteolytic processing in which their polypeptide chains are cut and peptides removed. Certain of this processing relates to whereabouts proteins are to be sent. Proteins come with peptide 'addresses' which direct them to cell compartments or to the environment external to their host cells.

Timeline of protein science: from Berzelius to BoltzGEN

---

Extraction and purification

The extraction and purification of proteins for in vitro studies was a challenge faced by many protein researchers (who were by then calling themselves biochemists). Proteins of interest were generally present within complex mixtures and often at very low concentration. Very many scientists: PhD graduates, post-docs and established scientists contributed to present day purification techniques. These techniques exploited the size and shape of proteins, solvent solubility, surface properties including net charge and their specific affinity for small or large ligands (now popularly called 'binders'!)

At first, fairly crude purification techniques were employed guided by assays to follow fractions containing the desired activity. Differential precipitation using inorganics like ammonium sulphate (salting out) was useful especially with bulk tissue extracts of plasma. Another early method was ultracentrifugation which separated protein molecules by size and shape. Thermostable proteins were sometimes recoverable by heating extraction media to precipitate thermally labile contaminants. In the 60s and 70s the substantial and pricey hardback 'tomes' of the Methods in Enzymology series were the 'go-to bibles' for purification of many protein molecules. This was, of course, before the democratisation of methods brought in by the World Wide Web!

Column chromatography

Affinity separation of biomolecules

---

The huge impact of UV / Visible light spectroscopy

1935 Arnold Beckman created the National Technologies Laboratories—later named Beckman Instruments. Triggered by the American government’s interest in measuring vitamin content in soldier's rations using UV and Vis light, research led to commercialisation of UV-Vis spectrophotometers in the early 1940s. Of these, the Beckman DU spectrophotometer—first sold in 1941—distinguished itself from competing products by delivering more accurate results and reducing analysis time from hours, or even weeks, to minutes.

My old Head of Department, Gilbert Beavan at MRC National Institute for Medical Research UK was a consultant in the development of the Beckman DU spectrometer, applying this to the measurement of proteins. Light intensity is measured from UV-Vis source lamps before and after the light passes through a sample contained in quartz cuvettes or capillaries. The amount of light absorbed corresponds to the molecular concentration in the sample. The instruments are now ubiquitous for measurements and detection of pharmaceuticals and biopolymers.

Nobel laureate Bruce Merrifield (developer of solid phase peptide synthesis) referred to the UV-Vis spectrophotometer as “probably the most important instrument ever developed toward the advancement of bioscience.”

The later development of fluorescence spectroscopy - in essence the emitted light from UV/Vis illumination of biopolymers - has similarly become indispensable to biochemists.

Fluorescence spectroscopy and microscopy

Circular dichroism spectroscopy

Stopped flow method for following protein interactions

Caged molecules for measuring real time response

Further knowledge about protein function, especially enzyme kinetics and conformational change has been obtained using caged ligands - especially ATP. After a set up containing a protein of interest has reached equilibrium, the caged reactant molecule is released by a laser light pulse. Fast detector output can be measure using spectroscopic systems or high brilliance X-Rays from a source like the Diamond light synchrotron.

Exciting recent applications of caged molecules have explored conformational states of membrane proteins trapped in transition stages using CryoEM to resolve 3D structures. The mechanism of action of receptors and membrane channels can be dissected in this manner.

---

Assessment of purity and characterisation

Gel electrophoresis

HPLC and peptide mapping

Electrospray Mass spectrometry

---

Amino acid sequence determination

1955 Prior to this time, chemists had determined the amino acid composition of proteins. The technique used was 'force majeure' : heating protein in concentrated HCl at 100^oC to break the peptide bonds. The released amino acids were then separated and quantitated by ion exchange chromatography using the amino group reaction with ninhydrin to yield a blue adduct. But the sequence of amino acids was only determined in short peptides using a variety of chemical and enzymatic methods. The first protein sequenced was insulin by the Cambridge chemist Fred Sanger. A small protein, but still an heroic achievement especially as insulin is made up of two polypeptide chains linked by disulphide bridges.

Sanger used the 'Sanger Reagent' fluorodintrobenzene (FDNB) to react with free amino groups of insulin - this gives a yellow reaction product. The labelled insulin was then cleaved into peptides by proteases or HCl. The labelled peptide amino acid compositions were identified by the ion exchange method described. Peptides were treated again with FDNB and once again hydrolysed. The labelled N-terminal amino acids were separated and identified by paper chromatography. The process was repeated many times to build up a complete picture of the sequence. This approach was also used to obtain the complete amino acid sequences of human and bovine serum albumin (over 500 residues!) by James Brown and co-workers in the 70s.

Manual sequencing methods were increased in sensitivity by the introduction of dansyl chloride to label the amino termini of proteins or peptides. The resulting dansyl derivative is highly fluorescent and can be released by acid hydrolysis and detected by 2D thin layer chromatography (TLC). Later a coloured reagent, dabsyl-isothiocyanate (DABITC) was introduced which replaced PITC in the Edman degradation described below, releasing intensely red coloured thiohydantoin derivatives which were again identified by 2D TLC. There is an excellent description of manual protein sequence analysis by Yarwood (Dept of botany, Durham University) in our Protein Sequencing book illustrated below.

The Edman sequence method

Mass spectrometric amino acid sequence analysis

---

Protein structure determination

X-ray crystallography

A 2025 open source review (from the Journal of Plant Physiology) by Veronica Maurino is worth a look: "Next generation technologies for protein structure determination: Challenges and breakthroughs in plant biology applications"

It's beyond doubt that cryoelectron microscopy will solve many questions in cell biology. A considerable number of these concern the structure and action of membrane proteins which enable cell communication with the outside world.

For example, all opioid receptors are 7-transmembrane G-protein coupled receptors with an extracellular binding domain and an intracellular signalling domain. Opioid binding to the extracellular domain activates cytoplasmic heterotrimeric G proteins by opening the Gα α-helical domain (AHD - see image) to enable GDP–GTP exchange (the signal). There has been an unsolved question about how the opiate antagonist naloxone rapidly mitigates opiate signalling. Naloxone is an acute drug for reversing opiate poisoning, yet its mechanism of action had r\until recently remained unknown. Opiate drugs act through activating the μ -opioid receptor (MOR) and its coupled G protein. Saif Khan et al. ( 2025 Nature 648 755) used CryoEM to capture snapshots of intermediate conformations of MOR - G protein with bound opioid agonist or naloxone. Their work provided insights into G-protein coupled receptor pharmacology.

---

Protein structure prediction, interactions and design

From the early days of X-ray crystallographic protein structure determination scientists all over the world attempted to predict the 3D folding of proteins of unknown structure. As the database of experimentally determined structures grew, this looked feasible.

Around 1970 Researchers Peter Chou and Gerald Fasman developed an algorithm to predict runs of secondary structure (α helix and β sheet) using indices calculated from the frequency of appearance of residues in experimentally reported protein structures. They termed these indices propensities. For instance, the amino acids Ala, leu and ile would appear often in α helix, whereas glycine and proline would appear more frequently at the termini of helices. This approach was modestly successful and I developed a further slight improvement in the predictive accuracy of amino acid propensities in 1980 using the then increased number of known structures and splitting them into α or β class proteins.

Secondary structure could be approximately located in sequences and homologies with known protein structures were encouraging, but tertiary structure determination remained elusive. Attempts to fold proteins ab initio using molecular dynamics approaches ran into computation time limitations. Yet, logically, proteins fold rapidly following biosynthesis and also rapidly following experimental unfolding by denaturants in the lab. What did nature know that we didn't?

Critical Assessment of Techniques for Protein Structure Prediction

2018 saw AlphaFold generated protein structure predictions enter the international CASP (CASP13) competition. The results led founder and organiser, John Moult from the University of Maryland, USA, to pronounce that the long time protein folding prediction had finally been cracked. Timeline: CASP 13 AlphaFold1 led in the most difficult category, where no homologous protein structures were available which could act as templates. CASP14 (2020) AlphaFold2 results led the organisers to declare that the protein folding problem had been solved for single chain proteins.

Since AlphaFold’s breakthrough in 2020, the landscape of protein structure prediction has diversified rapidly. Although the management of the originating commercial business, Deep mind, might have restricted access, they have made the programme open source in the form of the AlphaFold server and created the AlphaFold database of millions of predicted protein structures. In the spirit of this openness, follow on AI supported software is similarly available through online access. DeepRosettaFold, ESMFold, OmegaFold, AlphaFold3, and BETA are the most notable entrants, each pushing the boundaries of speed, accuracy, and biological relevance.It has come to cover protein and nucleic acid structures and interactions. These tools extend or compete with AlphaFold’s deep learning approach, each with unique strengths in speed, accuracy, or integration with other biological data.

Programme	Source/Year	Technology	Pros	Cons
RosettaFold	Baker Lab, Univ. of Washington (2021)	Three‑track neural network (sequence, distance, coordinates)	Flexible, good for diverse tasks, integrates with Rosetta ecosystem	Slightly less accurate than AlphaFold2 on benchmarks
ESMFold	Meta AI (2022)	Large language model trained on protein sequences (Evolutionary Scale Modelling)	No need for multiple sequence alignments (fast for novel proteins), scalable	Accuracy lower than AlphaFold2 for well‑studied proteins
OmegaFold	Helixon (2022)	Transformer‑based deep learning	Lightweight, rapid predictions, competitive accuracy	Less validated on large benchmark sets compared to AlphaFold
AlphaFold3	DeepMind (2025)	Expanded deep learning framework	Improved accuracy, predicts complexes and interactions, broader biological scope	Computationally intensive, still evolving in accessibility
BETA	Launched alongside AlphaFold3 (2025)	Specialized deep learning	Focuses on flexible/disordered proteins, challenging cases	Early stage, less widely adopted yet

Predicting protein folding and interactions

Amino acid sequences determine polypeptide 3D folding and function

The folding of polypeptide chains is dependant upon their amino acid sequences. The dominant forces which structure polypeptides in 3D are: hydrogen bonds mainly between the atoms which make up the polypeptide chain (often called the backbone in the literature); Hydrophobic interactions between non-polar amino acid side chains which are generally buried away from solution; Salt bridges between polar side chains and Disulphide bridges formed between cysteine amino acid side chains. There are other interactions too called π - π between buried aromatic amino acid side chains. Extensive disulphide bridging can be seen in the serum albumin molecule illustrated below.

Studies of proteins from organisms which live in very hot or high salt environments (extremophiles) generally have more extensive hydrogen bonds, disulphide links or salt bridging between polar side chains. But all these interactions are absolutely dependant upon an aqueous environment. Proteins won't fold in organic solvents. You just end up with an aggregated, structureless mess. It's this universal observation which underlies the (expensive!) search for water on cosmological bodies. Without water, life, at least in the way we have defined it above, just couldn't exist.

Some proteins are exceptionally highly conserved.

Such proteins are considered universal because they have changed very little in amino acid sequence and 3D structure throughout nature. They're especially conserved from microbes to man because their properties and functions are fundamental to life. They are key players fulfilling the eight criteria for life listed in the introduction to this article. Evolution has 'tinkered' extensively with many proteins, changing their amino acid sequences and shape and duplicating or deleting segments of polypeptide chains. But not these proteins: the penalty for even single amino acid changes through mutation in these is usually death! A particularly appropriate maxim here being 'If it ain't broke, don't (even think to) fix it'.

Knowing polypeptide 3D structures is essential to explain their properties and functions

To understand what makes life 'tick' it's necessary to determine protein structure and function. Knowing the 3D structure of proteins is critical for enabling applications like the development of drugs. The recognition that proteins have distinct folding of their polypeptide chains was first proved for myoglobin and haemoglobin by Max Perutz and John Kendrew at Cambridge in the 1960s. It's worth noting that the isolation and purification of proteins at that time was still a considerable challenge. With present day techniques and gene cloning, just about any protein can be prepared in pure form for structural studies. 3D protein structures can then be determined by X-ray crystallography or cryoelectron microscopy. Remarkably accurate computer protein structure prediction methods have been developed in the last few years as described later.

Other regions of polypeptide chains aren't characterised by such regular structures, but nevertheless adopt similar conformations in homologous proteins. The association of these secondary structural elements to give the folded proteins seen in the PDB is called tertiary structure. A large number of proteins are formed by the further non-covalent association of separate polypeptide chains as in haemoglobin and transthyretin, described below. This association is known as the protein's quaternary structure.

Natural proteins can broadly be classified into structural classes

Globular proteins can be loosely described as alpha class if they're folded almost completely with alpha helix secondary structure as shown for human serum albumin and the annexins. Or they can be regarded as beta class if their polypeptide chain is almost all folded as beta secondary structure as first shown in human transthyretin, whose structure I solved by X-ray crystallography in 1970 (illustrated below). These classes of protein are not particularly distinguished by their amino acid content. It;s their amino acid sequences that lead to the distinct alpha or beta classes. With a huge number of protein structures now determined, it can be seen that these proteins represent two architectural extremes. Many proteins have a mixture of these secondary structures, or sometimes none at all.

Fibrous proteins (like collagen) which frequently have monotonously repeating amino acid sequences with few hydrophobic side chains, tend to form extended alpha helices leading to their description as 'coiled' proteins. Such extended helices often self-associate as dimers or trimers which twist around each other. These assemblies are correspondingly described as coiled coils. Although fibrous proteins are generally thought of as having structural functions, some of these coiled coil assemblies participate in key cell membrane biology. One example are the members of the BAR protein family. These self associate as coiled coil structures which bind to membrane lipids and support the bending force important in pinching off cell membrane as vesicles in cell transport processes.

Closely similar (homologous) polypeptide domains which evolved for specific functions are frequently found in very different proteins. This is evident in the annexin family of calcium-regulated lipid membrane binding proteins. The image (left) above highlights the four annexin Ca²⁺ and lipid binding domains (differently coloured for emphasis). In the right hand image annexin amino terminal sequences shown (in red) are completely dissimilar in fold and function from the highly conserved helical core regions (highlighted in yellow). It's now known these amino terminal amino acid sequences are highly flexible and specifically bind different biopolymers. These include proteins, RNA and glycans. Here nature is showing us duplication and repurposing of an important functional module (called the 4 domain annexin core). The critical function of this core is to bind to cell membranes. The very different amino terminal 'domains ' then enable annexins to link other cell components to cell membrane surfaces. The annexin family is described in depth in another section of this website.

The repeating homologous polypeptide loop structure of HSA and the presence of 3 structural (and functional) domains was described earlier. Another very important class of proteins which are built up of distinct domains are the proteins of our immune system which include the antibodies. In this case the antibody domains have similar folds but quite distinct function. There's abundant evidence that polypeptide domains have been shared by 'horizonal gene transfer' both within and between organisms. Bacteria share genes in this manner which is often the basis of transfer of antibiotic resistance. In the past viral infection may have been responsible for such gene transfer in higher (eukaryotic) organisms too.

Polypeptide chains adopt 3D folds but these are dynamic.

Biophysicist: 'What skills did evolution teach you in folding lessons?'
Intrinsically Disordered Protein: 'Reeling & Writhing & fainting in coils'
quote - Lewis Carroll: Alice's Adventures in Wonderland

Polypeptide chain flexibility, especially at the beginning and end of polypeptide chains, is common and is often critical to protein function. Protein structures determined by X-ray crystallography or cryoelectron microscopy (CryoEM), by the very nature of these methods, show 'frozen' or time-averaged views of the polypeptide chain. Proteins in the PDB (Protein Data Bank) are shown with static folds, but that isn't how they exist in cells or researcher's test tubes. Under physiological conditions proteins are essentially dynamic. This includes movement of buried secondary structure. Some regions of polypeptide chain are particularly mobile. These are often the polypeptide chain ends (the amino or carboxy terminus). Chain termini can be in such rapid motion they are invisible in structures determined by X-ray crystallography.

Unlike crystals of small small molecules, protein molecules are surrounded by water in their crystals which permits chain motion as well as the free diffusion of ions and organics. In solution during lab experiments and within cells the chain termini and sometimes the entire protein molecule may lack regular order. These are now referred to intrinsically disordered proteins (IDPs). The amino termini of the three proteins I studied in depth (HSA, transthyretin and annexin) were each 'unstructured' under particular conditions. Each region has an important role to play in these proteins functions.

Group hugs!

Polypeptides with the intrinsically disordered regions (IDRs) mentioned in the last section are frequently involved in assembling molecular machines and subcellular compartments. These assemblies may comprise a single type or more usually a number of different polypeptides.

Capsids, chaperones and condensates

The earliest type of protein assemblies observed by electron microscopy were the 'capsids' which encase the genomes of many viruses. polypeptides in capsids generally form regular polygons: the symmetry arising from packing together many similar polypeptides. In bacteria, assemblies of polypeptides form the ''outboard motors' which give these single cell organisms motility.

Notable structures formed by polypeptide self association are the chaperonins which form hollow chambers where damaged or nascent proteins can fold or refold. Also proteosomes where proteins enzymically tagged as 'end of use' can be disassembled, cut up by proteases and the amino acids recycled. A more recently identified structure within human and other animal cells is the 'vault' illustrated in the introduction to this article. There are many more instances, known and probably yet-to-be-discovered molecular machines are built up of dissimilar proteins, with different functions. Very well understood examples of molecular machines include the proteins involved in the replication and regulation of DNA transcription and the separation of daughter chromosomes in mitosis.

Proteinaceous compartments

Prokaryotes carry out most of the functions listed in the earlier definition of life within the cell cytoplasm. The more complex eukaryotic cells require compartmentation of functions. This is addressed by many different membrane enclosed organelles (the mitochondrion, nucleus, lysosomes etc.). But some important eukaryotic functions occur within membrane-less regions called biomolecular condensates (the nucleolus is one example). Proteins and RNA within such regions assemble through multivalent association including most of the types listed in the earlier section on protein interactions.

These regions are said to be liquid-liquid phase separated. Diffusion in and out of these regions is not restricted by membranes, which require transport systems for proteins or RNA to cross them. Maybe the biomolecular condensates represent a legacy from the earlier prokaryotes which lack organelles. The intrinsic disorder of polypeptides which are involved in biomolecular condensates is important to their formation. See: A Framework for understanding functions of biomolecular condensates on molecular to cellular scales Lyon A. et al.At least one of the annexins - AnxA11 - participates in biomolecular condensates. AnxA11 has a very extended amino terminus rich in proline residues which leads to a large disordered region (IDR).

Modification of polypeptide amino termini is especially important in biological functions

Intrinsically disordered regions of proteins can be sites of enzymatic addition or removal of phosphate or other chemical groups (known as post-translational modification. By this means cells exert regulatory control on proteins or protein assemblies. The amino termini are frequently subject to such modification. Amino termini of histone proteins which bind and compact segments of DNA within chromosomes are modified by acetylation and other covalent change. This contributes to theepigenetic regulation of gene expression. Amino terminal sequences are also frequently subject to alteration by alternative splicing of the mRNA which specifies them. Addition or removal of sections of polypeptide chain in this manner changes the functions, distribution or interaction with other biopolymers in cells. This is especially seen in the annexins. This family of proteins is highly 'promiscuous' in biopolymer interactions through post-translational modification and alternative splicing!

Intrinsically disordered regions may also be sites of enzyme cleavage of the polypeptide chain. Many proteins are initially synthesised with 'signal peptide sequences' which are recognised as transport addresses directing proteins to subcellular compartments or secretion pathways. These peptides are excised during transport. Other proteins - especially hormones directed to secretion pathways - have 'propeptide sequences', which are similarly subject to enzymatic removal prior to release from the cell.

How many natural proteins has evolution created?

Most natural proteins range in size between 80 and 1000 amino acids (often called residues by biochemists), but there is a noticeable size peak at ~340 residues which corresponds with the sizes of the large family of receptor proteins which sit in our cell membranes. Working out an estimate for the number of different protein sequences is a combinatorial exercise. Taking a polypeptide chain of a modest 200 amino acids, as an example, since 20 different amino acids can be present at each point, there are theoretically 1.6 x 10²⁶⁰ different natural polypeptides.Even given such a vast theoretical number of unique polypeptide sequences, there is further diversity in nature to be accounted. Over 200 'post-translational' chemical modifications have been recognised. These include the attachment of chemical groups like phosphate or sugars.

Prediction of polypeptide and protein 3D folding

The stumbling block in polypeptide and 3D protein structure prediction (for many decades) was to determine how localised secondary structures (alpha helix and beta sheet) and interconnecting peptide chain associated to form the folded protein tertiary structure. It was evident that the amino acid sequence determined the folding of polypeptide chains. Many purified proteins could be unfolded and refolded with full regain of function. A race was on to try to determine the folding of proteins from their amino sequences. Success was seen as 'the holy grail' of protein science.

Prediction of regions of polypeptide chain which were likely to form secondary structures was tackled first with modest success by a number of groups. The approach assigned each amino acid a propensity (a preference) to be present in alpha, beta or no secondary structure. This was essentially a statistical method.

Structure Prediction - a blast from the past!

Following the lead of workers Chou and Fasman, on the 70s, I recalculated these amino acid propensities for all alpha or all beta protein classes of experimentally determined proteins. The values I obtained gave a modestly improved prediction of secondary structure. Much later, with Dr Willie Taylor ( then at Birkbeck College London), We were able to localise the alpha helical regions of the annexins I was working on Our helix locations were later shown to correspond reasonably well with the structure of annexin determined by X-ray crystallography (Image below). Prior to the publication of the structure of and annexin by Robert Huber in Germany, Willie and I then attempted to predict the 3D folding of one of the annexin domains using the latest prediction methods. We used the known 3D structure of the known Ca²⁺ binding proteins, which also having a bundle of alpha helices as template for our prediction. The Ca²⁺ binding site was at at turn between tow helices the called an 'EF hand' When the structure of rat Annexin A5 was published we could see that our prediction of a bundle of 5 alpha helices was correct, and or location of the Ca²⁺ binding site was also correct, but our estimate of the tertiary structure (the overall fold) was incorrect. In fact the annexins were found to belong to a new class of Ca²⁺ binding, proteins. We were close, but no cigar! As described later in this article, protein secondary and tertiary structure prediction has now become relatively routine.

Developments in AI have revolutionised protein structure prediction

An international competition CASP was held biennially where researchers submitted predicted 3D protein structures subsequent to publication of experimentally determined ones. Over decades these predictions improved slowly. The big breakthrough came with a computer programme AlphaFold. This approach by researchers from the Alphabet company Deep Mind used Machine Learning and their approach has been applied to structure all proteins from the sequences in public databases. The 3 researchers behind this advance were awarded the 2024 Nobel Prize in chemistry. From the deluge of research papers and open access resources following this, we might describe this AI breakthrough as a tipping point for protein structure determination and de novo protein design.

There are now a number of competitive offerings online, mainly open source, for predicting polypeptide folds from amino acid sequence. All methods are subject to continuous improvement through learning from ever-increasing databases of experimentally determined structures and physical realities of atom-atom interaction. As with the genetic code itself, there's some amino acid redundancy in folded proteins. Why do I claim this? Well in the rapid mutation rate seen in bacteria and viruses, through selection pressure, many changes of individual or even a number of amino acids do not destroy 3D protein structure or function and may even enhance infectivity. Protein engineers and designers have noted considerable tolerance of 3D structure to minor sequence changes.

1. Highly accurate protein structure prediction with AlphaFold2
2. Accurate prediction of protein structures and interactions using a three-track neural network [RoseTTAFold]
3. The Alphafold Protein Structure Prediction Database
4. The AlphaFold 3 server
5. RoseTTAFold All Atom protein-ligand structure prediction that can wide range of ligands and covalent amino acid modifications

Computational tools will be vital as we explore polypeptide space. Of course, many 'designer proteins' which might exist stably in vitro might be incapable of folding or otherwise not be viable in vivo where they have to interact with the intracellular gamut of biomolecules. Very many sequences in polypeptide space may also be impractible due to constraints such as insolubility. Most proteins presently have to be synthesised through expression of recombinant DNA in bacteria. Problems in the structure and stability of their nucleic acid templates. or inherant toxicity may also block the production of certain protein sequences. But amongst these 'designer polypeptides' will be molecules that evolution has not had time to express, but have valuable properties for both research and commerce. Some polypeptides are highly toxic to humans. Ricin is one of many examples. The DNA synthesis industry will have to monitor requests for certain RNA or DNA sequences which might have lethally toxic consequences. Commercial suppliers are already taking this threat into account.

From genomic DNA sequences, without the need to isolate proteins, we can infer the amino acid sequences of every life form from microbes to man. The human genome encodes the sequences of around 20,000 proteins, other life forms express even more. Ambitious initiatives to sequence the genomes of all lifeforms on earth are in progress.

Knowing the 3D structures of proteins is critical for understanding function and development of drugs. In just the last few years, we have learnt how to exploit machine learning (ML) to predict the folding of proteins and protein assemblies. We still presently need to confirm the 3D structures of proteins significant for drug development (where knowing precise structure at an atomic level is critical) by X-ray crystallography or cryo-electron microscopy. This has happened (for example) with structure determination of the 'spike protein' of SARS-Coronavirus-19 as part of the global effort to develop therapeutics for COVID.

Changes in amino acids at one or many more points in a polypeptide chain often doesn't alter the overall folding of a polypeptide. We've see many changes through genome mutations in the amino acid sequence of the coronavirus spike protein without loss of structure or function (infectivity). But some single amino acid changes may destroy function without altering the overal 3D folding. Mutation of a residue in an alpha helix to proline can disrupt this secondary structure. Loss of a single 'catalytic' amino acid (often serine or histidine) in proteolytic enzymes like trypsin renders the molecule unable to cut polypeptide chains.

Developments in Machine Learning have opened access to de novo polypeptide &binder design

Designing proteins and peptides for medicine

The earliest attempts at polypeptide design were targetted (unsurprisingly) at developing therapeutics. Right now this has become a huge and diverse endeavour by academic establishments, public sector research organisations, biotech and pharmaceutical industries. Some of the areas of pathology being addressed are listed here:

• Cancer
• Autoimmunity
• Infection and immunity
• Antibiotics and microbial resistance
• Rare conditions
• Dementias
• Neurodegenerative disease
• Aging

Many (but not all) of these conditions arise from protein dysfunction, inappropriate protein expression or exploitation of normal cell surface proteins for attachment or entry leading to infection by microorganisms and viruses. Certain pathologies previously considered 'undruggable' quot; are now subject to new approaches which include silencing or knockdown of expression of cell proteins through gene editing in vivo. BUT developments are now arriving so quickly would be difficult to provide anything apart from a brief summary here. It's immediately evident from the above list that manipulation of the proteins in the human immune system is one of the most fertile fields for protein and peptide design and the search for inhibitory drugs. The modular form of immunoglobins, especially IgG, has lent itself to the exporation and modification of antibodies through protein engineering.

De novo protein and protein binder design

To design amino acid sequences that will support a target 3D fold, the ProteinMpNN method has been developed. Together these deep learning techniques are being applied to protein and binder design projects with considerable success. In the polypeptide space section of this article, these methods are used with a minimal 'archaeic' amino acid set to map designed amino acid sequence onto chosen polypeptide folds.

Searching biological databases

The latest protein structure prediction, modelling and design technologies rely on the vast databases of molecular biological data as input to AI neural networks (Machine Learning). This approach gave rise to AlphaFold protein structure prediction software which learnt from the databases of experimentally determined protein structure. AI-assisted searching of genome sequence, protein sequence and taxonomic data has led to the development of a search engine: 'Google for DNA' brings order to biology's big data & MetaGraph: Petabase-Scale Search for Genomics & Beyond
Making sense of the global, dispersed and variously presented biological data to apply to the discovery and development of potential therapies will require powerful approaches of this kind to bring molecular and macroscopic knowledge together. But a note of caution is required here. As AI generated knowledge is added, biases could easily to be introduced into the biological databases used as the teaching resources for that same AI.

Tools for searching the biological databases are proliferating rapidly and it may be hard for researchers to focus on the best (for them) technologies. As an example of a recently released Natural Language search model I include: BioModelBench which aims to support intelligent searches for computational biology models, tools and biodata sets.

Developing novel molecular interactions: protein co-folding and binder discovery and design

New therapeutics often require the development of small or large molecular binders (a.k.a drugs) which bind to protein targets with nanomole to picomole affinity. This of course has been the basis of the pharmaceutical industry's endeavours from the start. Drug screening in in vitro and in vivo assay has long been the gold standard for such drug development. We have entered an era where the binders can themselves be peptides or polypeptides. Computational approaches employing known 3D structures had formerly relied on molecular dynamics (MD) to fit binders to proteins. Now we have generative protein and peptide design. This can be applied to design protein or peptide binders for a very diverse range of targets including 'minibodies' , peptides and small organic drugs. This is very reliant upon the data from which designer AI can learn. But ligands interact with protein targets in many ways. There are numerous examples of experimentally determined protein structures in the PDB with bound ligands at 'active' (orthosteric) sites, but few examples available of protein structures with ligands bound at 'inhibitor' (allosteric) sites. Researchers often comment that there are vastly more solved protein structures with ligands bound at both orthosteric and allosteric sites in the possesion of 'Big Pharma', but these are presntly not public domain, relating as these do to the core businesses of the pharmaceutical industry.

• BoltzGen Generative protein binder programme
• BoltzGen: Towards Universal Binder Design
• Protein Hunter: exploiting structure hallucination within diffusion for protein design
• Co-folding, the future of docking: prediction of allosteric and orthosteric ligands
• The Continuing Evolution of Protein-Ligand Co-Folding Methods

Peptide framework design

How many unique polypeptides might there be in a synthetic (man made) polypeptide space?

Even in the more than 3 billion years of life on earth, it seems unlikely that there's been time for evolution to access more than a minute fraction of potential polypeptide space. This raises the question: what applications might we make given the rapidly increasing number of ways to synthesise polypeptides now available?

Proteins already exist with highly useful thermodynamic, enzymatic and therapeutic properties - the DNA polymerase enzymes used in Polymerase Chain Reaction (PCR) to give just one example. Birds appear to have sensory proteins (cryptochromes) in their eyes which can detect earth's magnetic field to use as aids to navigation. Lifeforms exist which have proteins which generate light or electric currents. But can we use novel proteins and peptides in medicine? Apart from the challenge of immune response to novel entities introduced into our bodies, there is the problem of stability: polypeptides are subject to metabolism including proteolytic attack and rapid clearance by filtration in the kidneys. N-acetylation or cyclisation has been used to minimise peptide ligand loss. What if we built some or all of potential binders with D-amino acids, which would be resistant to proteolytic attack and have very different pharmacokinetics?

The inventory of polypeptide structures and functions is already very long, but may represent a drop in the proteinaceous ocean so far. So how might we enter polypeptide space ourselves and 'boldly go where nature has not been before'?

For general polypeptide design, we aren't restricted to evolution's choice of amino acids!

Going back in deep time - when there were only the ancestors of today's prokaryote micro-organisms, we don't really know when the familiar 20 amino acid building blocks of proteins came on the scene. We don't even know if their precursors were the same: going back that far in time, no chemical evidence remains. With a degree of inference, it has been assumed that a small number of amino acids were available during the early development of life. In an intriguing study, a research team asked if proteins designed with a very limited amino acid 'alphabet' 'Ancient folds' could yeild stable structures and possess functions. They selected a limited 'alphabet' of 10 amino acids, rimarily small, polar, and aliphatic residues such as the canonical Gly, Ala, Asp, Glu, Val, Ser, Ile, Leu, Pro, and Thr. So important residues in modern proteins like aromatic and basic side chains were absent. The minimal amino acid set was encoded in DNA in bacterial expression systems. The authors used the ProteinMpNN and RFdiffusion methods described in the section on protein design to generate sets of amino acid sequences that would fit into template 3D folds. Using a variety of structure assessing methods, the reduced alphabet proteins were found to have stable folds, adopting the designed fold and demonstrating good chemical and thermal stability. A metal binding function was created. Giacobelli et al. The minimal amino acid set would not serve all the functions of modern proteins, but does suggest that an evolutionary path to today's proteins was perfectly possible.

A recent article described proteins that assemble and disassemble on command, using an incorporated light-sensitive amino acid phenylalanine-4′-azobenzene 𝗔𝘇𝗼𝗙. They describe a computational approach to design protein–protein interactions regulated by non-canonical amino acids, focusing on the light-responsive phenylalanine-4′-azobenzene (AzoF). Using this approach, light-responsive cyclic homo-oligomers and heterodimers, which only assemble in AzoF’s trans configuration and disassemble when AzoF photoisomerizes to the cis configuration. Biophysical characterization confirms the light-responsive assembly and disassembly of these complexes, and the crystal structures match the design models with atomic accuracy. These light-responsive proteins can be used in constructing light-responsive hydrogels and engineering synthetic ligand receptors to optocontrol cell signalling in mammalian cells.

If we fully venture into synthetic polypeptide space, around 500 diferent amino acids are known to date. Considering these as potential monomeric units, the number of possible polypeptide molecules would be essentially, infinite. The huge challenge would be to identify small subsets of polypeptides which would teach us about biological function or have meaningful applications. I have no doubt that to access such vast molecular spaces, machine Learning would be essential here.

We might intuitively visualise a vast type of library (described in the literature as the 'Library of Babel') where the books could carry a different polypeptide sequence on every page, differing by one residue from the sequence on the neighbouring page. In this imaginary library, every known natural polypeptide would be present in books on the shelves. Such a library would also carry a vast number of polypeptide sequences which have valuable functions, not present in nature, or still remain undiscovered. But without a 'map' it could take many lifetimes (and enormous funds!) to locate them.

To represent this vast collection of molecules as a space that might be explored and manipulated we have to turn to maths and extremely fast computers. A mathematically interrogatable polypeptide space might be multidimensional with a dimension for at least each of the 20 common amino acids and chain length. First imagine all natural protein sequences (known by experiment or deduced from genome sequences) of up to, say 1000 residues) positioned as points in such a multidimensional (tensor) space. It would be expected that large regions of this space will be unoccupied or very sparcely populated. Homologous amino acid sequences differing in one or a few residues would be closely grouped. Fibrous proteins with highly repetitive sequences (collagen, spiderin) would appear in regions of polypeptide space remote from the bulk of globular proteins. Others might either be closely grouped (leaving vast regions of polypeptide space empty) or randomly distrubuted throughout. Just visualising existing proteins mapped into this space could be revealing about the rules and options of nature to date!

References

I have provided hyperlinks within this article to selected papers, reviews and resources about proteins. The ones I have included are mostly open source. This is such a vast field it's impracticable to even partly cover the main aspects of it. Here I have (immodestly) given references to my own published research on the proteins I worked with that I have used as examples in this review.

1. Structure of Prealbumin [transthyretin]: Secondary, Tertiary and Quaternary Interactions Determined by Fourier Refinement at 1.8 Å Blake C.C.F., Geisow M.J., Swan I.D.A., Oatley S.J. Rérat B. & Rérat C. (1978) J. Mol. Biol. 121, 339-356.
2. Physical and binding properties of large fragments of human serum albumin". Geisow M.J. & Beaven G.H. (1977) Nature 270, 476-477.
3. Polypeptide secondary structure may direct the specificity of prohormone conversion. Geisow M.J. (1978) FEBS Lett 87, 111-114.
4. Proteolysis of prohormones and proproteins. Geisow M.J. (1980) in: Post-translational Modifications of Proteins. (Freedman R.B. ed.) Academic Press.
5. Amino acid preferences for secondary structure vary with protein class. Geisow M.J. & Roberts R.D.B. (1980) Int. J. Biol. Macromol., 2, 387-389.
6. Annexins - a new family of Ca2+-regulated phospholipid-binding proteins. Geisow M.J., Walker J.H., Boustead C. & Taylor W. (1987) Biosci. Rep. 7, 289-298.
7. Predicted structure for the calcium-dependent membrane-binding proteins p35, p36 and p32" Taylor W. & Geisow M.J. (1987) Protein Engineering, 1, 183-187.