Computational Biology

(BIOSC 1540)

Mar 25, 2025

Lecture 11A

Protein structure prediction

Foundations

Announcements

Assignments

  • P02B is due Mar 28
  • P02C is due Mar 28
  • P03A is due Apr 4

Quizzes

  • The final exam is on Monday, Apr 28, at 4:00 pm in 244 Cathedral of Learning

Final exam

After today, you should have a better understanding of

Identify what makes structure prediction challenging

Protein structure is essential for understanding biological function

Proteins are molecular machines; their 3D shape determines how they interact with substrates, DNA, other proteins, etc.

Experimental methods for structure determination are powerful but limited

As of 2021, >200 million protein sequences exist, but <200,000 structures are known​

X-ray crystallography, NMR, and cryo-EM provide high-resolution data.

However, these methods are time-consuming, expensive, and often fail for specific proteins.

Protein structure prediction fills a crucial gap in biological discovery

Enables structural understanding of sequences with no experimental structure. This accelerates many research fields and democratizes access to atomistic insights

Example: Our collaborators (Dr. Cahoon) crystallized Lm PrsA1 in 2016, but we need a structural model of PrsA2. Instead of potentially years, AlphaFold 3 gives us a decent prediction within minutes

PrsA1 (X-ray)

PrsA2 (AF3)

Protein folding is computationally hard due to the vast conformational space

Proteins fold in milliseconds—implying nature doesn't sample all conformations.

Levinthal’s Paradox: A protein can’t sample all conformations in a biologically reasonable time, yet it folds quickly

Example: A protein with 100 amino acids, each capable of adopting about 3 torsion angles, results in ~         possible conformations3100

3^{100}

Proteins fold into their native structure by minimizing free energy

Scoring functions attempt to model this with statistical or physics-based potentials​

Proteins adopt conformations that minimize thermodynamic free energy

A potential energy surface (PES) represents the energy of a system as a function of the positions of its atoms

Allows us to understand how the system's energy changes upon reactions or movements

Flexible and disordered regions add complexity to structure prediction

Many proteins exist in ensembles of structures or are natively disordered.

Their function may depend on transient interactions or induced folding.

Environmental context dramatically impacts protein folding

Proteins fold differently in different environments

Predictions need to capture interactions with solvent molecules, ions, and cofactors

7MHX

Example: Predicting transmembrane protein structures, where the lipid bilayer plays a key role in folding, is particularly complex.

AlphaFold 3

pH-gated K+ channel

Environmental context dramatically impacts protein folding

PTMs such as phosphorylation, glycosylation, and methylation can alter protein folding and function

Example: eIF4E is a eukaryotic translation initiation factor involved in directing ribosomes to the cap structure of mRNAs

Ser209 is phosphorylated by MNK1

AlphaFold 3 accurately predicts these changes when they are already known

After today, you should have a better understanding of

Homology modeling

Hidden Markov Model alignments

Homology modeling predicts structure using evolutionary relationships

Requires a template—a known structure with detectable sequence similarity.

Often the first modeling strategy attempted due to simplicity and reliability.

Based on the principle that proteins with similar sequences tend to adopt similar structures.

The first step in homology modeling is to search for similar sequences

You begin with a query sequence
(the protein you want to model)

> PrsA2
CGGGGDVVKTDSGDVTKDELYDAMKDKYGSEFVQQLTFEKILGDKYKVSDE
DVDKKFNEYKSQYGDQFSAVLTQSGLTEKSFKSQLKYNLLVQKATEANTDT
SDKTLKKYYETWQPDITVSHILVADENKAKEVEQKLKDGEKFADLAKEYST
DTATKDNGGQLAPFGPGKMDPAFEKAAYALKNKGDISAPVKTQYGYHIIQM
DKPATKTTFEKDKKAVKASYLESQLTTENMQKTLKKEYKDANVKVEDKDLK
DAFKDFDGSSSSDSDSSK

Sequence alignment maps residues from the target onto the template structure

This is an MSA of Listeria monocytogenes PrsA2 to related proteins

Basic alignment algorithms are too simplistic for distant homolog detection

Methods like Smith-Waterman use direct pairwise alignment based on similarity scores (e.g., BLOSUM62).

They do not consider evolutionary variation, insertions, or residue-level probabilities.

We need methods that detect evolutionarily distant but structurally conserved relationships.

Profile-based aligners improve sensitivity by modeling residue variability at each position

A profile captures how conserved each position is across an MSA

Instead of a single residue, each position becomes a probability distribution over all 20 amino acids.

Made with skylign

HHblits starts by converting the query sequence or MSA into a profile HMM

A profile HMM models the amino acid probabilities at each position, plus insertion and deletion likelihoods.

The result is a probabilistic model that captures both conservation and structural variability.

Sequence likelihood can be computed by walking along the profile HMM

HHblits performs accurate HMM–HMM alignments on the best candidates

Full alignments are done using the Viterbi algorithm to find the best path through the HMM state space.

A maximum accuracy (MAC) alignment is also computed to optimize for correct residue–residue matches.

These alignments return E-values to estimate match confidence.

Only statistically significant hits (e.g., E < 1e-3) are retained for the next iteration.

The alignment is represented as red path through both HMMs

Homology modeling is most accurate when sequence identity is >30%

>50% identity: high-accuracy models (~1 Å RMSD) are achievable​

Between 30–50%: moderate accuracy; errors appear in loops, side chains

<30% identity: The "twilight zone" where structural similarity is uncertain

HHblits compares HMMs using a fast, two-step prefilter before full alignment

Step 1: Converts HMM columns into a discretized profile alphabet of 219 letters to speed up comparison.

Step 2: Performs a fast local alignment of the query HMM to these compressed database representations.

These steps eliminate ~99.9% of irrelevant comparisons—reducing millions of alignments to thousands​

HHblits iteratively refines the profile HMM by adding new homologs from each search round

Sequences from matched HMMs are added to the query MSA.

A new query HMM is built from this updated MSA.

Each iteration improves sensitivity by capturing more distant, diverse sequences.

This process continues for 1–4 rounds, depending on the desired depth and quality.

After today, you should have a better understanding of

Homology modeling

Template building

The model is built by copying the template structure and modeling variable regions

Conserved regions: Backbone atoms are copied from template directly.

Variable regions (loops, inserts): Built using fragment libraries or loop modeling algorithms.

Side chains are adjusted using rotamer libraries to fit target sequence.

Model refinement improves geometry and resolves steric clashes

After model construction, the structure often contains bad bond angles, clashes, or unrealistic torsions.

Refinement includes energy minimization using force fields or statistical potentials.

Some tools use molecular dynamics or Monte Carlo sampling.

Model validation helps assess confidence and detect errors

Ramachandran plots visualize backbone torsion angles.

Statistical scores (e.g., DOPE, QMEAN, GA341) evaluate nativeness.

Residue-by-residue assessment helps identify weak regions (e.g., VERIFY3D, ERRAT).

Good models have:

  • Most residues in favored Ramachandran regions

  • Low-energy scores

  • No large clashes

Homology modeling works best when you iterate and re-evaluate

If a model fails validation, revisit earlier steps:

  • Try a different template

  • Refine the alignment

  • Adjust loop modeling parameters

Multiple models are often built and ranked—choose the one with the best validation metrics.

SWISS-MODEL

swissmodel.expasy.org

MTLSILVAHDLQRVIGFENQLPWHLPNDLKHVKKLSTGHTLVMGRKTFESIGKPLPNRRNVVLTSDTSFNVEGVDVIHSIEDIYQLPGHVFIFGGQTLFEEMIDKVDDMYITVIEGKFRGDTFFPPYTFEDWEVASSVEGKLDEKNTIPHTFLHLIRKK

DHFR (UniProt)

SWISS-MODEL

Novel proteins are too challenging

After today, you should have a better understanding of

Know when to use threading instead of homology modeling

Why Use Threading?

In cases where sequence similarity to known structures is low (< 30%), homology modeling becomes unreliable

Phyre2, RaptorX, MUSTER, and I-TASSER are commonly used for threading and takes much longer than homology modeling

Threading matches sequences to known structural folds based on structural rather than sequence similarity

Identifying the Right Fold

After today, you should have a better understanding of

Interpret a contact map for protein structures

Contact Maps Visualize Residue Interactions in Proteins

A contact map is a 2D representation of which residues are in close proximity

Each point on the map corresponds to two residues that are close in 3D space

mapiya.lcbio.pl

Contact Maps Represent Spatial Proximity, Not Sequence Order

Contacts are determined by spatial proximity, typically within a certain distance threshold

Residues far apart in the sequence can still be close in the 3D structure, reflected in the contact map

Residues on the diagonal are adjacent in sequence (and spatially)

After today, you should have a better understanding of

Comprehend how coevolution provides structural insights

The Rise of Machine Learning in Structural Biology

Traditional methods like homology modeling and threading rely on templates and known structures

AlphaFold (DeepMind) and RosettaFold (Baker Lab) lead the charge in this area

ML predicts 3D structures only from sequence data

What is AlphaFold?

Developed by DeepMind, AlphaFold predicts protein structures with atomic accuracy by using deep learning models trained on large structural datasets

Breakthroughs

  • AlphaFold 2 achieved near-experimental level accuracy in the 2020 CASP14 competition (Critical Assessment of protein Structure Prediction)
  • AlphaFold 3 (2024) predicts proteins, DNA, RNA, ligands, and post-translational modifications

Coevolving residues mutate in a correlated manner

Mutations in one residue often result in compensatory mutations in its interacting partner

This is observed across species through analysis of homologous protein sequences

Correlated mutations indicate functionally significant residue pairs

Arg (positive)

Asp (negative)

Lys (positive)

Glu (negative)

Trp (hydrophobic)

Val (hydrophobic)

Evolution

Evolutionary Analysis Reveals Structural Insights

Coevolution analysis helps predict which residues are close in the 3D structure

Residues showing correlated mutations are likely to be spatially close in the folded protein

This is particularly useful when no experimental structure is available

Multiple Sequence Alignments Enable Coevolution Detection

Coevolution is detected using large MSAs from homologous proteins

The more diverse the sequences in the MSA, the better the resolution of coevolving residues

Evolutionary information from MSAs guides predictions for residue-residue contacts

evcouplings.org

Coevolution example: DHFR

Residues with a high Score (i.e., coevolve) are near each other in the protein's structure (i.e., small distance)

Val14 and Gly120 coevolved

Models predict these residues are spatially close

evcouplings.org

Coevolutionary signals can be noisy

Not all correlated mutations are due to direct physical interactions; some may be indirect

Noise in the data can come from random mutations or insufficient evolutionary diversity.

Large and diverse sequence data sets are needed for reliable coevolution predictions.

Machine learning leverages coevolution for high-accuracy predictions

AlphaFold and RosettaFold utilize coevolutionary data from MSAs to predict residue interactions

These models incorporate evolutionary information along with structural features, leading to highly accurate predictions

After today, you should have a better understanding of

Explain why ML models are dominate protein structure prediction

AlphaFold pipeline, simplified

Given the following data

MTLSILVAHDLQRVIGFENQLPWHLPNDLKHVKKLSTGHTLVMGRKTFESIGKPLPNRRNVVLTSDTSFNVEGVDVIHSIEDIYQLPGHVFIFGGQTLFEEMIDKVDDMYITVIEGKFRGDTFFPPYTFEDWEVASSVEGKLDEKNTIPHTFLHLIRKK

Input sequence

Multiple Sequence Alignment

Predict

Atomistic structure

ML models

AlphaFold 2 pipeline: Evoformer

Using MSAs and contact maps, DeepMind trained a model to predict protein structures

Contact maps are converted into dihedral angles

AF2 iterations

What is new in AlphaFold 3?

Biggest change is the use of a diffusion model

Diffusion models essentially learn to unscramble atoms into a structure

AlphaFold 3 is supercharged for any biomolecule

Proteins, DNA, RNA, ligands, PTMs, protein-proteins, etc.

AlphaFold 3

MTLSILVAHDLQRVIGFENQLPWHLPNDLKHVKKLSTGHTLVMGRKTFESIGKPLPNRRNVVLTSDTSFNVEGVDVIHSIEDIYQLPGHVFIFGGQTLFEEMIDKVDDMYITVIEGKFRGDTFFPPYTFEDWEVASSVEGKLDEKNTIPHTFLHLIRKK

DHFR (UniProt)

alphafoldserver.com

MGKKEVILLFLAVIFVALNTLVVAVYFRETADEQVVYGKNNINQKLIQLKDGTYGFEPALPHVGTFKVLDSNRVPQIAQEIIRNKVKRYLQEAVRIEGTYPIVDGLVNAKYTVANPNNLHGYEGFLFKDNVPLTYPQEFILSNLDGKVRSLQNYDYDLDVLFGEKEEVKSEILRGLYYNTYTRAFSPYKL

Novel protein (ChatGPT)

AlphaFold 3 is a breakthrough, not the final solution

Caveat: Proteins are dynamic

What about intrinsically disordered proteins?

At least 40% of proteins have disordered regions

LARP1

AlphaFold (and all other methods) struggle with disordered regions

Before the next class, you should

Lecture 11B:

Protein structure prediction -
Methodology

Lecture 11A:

Protein structure prediction -
Foundations

Today

Thursday

A Markov model predicts outcomes based on transitional probabilities

Suppose I collect weather data in Pittsburgh for the past 30 days: Sunny, Cloudy, or Rain

I want to figure out how to predict tomorrow's weather based on today's

Today's weather

Tomorrow's weather

Transition probability

Example: If today is cloudy, there is a 57% chance it will be Sunny tomorrow

We can represent these states and probabilities as a (cursed?) graph

Each edge represents the probability of transitioning from one state to the next

Hidden Markov models also include additional information in "hidden states"

Suppose my friend lives in a remote location where it is either Rainy or Sunny

I cannot look up the weather but I have last year's weathers reports

  • Walking
  • Shopping
  • Cleaning

We know how weather patterns transitions, but we don't have this information from our friend

Obervables

Hidden states

Note: If we had previous observable data, we could fit/learn transition probabilities of hidden states

My friend can only tell me

Hidden Markov Models (HMMs) Capture Evolutionary Patterns in Proteins

HMMs are statistical models representing sequences using probabilities for matches, insertions, and deletions

Essentially more robust alignments

HMMs Model Protein Sequences as a Series of Probabilistic States

Hidden states represent the underlying biological events that are not directly observable

Observables are the actual amino acids (residues) in the protein sequence that we can observe

  • Match states: conserved positions in the sequence
  • Insertion states: positions where extra residues are added
  • Deletion states: positions where residues are missing

SWISS-MODEL

SWISS-MODEL

What happens with a novel protein?

MGKKEVILLFLAVIFVALNTLVVAVYFRETADEQVVYGKNNINQKLIQLKDGTYGFEPALPHVGTFKVLDSNRVPQIAQEIIRNKVKRYLQEAVRIEGTYPIVDGLVNAKYTVANPNNLHGYEGFLFKDNVPLTYPQEFILSNLDGKVRSLQNYDYDLDVLFGEKEEVKSEILRGLYYNTYTRAFSPYKL

Novel protein (ChatGPT)

Prediction methods depend on known structures and evolutionary information

Example: AlphaFold has made strides, but predicting de novo structures remains challenging, especially for proteins with no templates

Our predictions rely on similarity to known structures, but novel sequences or folds (for which no homologous structures exist) are difficult to predict accurately