AlphaFold2 & ColabFold: Next-Gen Protein Structure Prediction
For CS 7150

Why Is Protein Structure Important?

Proteins are the molecular machines that carry out most cellular functions, from catalyzing biochemical reactions to providing structural support. Their function is intimately tied to their three-dimensional structure - the precise spatial arrangement of atoms that determines how a protein interacts with other molecules. Understanding protein structures is crucial for:

• Understanding disease mechanisms at the molecular level
• Rational drug design and development of therapeutics
• Engineering proteins with novel functions for biotechnology
• Deciphering biological processes and cellular machinery

The classic paradigm "sequence determines structure determines function" highlights why predicting structure from amino acid sequence has been a grand challenge in molecular biology for over 50 years.

What are the Experimental Methods of Getting Protein Structure?

Traditional experimental methods for determining protein structures include:

• X-ray Crystallography: The gold standard for high-resolution structures (often <1.5Å resolution), but requires protein crystallization which can be difficult for many proteins
• Nuclear Magnetic Resonance (NMR) Spectroscopy: Works with proteins in solution but typically limited to smaller proteins (<50kDa)
• Cryo-electron Microscopy (cryo-EM): Recent advances allow near-atomic resolution, especially valuable for large complexes
• Small-angle X-ray Scattering (SAXS): Provides low-resolution shape information

These methods are time-consuming (often taking months to years), expensive, and technically challenging, creating bottlenecks in structural biology research. This gap between known protein sequences (~200 million) and experimentally determined structures (~180,000) motivated computational prediction approaches.

What is CASP (Critical Assessment of Structure Prediction)?

CASP (Critical Assessment of Structure Prediction) is a biennial community-wide experiment established in 1994 to objectively evaluate the state of the art in protein structure prediction. Key aspects include:

• Blind prediction format: participants predict structures of proteins whose experimental structures are determined but not yet published
• Standard metrics: primarily GDT_TS (Global Distance Test) and RMSD (Root Mean Square Deviation) to quantify prediction accuracy
• Diverse targets: ranging from easy comparative modeling cases to difficult free modeling scenarios
• Independent assessment: ensures unbiased evaluation of different methods

CASP has been the gold standard benchmark for progress in the protein structure prediction field for nearly three decades, chronicling incremental advances until AlphaFold2's breakthrough.

How Did AlphaFold2 Revolutionize Prediction?

AlphaFold2 represented a paradigm shift in protein structure prediction through several key innovations:

• End-to-end differentiable architecture: Unlike previous approaches that combined disparate methods, AlphaFold2 integrated all components into a single neural network
• Attention mechanisms: Leveraged transformer architectures to capture long-range dependencies and evolutionary patterns between residues
• Structure module: Directly predicted 3D coordinates through iterative refinement rather than assembling fragments
• Evolutionary information: Effectively leveraged multiple sequence alignments to extract co-evolutionary signals
• Self-confidence scoring: Developed pLDDT metric to assess prediction quality without ground truth

The result was a system that achieved unprecedented accuracy in CASP14 (2020), with a median GDT_TS of 92.4 for all targets and even achieving competitive results for the most difficult free modeling cases. This performance approaches experimental accuracy, effectively solving a 50-year-old grand challenge in biology.

Team at AlphaFold

The AlphaFold breakthrough came from DeepMind, a London-based AI company (acquired by Google in 2014). The interdisciplinary team combined expertise from machine learning, structural biology, physics, and bioinformatics:

• John Jumper: Lead researcher and first author of the AlphaFold2 paper
• Demis Hassabis: CEO and co-founder of DeepMind, with backgrounds in neuroscience and AI
• Key team members: Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, and many others

The team collaborated with structural biologists and validated their approach through rigorous testing before the CASP14 competition. The project demonstrates the power of interdisciplinary research combining deep learning expertise with domain knowledge in biology to solve complex scientific problems.

AlphaFold2 Architecture: A Deep Dive

The AlphaFold2 architecture represents a revolutionary approach to protein structure prediction, combining evolutionary information with deep learning to achieve unprecedented accuracy. Unlike previous methods that relied on fragment assembly and physics-based simulations, AlphaFold2 uses an end-to-end neural network approach to directly predict 3D coordinates.

Key Components

1. Multiple Sequence Alignment (MSA) Processing

The first critical component of AlphaFold2 is the processing of evolutionary information through multiple sequence alignments:

• Searches genetic databases to find evolutionarily related sequences
• Aligns these sequences to capture conservation patterns
• Generates a rich representation that captures which amino acids tend to co-evolve
• This evolutionary information is crucial, as co-evolving residues often indicate spatial proximity in the folded structure

2. Template Processing

While not always necessary for accurate predictions, AlphaFold2 can use information from known protein structures:

• Searches structural databases (PDB) for proteins with similar sequences
• Extracts distance maps and structural features from these templates
• Integrates this information with sequence-based predictions
• This component helps guide the model, especially for proteins with known homologs

3. Evoformer

The core of AlphaFold2's architecture is the Evoformer, a novel transformer-based neural network:

• Contains 48 identical blocks that iteratively refine representations
• Processes both MSA and pairwise representations simultaneously
• Uses specialized attention mechanisms to update information bidirectionally between these representations
• Key mechanisms include:
  • Row-wise gated self-attention: processes evolutionary information across sequence positions
  • Column-wise gated self-attention: processes evolutionary information across aligned sequences
  • Triangle multiplication: models higher-order interactions between residues
  • Transition layers: non-linear processing that integrates information

4. Structure Module

The final stage of AlphaFold2 is the Structure Module, which converts refined representations into 3D coordinates:

• Consists of 8 blocks with shared weights
• Takes the representations from the Evoformer and gradually builds the 3D structure
• Uses Invariant Point Attention (IPA): a novel attention mechanism that respects the physics of 3D space
• Predicts backbone frames (rotations and translations)
• Computes atom positions based on these frames
• Predicts torsion angles to determine side chain conformations

5. Recycling and Confidence Estimation

AlphaFold2 employs two additional techniques that significantly improve its performance:

• Recycling: The entire prediction process is repeated multiple times (typically 3), with each iteration using the output of the previous one as input. This allows the model to refine its predictions based on emerging structural information.
• Confidence Estimation: AlphaFold2 provides two key metrics:
  • pLDDT (predicted Local Distance Difference Test): per-residue confidence scores from 0-100, with higher values indicating greater confidence
  • PAE (Predicted Aligned Error): estimates the expected position error between any two residues, providing insight into the reliability of domain arrangements in multi-domain proteins

This sophisticated architecture enables AlphaFold2 to achieve remarkable accuracy, with many predictions reaching experimental-level quality (GDT scores > 90). The integration of evolutionary information with geometric reasoning through deep learning represents a fundamental breakthrough in protein structure prediction.

Methods

Our approach utilizes ColabFold, an open-source adaptation of AlphaFold2 designed for accessibility and efficiency. Developed by Mirdita et al., ColabFold significantly reduces the computational barriers to protein structure prediction while maintaining AlphaFold2's accuracy. Key advantages of ColabFold include:

• Accelerated MSA generation using MMseqs2, reducing search times from hours to minutes
• Optimized implementation for Google Colab's free GPU resources
• Streamlined preprocessing and reduced memory requirements
• Support for both protein monomer and complex predictions
• Interactive visualization tools for structural analysis
• Batch processing capabilities for multiple sequences

ColabFold preserves the core AlphaFold2 architecture while making it accessible to researchers without specialized computing infrastructure. Our implementation follows the standard ColabFold pipeline, which employs a multi-stage neural network architecture consisting of:

1. Multiple Sequence Alignment (MSA) generation using MMseqs2 to capture evolutionary information
2. Template structure identification from the PDB database
3. Neural network processing through the Evoformer (48 blocks) which refines sequence representations
4. Structure module (8 blocks) that predicts backbone coordinates and side chain orientations
5. Confidence assessment using pLDDT (predicted Local Distance Difference Test) scores

For our experiments, we configured ColabFold with the following parameters:

• Model type: "auto" (automatically selects appropriate model based on input)
• MSA depth: 512 sequences for optimal runs, reduced to 2 sequences for limitation testing
• Extra MSA depth: 1024 sequences for optimal runs, reduced to 0 for limitation testing
• Number of recycles: 6
• Early stopping tolerance: 0.5 RMSD
• Use of templates: Enabled with PDB70 database for optimal runs, disabled for some limitation tests

Our experiments covered two distinct scenarios. First, we predicted the structure of lysozyme (129 amino acids), achieving very high confidence scores (pLDDT ~0.98) that compared favorably with experimentally determined structures (PDB ID: 1LYZ). The entire prediction process took approximately 1-2 minutes on a T4 GPU, with early stopping triggering after just 2 recycles due to structural convergence. Second, we conducted limit-testing experiments on a larger fragment (~1000 amino acids) with minimal MSA and no templates, which resulted in dramatically lower confidence predictions (pLDDT scores averaging 30-50%), demonstrating the critical dependence of AlphaFold2 on sufficient evolutionary and template information, especially for larger proteins.

Link to Code

Our implementation and demonstration can be found in our GitHub repository: AlphaFold2-Demo. The repository includes our Jupyter notebook with step-by-step execution of the ColabFold pipeline, visualization code, and explanations of each component. We've also incorporated auxiliary scripts for structural comparison between predicted models and experimentally determined structures.

Click the button above to open our demo notebook directly in Google Colab where you can run the protein structure prediction pipeline yourself.

Experiments

Experiment 1: Lysozyme Prediction (Optimal Conditions)

We conducted structure prediction experiments on lysozyme (129 amino acids), a well-characterized enzyme found in egg whites and human tears. Our implementation:

1. Generated deep multiple sequence alignments (>2500 sequences)





2. Identified suitable templates (1ior, 1ioq, 1iot, 1kxw)



3. Ran predictions with 5 recycles across multiple AlphaFold models



4. Achieved exceptional prediction quality with pLDDT scores averaging 0.984 and ptm scores of 0.913



5. Visualized predictions using both rainbow coloring (by residue position) and confidence coloring
6. Compared our predicted structure with the experimental structure (1LYZ)

The high confidence score (blue coloring in pLDDT visualization) indicates a highly reliable prediction across the entire protein structure, demonstrating AlphaFold2's remarkable ability to predict protein structures with near-experimental accuracy.

Experiment 2: Exploring the Limitations of AlphaFold2

To understand the boundaries of AlphaFold2's capabilities, we conducted two challenging prediction scenarios:

Limited Evolutionary Information for long protein

We tested how the prediction quality changes for a significantly larger protein fragment( ~1000 amino acids) under challenging conditions when evolutionary information is restricted:

• Reduced MSA depth from 512 to just 2 sequences
• Maintained all other parameters
• No structural templates available
• Observed significant degradation in prediction quality:
  • pLDDT scores dropped from 0.984 to 0.433 at best
  • Required full 6 recycles (no early stopping)
  • High predicted alignment errors (predominantly red PAE matrix)
  • Unstable structures across recycles with large RMSD variations
  • Different models produced substantially different predicted structures

Key Insights from Experiments

These experiments collectively demonstrate that AlphaFold2's prediction quality depends on multiple interdependent factors:

• Evolutionary information (MSA depth) plays a crucial role in achieving high-confidence predictions
• Template availability provides important structural guidance
• Prediction quality degrades substantially under challenging conditions with limited information
• The combination of protein complexity and limited evolutionary information presents the greatest challenge
• Under optimal conditions, predictions can achieve near-experimental accuracy

Conclusion:

Our hands-on exploration of AlphaFold2 through ColabFold has revealed both the impressive capabilities and important limitations of this revolutionary technology. While we achieved remarkable accuracy with lysozyme prediction (pLDDT scores of 0.984), our experiments with challenging scenarios demonstrated clear performance boundaries. When working with a protein fragment under constrained conditions (minimal MSA depth and no templates), prediction confidence dropped significantly, with pLDDT scores averaging only around 30-50% and high predicted alignment errors as visualized in the red confidence maps.

These findings highlight a crucial insight about the interdependent factors affecting prediction quality: the absence of sufficient evolutionary information (limited MSA) combined with lack of structural templates dramatically impacts performance, especially for challenging proteins. Our experimental comparison demonstrates how AlphaFold2's success depends on the integration of multiple data sources rather than any single algorithmic innovation. This Nobel Prize-worthy breakthrough represents a sophisticated synthesis of evolutionary information, structural knowledge, and deep learning rather than a complete departure from traditional approaches.

References

[1] Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583-589 (2021). This groundbreaking paper introduced AlphaFold2 and its revolutionary approach to protein structure prediction. It details the neural network architecture that solved a 50-year-old grand challenge in biology with near-experimental accuracy.

[2] Mirdita, M., Schütze, K., Moriwaki, Y. et al. ColabFold: making protein folding accessible to all. Nat Methods 19, 679-682 (2022). This paper describes an optimized implementation of AlphaFold2 designed to run efficiently on consumer hardware. ColabFold democratizes access to state-of-the-art protein structure prediction by reducing computational requirements while maintaining high accuracy.

[3] Google DeepMind. Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry. DeepMind Blog (2024). Retrieved April 21, 2025. This announcement highlights the unprecedented scientific impact of AlphaFold2, which earned a Nobel Prize just four years after publication. The recognition underscores how AI can solve fundamental scientific problems previously thought to require decades more research.

[4] Ahdritz, G., Bouatta, N., Kadyan, S. et al. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nat Methods (2024). This paper presents OpenFold, an open-source reimplementation and retraining of AlphaFold2 that provides deeper insights into how the model works. The study reveals important details about AlphaFold2's learning mechanisms and its ability to generalize to novel protein structures.

Team Members (Authors)

Rishi Mule

Dhanshree Baravkar