Understanding protein sequence analysis, structure prediction and their role in drug discovery
By Dr. Neil Taylor
Proteins do almost everything in a living cell. They carry signals, catalyse reactions and give tissue its shape. Most drugs work by binding to a protein and changing what it does. To design a drug well, it helps enormously to know what the target protein looks like in three dimensions.
That is what a protein model provides. This article explains the basics of protein structure modelling: how a protein's shape is described, how it is determined or predicted, and where AI tools like AlphaFold fit into modern drug discovery.
From protein sequence to three-dimensional structure
Every protein starts as a chain of amino acids. There are 20 common amino acids, and the order in which they appear is the protein's sequence. Protein sequence databases such as UniProt now hold hundreds of millions of these sequences.
A sequence alone is a one-dimensional string of letters. The protein only becomes functional when that chain folds into a specific three-dimensional shape. Biologists describe protein structure at four levels:
- Primary structure: The amino acid sequence itself.
- Secondary structure: Local, repeating patterns. The most common are the alpha helix, a tight coil, and the beta sheet, where strands of the chain line up side by side.
- Tertiary structure: How the helices, sheets and loops pack together into the complete folded protein.
- Quaternary structure: How several folded chains assemble into a larger complex.
Protein sequence analysis and secondary structure prediction
Protein sequence analysis was the first window into structure. By comparing a new sequence with known ones, scientists can identify its protein family using a protein family database such as Pfam or InterPro. Proteins in the same family usually share a similar fold and often a similar function.
Secondary structure prediction goes a step further. It estimates, residue by residue, whether each part of the chain is likely to form a helix, a strand or a loop.
Early methods relied on simple statistics about which amino acids prefer which shapes. Modern predictors use machine learning and reach accuracies of around 80 to 85%. These predictions are useful, but they do not tell you how the whole protein folds.
Determining protein structures experimentally
For decades, a reliable 3D protein model meant an experiment. The main methods for experimental protein structure determination are:
- X-ray crystallography: The protein is crystallised and its diffraction pattern used to calculate an atomic model. This is still the workhorse of structure-based drug design because it can show a drug molecule sitting in its binding pocket.
- Cryo-electron microscopy (cryo-EM): Frozen protein samples are imaged directly. It has transformed the study of large complexes and membrane proteins that are hard to crystallise.
- NMR spectroscopy: Useful for smaller, flexible proteins in solution.
The public Protein Data Bank (PDB) now holds more than 200,000 experimental structures. Pharmaceutical companies hold many more of their own, often showing their target proteins bound to proprietary compounds.

What AlphaFold changed
In 2020, DeepMind's AlphaFold 2 showed that AI could predict a protein's 3D structure from its amino acid sequence with accuracy often close to experiment. The AlphaFold Protein Structure Database has since made predicted models available for more than 200 million proteins. The work was recognised with the 2024 Nobel Prize in Chemistry.
AlphaFold 3, released in 2024, extended AlphaFold modelling to complexes that include DNA, RNA and small molecules. Open models built on similar ideas have followed, including Boltz, Chai and OpenFold.
For structure-based drug discovery, this means a reasonable model of almost any target can now be available from day one, before any experiment.
What protein models still can't tell you
AI-predicted protein structures are a starting point, not the final word. A few limitations are worth keeping in mind:
- Confidence varies. AlphaFold reports a per-residue confidence score. Flexible loops and disordered regions are often low-confidence, and these can sit right next to a binding site.
- Proteins move. A single predicted structure is a snapshot. Many drug targets change shape when a molecule binds.
- Small-molecule detail matters. The precise position of a ligand, water molecules and subtle side-chain shifts can decide whether a compound is potent. Experimental structures remain the gold standard here.
In practice, the most productive research teams use predicted and experimental protein structures together. Predictions guide early work and fill gaps; experiments confirm and refine.
Bringing it together
As protein models multiply, the challenge moves from obtaining a structure to managing many of them.
Researchers need to compare predicted and experimental models, track which ligands bind where, and search consistently across thousands of entries.
Effective protein structure data management is therefore an increasingly important part of structure-based drug discovery, helping research teams turn growing volumes of structural information into useful scientific insights.
About DesertSci
DesertSci's Proasis platform helps pharmaceutical research teams store, standardise and analyse protein structures and protein-ligand complexes at scale. It is used by more than 60% of the top 25 global pharmaceutical companies.
Learn more at desertsci.com and follow Dr Neil Taylor on LinkedIn for more insights.



