Arsha Vidya Mandir

114, Velachery Road, Guindy, Chennai, Tamil Nadu 600032

NAME OF THE STUDENT : KRITHIKA RAJASEKAR

CBSE ROLL NO :

CLASS : XII

SUBJECT : BIOLOGY

PROJECT TITLE : UNRAVELLING PROTEIN

STRUCTURES WITH AI

ACADEMIC YEAR : 2025-2026

1
 Arsha Vidya Mandir
BIOLOGY PROJECT

BONAFIDE CERTIFICATE

I hereby certify that this project titled “Unravelling Protein Structures with AI” is
a bonafide work done by Krithika Rajasekar of Grade XII of Arsha Vidya Mandir
during the academic year, 2025 – 26.

Date : Signature of Teacher in Charge

Roll No: _______________________.

Submitted for the SSCE practical examination held on _________________ at

Arsha Vidya Mandir.

Dr. Aruna A

Principal Internal Examiner External Examiner

School Seal

2
 ACKNOWLEDGEMENT

I thank the CBSE Board for giving me an opportunity to carry out a research project
and widen our knowledge on the chosen topic.

I thank my School management, Correspondent and Principal for their encouragement

and support in the implementation of my project.

I thank my Biology teacher, for providing me the required guidance and support. It is
my humble pleasure to acknowledge my deep sense of gratitude to her, for the constant
help and suggestions at each and every stage, without which it wouldn’t have been
possible to complete this project.

Last but not the least I thank the Almighty for giving me patience and perseverance to
successfully complete this work.

3
 TABLE OF CONTENTS

[Link]. Contents Page No.

1. Introduction 5

2. Aim 6

3. Objectives 6

4. Methodology 6

5. Amino Acids and Peptide Bonds: The Chemistry of Proteins 8

6. How Proteins are Made: From Structure to Function 10

7. ‘The Protein Folding Problem’ and Levinthal’s Paradox 13

8. Early Methodologies to determine Protein Structures 14

9. Hurdles in the Path to Determine Protein Structures 18

10. Rise of AI in Protein Folding with AlphaFold 22

11. Pioneering Protein Design with RF Diffusion 26

12. Applications of AI Driven Protein Science 30

13. Discussion 32

14. Conclusion 33

15. Bibliography 34

4
Introduction:

Proteins are essential biomolecule, involved in everything from muscle contraction to

timely immune responses. Understanding their structure is the key to breakthroughs in
fields like medicine, environmental science and materials engineering. However,
determining the 3D structure of a protein based solely on its amino acid sequence –
known as the ‘Protein Folding Problem’ – has been a major scientific challenge for
over fifty years. Traditional methods like X-ray crystallography and cryo-electron
microscopy provided insights into around 150,000 proteins, but these techniques are
expensive and time-consuming.

Recent advancements in AI have transformed this field. DeepMind’s AlphaFold 2 has

achieved near-perfect accuracy in predicting protein structures, mapping over 200
million proteins. At the same time, David Baker’s lab has developed RF Diffusion, a
model capable of designing entirely new proteins from scratch. This innovation opens
up new possibilities for addressing global challenges such as disease treatment, plastic
degradation and climate change. This report will examine the need for, development
of, and future potential in AI-driven protein science.

5
Aim:

To explore how AI, particularly AlphaFold 2 and RF Diffusion, has transformed our
ability to predict and design protein structures, and to understand the broader
implications of these advances on science.

Objectives:

• To explore the complexities of protein structure and its crucial role in

biological processes.
• To trace the evolution of protein structure determination methods and the
challenges each approach faced.
• To describe the key innovations behind AlphaFold 2, focusing on its use of
evolutionary data and structural modules to predict protein structures.
• To investigate how RF Diffusion is shaping the design of novel proteins, and
its possible applications.
• To evaluate the implications of AI-driven protein science on healthcare,
environmental sustainability and scientific innovation.

Methodology:

This report employs a qualitative analysis of secondary data sourced from scientific
literature, research papers, and documentaries.

The methodology involves the following key steps:

1. Literature Review:
A comprehensive examination of academic journals, review papers and
Nobel Prize citations that explore AlphaFold and advancements in synthetic
protein design.
2. Technical Breakdown:
Analyzing the architecture and training process of AlphaFold 2, with a focus on

6
 its use of deep learning techniques, evolutionary data and attention-based
transformer models for protein structure prediction.
3. Case Studies and Application Exploration:
Investigating the real-world implementation of AlphaFold and RF Diffusion,
including their roles in vaccine development, anti-venom production and the
design of enzymes for plastic degradation.

7
Amino Acids and Peptide Bonds: The Chemistry of Proteins

Proteins are heteropolymers composed of different kinds of monomers, specifically

amino acids. Each amino acid has a common core structure: a central α-carbon (alpha
carbon) bonded to four different groups:

• An amine group (–NH₂)

• A carboxyl group (–COOH)
• A hydrogen atom (–H)
• A variable R group (side chain), which differs between amino acids and
determines their chemical properties

Figure 1: Structure of a general amino acid showing the α-carbon, –NH₂, –COOH, –
H, and R group. Source: [Link]

The term ‘amino acid’ is derived from the presence of both an amine and a carboxyl
group in each molecule.

At physiological pH (around 7.4), amino acids generally do not exist in their neutral
form. Instead, they predominantly exist in the Zwitterionic state. In this form, the amine
group gains a proton to become –NH₃⁺, while the carboxyl group loses a proton to
become –COO⁻. This intra-molecular charge balancing makes the molecule
electrically neutral overall, but with both positive and negative charges present.

8
 Figure 2: Zwitterionic form of an amino acid, showing –NH₃⁺ and –COO⁻ groups
with respective charges. Source: [Link]

The zwitterionic nature of amino acids is crucial for peptide bond formation. During
a condensation reaction, the amine group of one amino acid reacts with the carboxyl
group of another – forming a peptide bond. This covalent bond links the two amino
acids together while releasing one molecule of water (H2O).

Figure 3: Peptide bond formation between two amino acids, with a water
molecule shown as a by-product. Source: [Link]

This reaction can repeat multiple times, allowing amino acids to link together and form
polypeptide chains, which fold into complex three-dimensional structures to form
functional proteins.

How Proteins are Made: From Structure to Function

9
Starting from the polypeptide chain, proteins exhibit four distinct levels of structural
organization, each contributing to their final shape and biological function: primary,
secondary, tertiary, and, in some cases, quaternary structures.

1. Primary Structure:
It refers to the polypeptide chain, with amino acids connected by peptide bonds.
This sequence is genetically determined and dictates all higher levels of protein
structure. The chain has free amine group (N-terminus) at one end and a free
carboxyl group (C-terminus) at the other.

Figure 4: Linear polypeptide chain showing peptide bonds and labeling of N-terminus
and C-terminus. Source: [Link]

2. Secondary Structure:
It arises from regular, repeating patterns formed by hydrogen bonding between the
backbone atoms of the polypeptide chain (not the side chains). The two most
common secondary structures are:
• α-Helix (Alpha Helix): A right-handed coil stabilized by hydrogen bonds
between the carbonyl oxygen of one amino acid and the amide hydrogen of
another four residues ahead in the chain. This results in a tightly coiled, spring-
like structure.

10
 • β-Pleated Sheet (Beta Sheet): Formed when two or more strands of a
polypeptide chain lie side by side, connected by hydrogen bonds. These sheets
may be parallel or antiparallel, and the structure appears folded or pleated.

Figure 5: Alpha helix and beta pleated sheet structures, with hydrogen bonds shown.
Source: [Link]

Only right-handed α-helices are found in naturally occurring proteins due to steric
constraints and stability considerations.

3. Tertiary Structure:
It is the three-dimensional folding of a single polypeptide chain into a compact,
globular shape. This folding is stabilized by various interactions between side
chains (R groups), including:
• Hydrogen bonds
• Ionic interactions
• Hydrophobic interactions
• Disulfide bridges (in some proteins)

This structure can be visualized as a folded ball of yarn, with many internal
interactions holding it together. Not all proteins go beyond this level; for many, the
tertiary structure represents their functional form.

11
 Examples include myoglobin (an oxygen-binding protein) and enzymes like
lysozyme or ribonuclease, which function independently in their tertiary form.

Figure 6: Illustration of a globular tertiary protein structure showing different

stabilizing interactions. Source: [Link]

4. Quaternary Structure (if present):

It arises when two or more tertiary-structured polypeptide subunits come
together to form a functional protein complex. These subunits are held together by
the same types of interactions seen in tertiary structure. Examples include
haemoglobin, composed of four subunits (two α and two β chains), each
maximizing its oxygen-carrying capacity.

Figure 7: Quaternary structure of hemoglobin, showing the arrangement of α and β

subunits. Source: [Link]

12
‘The Protein Folding Problem’ and Levinthal’s Paradox

The folding of a protein is determined by its amino acid sequence (the primary
structure), yet the journey from a linear chain to a functional protein remains one of
the greatest puzzles in molecular biology. This challenge, known as the ‘Protein
Folding Problem’, has both intrigued and frustrated scientists for over half a century.
So, why does this problem persist? If a protein's structure is determined by its primary
structure, shouldn’t it be straightforward to predict?

Apparently not. This paradox was first described by Cyrus Levinthal in 1969.
Levinthal argued that if a protein were to randomly sample all possible conformations,
even a small protein of 100 residues would take longer than the age of the universe to
fold correctly. Yet, in living cells, proteins consistently fold into their native structures
within milliseconds to seconds. This discrepancy, known as Levinthal’s Paradox,
revealed that folding is not a random process, but follows specific pathways and
principles — though the details remained elusive.

Figure 8: Comparing in vitro and in vivo protein folding situations using Levinthal's
Paradox. Source: [Link]

13
Also, understanding protein folding is not just an academic pursuit. Protein misfolding
is central to a range of human diseases, including Alzheimer’s, Parkinson’s,
Huntington’s, amyotrophic lateral sclerosis (ALS), and prion diseases like
Creutzfeldt–Jakob disease. In these conditions, proteins adopt abnormal
conformations that aggregate into toxic fibrils or plaques, damaging cells and tissues.
In prion diseases, misfolded proteins can even propagate their abnormal structure to
other molecules, making them particularly destructive.

Beyond neurodegenerative disorders, misfolding is also linked to cystic fibrosis

(caused by misfolded CFTR protein), sickle cell anaemia (due to abnormal
haemoglobin polymerization), and many cancers, where destabilizing mutations
disrupt tumour suppressor proteins. Thus, the ‘Protein Folding Problem’ is not only
a theoretical puzzle, but a biomedical challenge.

Figure 9: Abnormal haemoglobin polymerization in red blood corpuscles (RBCs),

leading to sickle cell anaemia. Source: [Link]

Early Methodologies to Determine Protein Structures

The quest to determine protein structures began long before the advent of AI. For much
of the twentieth century, structural biology relied on painstaking experimental
techniques that required immense technical skill, sophisticated instruments, and

14
years of effort. While revolutionary for their time, these methods were also limited,
slowing the pace of discovery.

Here, we discuss three key methods: X-ray crystallography, nuclear magnetic

resonance (NMR) spectroscopy, and cryo-electron microscopy (cryo-EM).

X-ray crystallography has been the dominant technique for protein structure
determination since the 1950s. It works on a straightforward but technically demanding
principle: proteins are crystallized into an ordered lattice, then bombarded with X-
rays. The way X-rays scatter or diffract off the crystal provides data that can be
mathematically transformed into an electron density map. By fitting the amino acid
sequence into this density, researchers can reconstruct the protein’s three-dimensional
atomic structure.

Figure 10: Steps of the X-ray crystallography process.

Source: [Link]

The first protein structures solved by X-ray crystallography were myoglobin and
haemoglobin, published in the late 1950s and early 1960s by John Kendrew and Max
Perutz, respectively. These achievements earned them the 1962 Nobel Prize in
Chemistry and demonstrated the enormous potential of crystallography to reveal the
molecular basis of life.

Though X-ray crystallography posed significant challenges, it produced the lion’s

share of protein structures in the Protein Data Bank (PDB) for decades and remains a
cornerstone of structural biology.

15
In the 1980s, NMR spectroscopy emerged as an alternative to crystallography,
especially for proteins that could not be crystallized. NMR exploits the magnetic
properties of atomic nuclei, particularly hydrogen, carbon (13C), and nitrogen (15N).
When placed in a strong magnetic field and exposed to radiofrequency pulses, these
nuclei resonate in ways that depend on their chemical environment. By analysing these
signals, researchers can infer interatomic distances and angles, helping to construct
three-dimensional structures.

NMR offers several unique advantages:

• Dynamic insights: Unlike crystallography, NMR is performed in solution,

allowing scientists to study proteins in conditions closer to their native state.
• Conformational flexibility: NMR can capture ensembles of structures,
revealing how proteins fluctuate between conformations.
• Small protein suitability: NMR is particularly effective for proteins under
~30 kDa, including many regulatory and signaling proteins.

Figure 11: Working of NMR Spectroscopy.

Source: [Link]

Soon, cryo-electron microscopy, or cryo-EM, revolutionized structural biology in

the 2010s, though its origins date back much earlier. In cryo-EM, proteins or protein

16
complexes are flash-frozen in vitreous ice and imaged with an electron beam.
Thousands to millions of images of randomly oriented molecules are computationally
combined to reconstruct a three-dimensional structure.

For many years, cryo-EM produced only low-resolution “blob-like” reconstructions,

earning it the moniker “blob-ology.” However, technological advances transformed the
field:

• Direct electron detectors dramatically improved image quality and sensitivity.

• Advanced algorithms for image alignment and reconstruction boosted
resolution.
• Automation increased throughput and reproducibility.

Figure 12: Steps of the Cryo-EM process.

Source: [Link]

By 2013–2015, cryo-EM entered what many called the “resolution revolution.”

Suddenly, it was possible to achieve near-atomic resolution for large protein

17
complexes that were previously intractable by crystallography or NMR. Structures such
as ion channels, ribosomes, and membrane-bound receptors became accessible,
ushering in a golden age of structural biology.

In 2017, the Nobel Prize in Chemistry was awarded to Jacques Dubochet, Joachim
Frank, and Richard Henderson for their contributions to cryo-EM. Cryo-EM was
widely used to study large, flexible, and membrane-associated proteins,
complementing crystallography and NMR.

Other methods like Small-angle X-ray scattering (SAXS), Mass spectrometry (MS)
and FRET (Förster resonance energy transfer) were also developed. However, they
were not widely used and did not create as much impact as the three methods mentioned
above.

Timeline of Progress and Breakthroughs:

The progress of experimental methodologies can be traced through milestones:

• 1958–1960s: First protein structures solved by crystallography (myoglobin,

hemoglobin).
• 1970s–1980s: Growth of crystallography and emergence of NMR spectroscopy.
• 1990s: PDB expands rapidly; cryo-EM still limited to low resolution.
• 2000s: Genome sequencing outpaces structural determination; structural
genomics initiatives launched.
• 2010s: Cryo-EM revolution enables atomic resolution of large complexes.

Hurdles in the Path to Determine Protein Structures

Determining protein structures has long been one of the most complex challenges in
biology. Though experimental techniques such as X-ray crystallography, NMR
spectroscopy and cryo-EM made significant strides in the 20th and 21st centuries,
numerous technical, computational and biological limitations continue to impede
scalability, accessibility and accuracy.

18
Experimental Challenges

1. Crystallization Bottleneck: Many proteins, particularly membrane proteins,

intrinsically disordered proteins (IDPs), and large complexes, resist forming the
well-ordered crystals required for X-ray diffraction. Even when crystals form,
they often diffract poorly, restricting resolution and delaying the process.
2. Static vs. Dynamic Representation: Most methods capture proteins in a single,
static conformation, which may not reflect their native, dynamic state. For
example, X-ray crystallography often locks proteins in a rigid crystalline lattice,
while cryo-EM averages multiple conformations, potentially missing important
structural details that arise in vivo.
3. Time, Labor, and Resource Intensive: The process of solving a protein
structure requires months to years of effort, involving complex steps like protein
expression, purification, crystallization, and computational refinement. The
techniques demand expensive equipment such as synchrotrons and cryo-EM
facilities, making large-scale structural studies challenging for many labs,
particularly in lower-resource settings.
4. Size and Complexity Limitations: NMR spectroscopy is effective for smaller
proteins (under 40 kDa) but struggles with larger proteins due to complex,
overlapping spectra. Cryo-EM, while transformative, initially had poor
resolution and required large quantities of purified protein, and still demands
expensive equipment and highly trained personnel.

Computational Challenges

1. Folding Complexity: Proteins are made of 20 amino acids, each with unique
chemical properties. A 100-residue chain could fold into an astronomical number
of conformations. This vast “folding space,” as pointed out by Levinthal's
paradox, makes it infeasible to simulate all possible structures through brute
force, even with powerful computational models.
2. Molecular Dynamics (MD) Limitations: MD simulations rely on computing
atomic motions based on physical laws but are constrained by computational

19
 power. Simulating protein folding, which occurs on timescales of milliseconds
to seconds, requires enormous computational resources. Even supercomputers
can only model small proteins over limited timescales.
3. Incomplete Force Fields: Current computational models depend on force fields
to describe atomic interactions. While they have improved, these force fields are
still imperfect, failing to capture long-range electrostatic interactions, solvent
effects, or entropy contributions, leading to inaccuracies in simulating protein
behavior over time.
4. Limited Training Data: Machine learning models require large, high-quality
datasets, but until recently, structural databases were incomplete, especially for
membrane proteins, disordered regions, and multi-subunit complexes. This data
scarcity hampers the accuracy of predictions, especially for underrepresented
protein types.

Biological Challenges

1. Intrinsically Disordered Proteins (IDPs): An estimated 30–40% of the

eukaryotic proteome consists of IDPs or intrinsically disordered regions (IDRs).
These proteins lack stable tertiary structures and instead adopt dynamic
ensembles of conformations. Traditional methods, designed for well-folded
proteins, struggle to capture their conformational flexibility, making them
difficult to study.
2. Post-Translational Modifications (PTMs): Proteins undergo a wide range of
chemical modifications after translation (phosphorylation, glycosylation,
ubiquitination, acetylation, etc.) that dramatically alter their structure, stability,
and function. However, experimental techniques often struggle to capture these
modifications, and many computational models fail to incorporate them, limiting
the biological realism of predictions.
3. Protein Complexes and Interactions: Many proteins do not function in
isolation but form transient, dynamic complexes with other proteins, nucleic
acids, or small molecules. Studying these complexes is much more challenging
than isolated proteins due to their complexity, variability, and transient nature.

20
 This is a major hurdle in understanding cellular processes, as structural data is
often needed for complexes, not just individual proteins.
4. Environmental Effects: Proteins fold and function in the crowded,
heterogeneous environment of the cell, which is vastly different from the
purified, dilute conditions typically used in experimental setups. Factors such as
molecular crowding, pH, ionic strength, and interactions with chaperone proteins
all influence folding and stability, but they are rarely accurately modeled in
traditional experimental or computational approaches.

Societal and Scientific Implications

The challenges of protein structure determination are not only scientific but also
societal. The cost and complexity of these methods have historically limited access to
structural biology, creating disparities between well-funded institutions and those in
low-resource settings. The slow pace of structural determination can also delay
responses to urgent global health issues. For instance, during the early stages of the
COVID-19 pandemic, structural biologists raced to solve the SARS-CoV-2 spike
protein's structure using cryo-EM. AI-based tools like AlphaFold, however,
accelerated research by quickly providing high-quality predictive models of proteins.

Figure 13: Structure of SARS-CoV-2 spike protein. Source: [Link]

21
Additionally, the inability to resolve protein structures at scale has hindered progress in
drug discovery, enzyme engineering and even basic biological research. With
millions of proteins in the human proteome alone, traditional methods would never
meet the growing demand for structural insights in an era of increasing scientific and
biomedical challenges.

Rise of AI in Protein Folding with AlphaFold

AlphaFold, an artificial intelligence system developed by DeepMind, represents one

of the most transformative breakthroughs in modern biology. By first outlining the
limitations of earlier structure-determination methods, we can better appreciate the
scale of AlphaFold’s impact.

This section explores how AlphaFold emerged, the innovations behind it, its validation
through the CASP competitions and the far-reaching implications of its success for
biology and beyond.

The Long Road to AlphaFold

Throughout the 1980s and 1990s, researchers tried to predict protein structures from
sequence using physical chemistry rules or homology modelling, where a protein’s
structure was inferred from a similar, already-known one. While somewhat effective,
these approaches lacked broad accuracy. The Critical Assessment of Protein
Structure Prediction (CASP) competition, launched in 1994, regularly highlighted
just how far computational methods lagged behind experimental ones.

For over two decades, progress at CASP was slow and incremental. Traditional
models were hampered by limited computing power, imperfect force fields and a
lack of large, high-quality datasets. By the early 2010s, enthusiasm had faded—
some questioned whether accurate prediction was even possible. It was in this climate
of stagnation that DeepMind’s AlphaFold made its debut.

22
The CASP Competitions

CASP provided the testing ground that ultimately validated AlphaFold. In CASP11
(2014) and CASP12 (2016), machine learning began to show promise, particularly in
predicting residue–residue contact maps based on evolutionary covariance. These
early breakthroughs hinted that statistical patterns in sequence databases contained
hidden structural information.

In CASP13 (2018), DeepMind’s first version of AlphaFold stunned the scientific

community by outperforming all other competitors. Using deep learning approaches
that combined evolutionary data with novel energy potentials, AlphaFold achieved
accuracy far beyond what had previously been possible. While still imperfect, its
performance marked a clear turning point, signalling that AI might finally provide the
long-sought solution to the ‘Protein Folding Problem’.

However, the true revolution came in CASP14 (2020), when AlphaFold2 delivered
unprecedented accuracy. For many targets, its predictions matched experimental
structures within the margin of experimental error. The results were so striking that
organizers declared the problem “essentially solved”.

Figure 14: CASP Scores for the years 2012, 2013 and 2014; showing AlphaFold's
rise. Source: [Link]

23
Architecture of AlphaFold2: A Conceptual Leap

AlphaFold2 was not simply an incremental improvement over its predecessor; it was
a conceptual and architectural leap. Its design integrated insights from machine
learning, structural biology and physics into a cohesive system. The following
aspects aid its accuracy:

1. Multiple Sequence Alignments (MSAs) and Evolutionary Covariation

AlphaFold2 starts with evolutionary signals. Because proteins evolve under
structural constraints, amino acids that are close in 3D space often co-mutate
to maintain function. By analyzing multiple sequence alignments (MSAs) from
large genomic databases, AlphaFold2 detects these patterns of covariation,
which reveal likely inter-residue distances and orientations — effectively
reconstructing a protein’s structural ‘skeleton’ from its evolutionary history.

2. The Evoformer Module

At the heart of AlphaFold2 is the Evoformer, a deep learning architecture
inspired by transformer models originally developed for natural language
processing. The Evoformer simultaneously processes MSAs and pairwise
residue representations, passing information back and forth through attention
mechanisms. This allows the network to capture both global sequence
relationships and local structural features.

3. End-to-End Differentiable Structure Module

Unlike earlier approaches that predicted intermediate features such as contact
maps, AlphaFold2 is end-to-end. Its structure module directly generates three-
dimensional atomic coordinates, refining them iteratively through gradient
descent. This differentiable approach integrates physical plausibility directly
into the learning process, producing models that obey stereochemical
constraints.

24
 4. Recycling Mechanism
A unique innovation in AlphaFold2 is its recycling mechanism. Predictions are
fed back into the network, allowing iterative refinement. This mimics the way
scientists refine models by repeatedly comparing them to experimental data, but
it is executed automatically within the AI.

5. Confidence Estimation (pLDDT)

AlphaFold2 not only predicts structures but also estimates the reliability of its
predictions. Its per-residue confidence score (predicted Local Distance
Difference Test, or pLDDT) provides users with an internal quality check,
allowing them to distinguish highly reliable regions from uncertain ones.

Together, these innovations enabled AlphaFold2 to surpass all previous attempts,

achieving accuracies previously thought unattainable.

25
26
Figure 15: Working of AlphaFold 2. Source: [Link]
Impact of AlphaFold2’s Breakthrough
1. Validation against previously determined Experimental Structures from
Physics (using root-mean-square derivations or RMSD).
2. Democratization of Protein Structures through public release of the AlphaFold
Protein Structure Database.
3. Acceleration of Biological Discovery due to availability of accurate protein
structures.
4. Complementarity with Experimental Methods (without rendering them
obsolete).

Limitations and Critiques

Despite its success, AlphaFold2 is not without limitations. It struggles with
intrinsically disordered proteins, multi-protein complexes and proteins requiring
cofactors or post-translational modifications. Its predictions represent static
structures, not the full dynamic ensembles that often define protein function.
Furthermore, while AlphaFold2 is accessible, its underlying training data and
computational resources remain centralized, raising concerns about equity and
reproducibility.

Pioneering Protein Design with RF Diffusion

While AlphaFold solved the decades-old challenge of predicting the structure of

proteins from amino acid sequences, it did not address a related and equally profound
challenge: how to design entirely new proteins with tailored properties that do not
exist in nature. Prediction is retrospective — it tells us what a sequence will fold into.
Design, by contrast, is prospective — it asks us to create sequences that fold into
structures optimized for human needs, such as breaking down plastics, neutralizing
toxins, or capturing carbon dioxide.

This is where RF Diffusion, a generative AI framework developed by David Baker’s

Institute for Protein Design (IPD) at the University of Washington, enters the

27
picture. Building upon the successes of AlphaFold and advances in deep generative
modeling, RF Diffusion represents a new phase in structural biology: not just
understanding proteins as they are, but engineering them as we wish them to be.

From Prediction to Design: A Paradigm Shift

The difference between prediction and design can be analogized to architecture.

AlphaFold is like a system that, when given a blueprint, can tell you what the finished
building will look like. RF Diffusion, on the other hand, is akin to an architect who
can draft entirely new blueprints to create buildings with specific features. This
transition from analysis to creativity marks a paradigm shift in protein science.
For decades, protein engineering relied on either directed evolution — iteratively
mutating proteins and selecting improved variants — or rational design, in which
scientists used biochemical knowledge to modify existing proteins. Both approaches
achieved successes, but they were constrained. Directed evolution could not easily leap
to radically new folds, while rational design often failed because of the complexity of
folding rules.
Generative AI models, such as RF Diffusion, circumvent these limitations by learning
the underlying patterns of protein structures directly from data. They can generate
entirely novel folds and functions that nature never explored, vastly expanding the
protein design space.

28
29
Figure 16: Comparison of AlphaFold 2 (Protein Folding) with RF Diffusion (Protein
Design). Source: [Link]
The Principles of Diffusion Models

RF Diffusion is built on the concept of diffusion probabilistic models, a class of

generative models that have revolutionized machine learning in recent years.
Originally developed for image synthesis (e.g., DALL·E and Stable Diffusion),
diffusion models operate by gradually corrupting data with noise and then training a
neural network to reverse the process. By learning how to denoise step by step, the
model can generate entirely new samples from random noise that follow the statistical
patterns of the training data.

Figure 17: Working of RF Diffusion technology. Source: [Link]

In the case of RF Diffusion, the data are not images but protein backbones — the
three-dimensional coordinates of atoms forming polypeptide chains. The model learns
to add and then remove noise from these structures, ultimately allowing it to sample
novel protein conformations consistent with the principles of folding.

RF Diffusion Architecture

RF Diffusion integrates several key innovations that make it uniquely suited for protein
design:
1. Rotationally and Translationally Equivariant Representations
Proteins are three-dimensional objects, and their orientation in space should not
affect how they are represented or understood. RF Diffusion employs SE(3)-
equivariant neural networks, which respect the symmetries of 3D space. This
ensures that the model treats a protein identically whether it is rotated,
translated or flipped, a critical feature for accurate structural learning.

30
 2. Conditioning for Functional Design
RF Diffusion can be conditioned to design proteins with specific properties. For
example, researchers can instruct the model to create a binding pocket with a
desired shape, or to generate symmetric assemblies useful for nanomaterials.
This conditional generation opens the door to targeted protein design for
biomedical and industrial applications.
3. Integration with Sequence Design Models
Once a backbone structure is generated, additional tools such as
ProteinMPNN (a sequence design network also developed by Baker’s group)
can assign amino acid sequences likely to fold into that structure. These
sequences can then be validated experimentally, creating a full pipeline from
design to laboratory realization.

Case Studies of RF Diffusion in Action

• Designing Symmetric Nanocages which self-assemble into predetermined
architectures. These nanocages can be engineered to carry drugs, antigens, or
imaging agents, offering a modular platform for nanomedicine.
• Enzyme Engineering for Plastic Degradation (using PETase enzymes) with
enhanced stability and catalytic efficiency. These enzymes could enable
industrial-scale recycling of plastics, reducing reliance on fossil fuels and
mitigating environmental pollution.

Figure 18: Structure of PETase enzyme. Source: [Link]

31
 • Therapeutic Binders that attach to viral proteins such as SARS-CoV-2 spike
protein and neutralize infection. Unlike antibodies, which are large and
complex, these small designed binders are easier to manufacture and potentially
more stable, offering new modalities in antiviral therapy.

Limitations and Challenges of RF Diffusion

Although RF Diffusion represents a remarkable leap forward, it is not without

limitations:
• Experimental validation required to test efficiency of designed proteins.
• Biases in Training Data (like Protein Data Bank and related datasets), which
are skewed toward certain protein classes.
• Dynamic and contextual functions in proteins are hard to account for.
• Ethical and biosafety concerns about design of harmful biomolecules.

Applications of AI-Driven Protein Science

The advent of artificial intelligence (AI) in protein science has resolved theoretical
challenges and unlocked a vast array of practical applications with great potential.
AlphaFold2 and RF Diffusion have the potential to reshape biotechnology, medicine,
environmental science and materials engineering.
This section explores the applications of AI-driven protein science across multiple
domains, highlighting concrete examples:

1. Medicine and Therapeutics

a. AlphaFold has accelerated drug discovery and development by providing
protein structures to which drugs can bind and act. For example, researchers
working on antimicrobial resistance have used AlphaFold to predict the
structures of bacterial enzymes that degrade antibiotics, allowing the
rational design of inhibitors.

32
 b. RF Diffusion has aided protein-based therapeutics, such as the production
of novel antibodies, enzymes and cytokines to mitigate disease.
c. Personalized medicine using patient-specific genomic data.

2. Infectious Diseases and Global Health

a. AlphaFold provided structural models of the SARS-CoV-2 spike protein,
aiding rapid vaccine development during the COVID-19 pandemic.
b. AI tools can help combat antimicrobial resistance by predicting structures
of resistance enzymes, such as β-lactamases and enabling the design of
inhibitors.

2. Environmental Sustainability
a. Plastic degradation using PETase enzymes synthasized by RF Diffusion.
b. Carbon capture and greenhouse gas mitigation using enzymes with
enhanced carbonic anhydrase-like activity.

3. Agriculture and Food Security

a. AlphaFold-predicted structures of plant proteins provide insights into how
crops respond to drought, pests and soil conditions. This knowledge can inform
the engineering of more resilient crop varieties.
b. Sustainable food production using enzymes designed with RF Diffusion.

4. Materials Science and Nanotechnology

a. Protein-based nanostructures with diverse potential in drug delivery,
biosensing, etc.
b. Synthetic biology platforms to enhance efficiency of biomanufacturing
processes, from biofuels to pharmaceuticals.

5. Advanced Theoretical Research to illuminate less understood regions of the

proteome, including proteins involved in neurodegeneration, immunity and cell
signalling.

33
Discussion:

When I began this project, I believed that AI might one day replace scientists
altogether. With so much discussion about automation and machines performing tasks
faster and more accurately than humans, it seemed inevitable that AI would eventually
take over scientific research. Learning that AlphaFold2 could predict millions of
protein structures in such a short time only strengthened that concern. It appeared that
if a computer could solve one of biology’s greatest mysteries, there would be little left
for people to do.

However, as I explored the topic further, I realized that AI is not replacing scientists
but supporting them. AlphaFold2 and RF Diffusion have not removed the need for
experiments; instead, they make research more focused, efficient and creative. AI can
analyse immense data sets and uncover hidden patterns; but human insight,
validation and ethical judgment remain essential.

Through this project, I learned that science today is defined by collaboration between
humans and machines. Every breakthrough — whether in protein prediction, enzyme
design or disease research — still depends on curiosity, teamwork and
responsibility.

In conclusion, I no longer see AI as a threat to scientific progress but as a powerful

collaborator that extends what humanity can achieve. Technology, I realized, is not
the end of human effort — it is the next step in it.

34
Conclusion:

The integration of AI into protein science marks a historic turning point in biology.
AlphaFold2’s success in predicting protein structures with near-experimental
accuracy solved a challenge that had frustrated scientists for decades, providing
structural insights across the entire proteome and democratizing access through the
AlphaFold Protein Structure Database. Meanwhile, RF Diffusion has extended the
frontier from prediction to creation, enabling the design of entirely new proteins with
functions tailored to medicine, sustainability and nanotechnology. Together, these
breakthroughs represent a paradigm shift from descriptive to predictive and
generative biology.

The implications extend far beyond the laboratory. Applications in healthcare,

environmental remediation, agriculture and materials science promise to address
some of humanity’s most pressing challenges. Yet these opportunities are accompanied
by ethical and societal questions about access, equity, safety, and governance. These
are not within the scope of the report.

In sum, the rise of AI in protein science is not merely a technical achievement but a
profound reimagining of what it means to study and engineer life. Its future trajectory
will be defined not only by scientific ingenuity but also by the collective wisdom with
which humanity chooses to wield this transformative power.

35
Bibliography:
• [Link]
• [Link]
model/#life-molecules
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• [Link]
• NCERT Biology – Textbook for Class XI (2025-26 Edition)

36
 TEACHER’S OBSERVATION

I acknowledge that this project is the original work, done by Krithika Rajasekar of Class
XII as per the guidelines of CBSE requirement. I also record that this project is a result
of adequate research and laboratory activities/ field visits made, which laid the
foundation for the completion of this project.

Teacher’s sign

37