In a clinic somewhere in the world, a patient finishes months of tuberculosis treatment. Every pill was taken on time. Every instruction followed. By the book, the infection should be gone.

But it isn’t.

The bacteria are still there, alive, unchanged, and no longer affected by the drugs that once worked. The patient didn’t do anything wrong. The medicine didn’t fail because it was weak. The bacteria simply adapted faster than the treatment could keep up.

In 2023, tuberculosis killed more people than any other infectious disease. This is hard to accept because TB is, in theory, curable. We have antibiotics that work. We can sequence the bacterium’s DNA. We understand how the disease spreads. And yet, people continue to die from an infection we know how to treat.

The problem isn’t a lack of medical knowledge, it’s how treatment responds to a living, evolving system. TB care still relies on fixed schedules: take these drugs, at these doses, for these many months. But bacteria don’t follow schedules. As treatment progresses, some TB cells mutate, survive, and pass those mutations on. Over time, the population shifts. Drug-resistant strains take over, and antibiotics that once worked become useless.

Drug resistance doesn’t just make TB harder to treat; it reshapes the biology of the infection itself. Resistant bacteria persist, spread, and turn what should be a manageable disease into a deadly one again.

This pattern isn’t unique to tuberculosis. Many diseases, from infections to cancer, involve systems that adapt under pressure. Our tools are powerful, but they are mostly reactive. We wait until resistance emerges, until treatments fail, until the system breaks, and only then do we respond.

But what if we didn’t have to wait?

What if, instead of chasing resistance after it appears, we could intervene at its source, rewriting the genetic changes that cause resistance in the first place, and restoring bacteria to a state where the drugs we already have can work again?

The Status Quo: Powerful Tools, But Working Separately

Today, we have tools that seem like they could solve this problem:

• DNA sequencing: We can read a bacterial genome in hours, identifying exactly which mutations cause resistance.
• Gene editing: Tools like prime editing allow scientists to change single letters in DNA, precisely and safely.
• Artificial intelligence: AI models can predict which gene edits will work best, which mutations matter most, and how genes interact.

Each tool is powerful, but they mostly operate in isolation. Gene editing experiments are designed and tested slowly. AI can predict targets, but those predictions rarely feed directly into experiments. Sequencing data exists, but it isn’t automatically turned into edits that fix the problem.

Meanwhile, diseases like TB are complex networks. Drug resistance isn’t caused by just one mutation. Mutations interact, feedback loops amplify effects, and the bacteria’s survival strategies make them resilient. Editing one gene without understanding the network around it can be ineffective or even harmful.

So the current system is limited. We can read, predict, or edit, but we haven’t yet combined all three in a smart, systematic way to reverse resistance.

This raises a key question:

“Which critical genes or mutations in the disease network, if we corrected them, could shift the entire system back to a healthy state, and how can AI guide us to make those corrections efficiently and precisely?”

Answering this question could allow us to move from reacting to resistance toward controlling it at the genetic level.

The Core Insight: Resistance Is Written in DNA

Drug resistance in TB often comes down to tiny changes in the bacteria’s DNA. Many of these changes are just single letters swapped in the genetic code, called point mutations.

These mutations are well-mapped and predictable. That means, in theory, we know exactly which letters need to be changed to make the bacteria sensitive to the drugs again.

The big insight here is simple but powerful: resistance isn’t always permanent; it’s a genetic state that we can target. Instead of trying to kill the bacteria with stronger or longer treatments, what if we could rewrite the mutations causing resistance, returning the bacteria to a drug-sensitive state?

This changes the whole approach to TB treatment. Resistance becomes something we can edit, not just something we fight.

The Idea: AI-Guided Prime Editing to Rewrite Resistance

Instead of trying to kill resistant TB bacteria with stronger drugs, we can focus on the genetic changes that actually cause resistance. By correcting these mutations, the bacteria could become sensitive again to antibiotics that are already proven to work.

This approach combines AI and prime editing: AI helps design precise DNA edits, and prime editing applies those edits to reverse the mutations in Mycobacterium tuberculosis. By targeting the root cause of resistance rather than its symptoms, we may be able to restore the effectiveness of existing treatments.

Step 1: Identify the mutation

The first step would be sequencing the bacteria from a patient. Genome sequencing today is fast and highly accurate, and many resistance-causing mutations are already well documented. For example:

• A single change in the rpoB gene can prevent rifampicin from binding.
• A point mutation in katG makes isoniazid ineffective.

Once we know the exact mutation responsible for resistance, we have a specific target for editing. This allows us to act precisely, rather than relying on broad-spectrum antibiotics that affect many bacterial functions at once.

Step 2: Design the edit with prime editing

Prime editing is a newer type of genome editing that allows us to change individual DNA letters without cutting the DNA completely (doubt strand break).

This precision is crucial. Traditional editing methods can trigger repair mechanisms that cause additional mutations or even kill the bacteria prematurely. Prime editing avoids these problems, making it a safe and reliable tool for reversing specific resistance mutations.

Step 3: Optimize the design with AI

Designing a prime edit isn’t as simple as choosing the right DNA letter to change. Each edit depends on a guide sequence called a pegRNA, which directs the cell’s editing machinery to the correct location and specifies the exact genetic correction. Even small differences in pegRNA design, sometimes just a few nucleotides, determine whether an edit works efficiently, works poorly, or fails entirely.

This design sensitivity makes prime editing a natural candidate for machine learning. Instead of testing dozens of pegRNA designs through slow and costly trial-and-error, AI models can be trained on large experimental datasets to predict which designs are most likely to succeed. OPED (Optimized Prime Editing Design) is one such framework developed specifically to improve prime editing outcomes. OPED evaluates key features of pegRNA design, including primer binding site length, reverse transcription template structure, and sequence composition, and ranks candidate designs based on their predicted efficiency. This allows researchers to prioritize high-probability edits before entering the lab, rather than relying on intuition or random selection.

AI-driven optimization also helps reduce unintended effects. In bacterial genomes like Mycobacterium tuberculosis, genes are densely packed, and poorly designed edits can interfere with nearby sequences or introduce off-target changes. By learning patterns from thousands of previous editing attempts, machine learning models can identify pegRNA designs that are more likely to cause errors and deprioritize them, improving precision and safety.

By integrating AI into pegRNA design, prime editing shifts from a trial-and-error process to a rational, design-driven workflow. Researchers can computationally screen and refine edits before experimentation begins, reducing wasted effort and increasing the likelihood that each edit successfully corrects a resistance mutation. This optimization step is critical for making genome rewriting a realistic strategy rather than a theoretical one.

Step 4: Apply the edit and test sensitivity

After choosing the optimized design, the edit is applied to the bacteria in the lab. The bacteria would then be tested with antibiotics to see if they have regained sensitivity.

If successful, this could restore the effectiveness of standard TB drugs, even against strains that were previously multidrug-resistant. This approach could reduce the need for longer, more toxic treatments and improve patient outcomes worldwide.

This method shifts treatment from reacting to resistance to actively controlling it at the genetic level. Instead of fighting the symptoms, we focus on fixing the root cause.

4. Why Current Treatments Can’t Solve Resistance

Even the most powerful antibiotics struggle because TB bacteria evolve faster than treatments can respond. Current TB therapies rely on killing bacteria with multiple drugs over long periods, but this approach has several limitations:

1. Resistance keeps evolving — Each round of antibiotics creates selective pressure. The bacteria that survive carry mutations that make them resistant. Over time, these resistant strains dominate, making treatment progressively harder. For example, multidrug-resistant (MDR) TB is already resistant to the two most important first-line drugs: rifampicin and isoniazid. Extensive drug-resistant (XDR) TB can resist even more drugs, leaving patients with few options.
2. Broad-spectrum drugs have side effects — Current treatments impact not only TB bacteria but also other beneficial microbes in the body. This can cause digestive problems, liver toxicity, and even long-term impacts on the microbiome. Patients often struggle with adherence because the side effects are severe, which can further drive resistance.
3. Treatment is reactive, not proactive — Doctors usually detect drug resistance only after standard treatment fails. By that point, the bacteria may have already evolved additional mutations or spread to others, increasing public health risks.

The core problem is that current methods fight symptoms, not the source. They attempt to kill resistant bacteria rather than fixing the underlying mutations that caused resistance in the first place. This creates a catch-up game that is expensive, slow, and often dangerous.

5. Why Prime Editing Works Here

Prime editing offers a way to directly fix the genetic changes that cause resistance, which is something current treatments cannot do. Here’s why it’s a good fit for TB:

• Precision: Prime editing can change individual DNA letters without cutting the DNA completely. Traditional CRISPR methods make double-strand breaks, which can trigger unpredictable repair mechanisms or kill the bacteria before the intended edit occurs. Prime editing avoids these problems, making it safer and more predictable.
• Flexibility: TB resistance comes from many small, single-letter mutations in key genes like rpoB (rifampicin resistance) and katG (isoniazid resistance). Prime editing can target these exact points instead of making a generalized solution and correcting the mutation without affecting other parts of the genome. This allows us to restore sensitivity to the same drugs that previously failed.
• Demonstrated potential in bacteria: Lab experiments have shown that prime editing can successfully edit bacterial genomes, including E. coli and other model organisms. Although TB is slower-growing and more difficult to work with, the underlying principles still apply. With optimized delivery systems, prime editing is feasible in lab-grown TB strains.
• Reduced risk of unintended changes: Because prime editing doesn’t cut both DNA strands, it’s less likely to introduce off-target mutations, making it a cleaner and safer approach for correcting resistance mutations.

In short, prime editing gives us a tool that can fix the source of drug resistance directly, instead of trying to outsmart bacteria with stronger drugs.

Feasibility

This idea is grounded in technologies that already exist and have been tested in real biological systems. The genetic basis of TB drug resistance is well understood, with mutations in genes such as rpoB and katG reliably linked to resistance against rifampicin and isoniazid. These mutations are extensively documented in genomic databases, meaning the targets for intervention are clearly defined before any editing begins.

Prime editing provides a practical mechanism for correcting these resistance-conferring mutations. Since its introduction, prime editing has been experimentally shown to support precise single-nucleotide substitutions, small insertions, and deletions with fewer unintended edits than traditional CRISPR-Cas9 approaches. This level of precision is especially important for correcting point mutations that cause drug resistance, where even small bystander changes could disrupt bacterial viability or create new resistance pathways.

The real challenge is not whether prime editing can introduce desired changes, but whether it can do so efficiently and reliably, and this is where OPED’s performance has strong experimental support. In the peer-reviewed study that introduced OPED, pegRNAs with higher OPED scores resulted in substantially higher editing efficiencies compared with lower-scoring designs across varied experimental conditions. Across multiple benchmark datasets, pegRNAs ranked highly by OPED showed between ~2.2-fold and ~82.9-fold increases in editing efficiency over poorly ranked designs, demonstrating that AI–guided optimization can produce dramatic improvements in actual editing outcomes.

Beyond predictive benchmarks, OPED-optimized designs were also tested in real editing experiments involving installation of ClinVar pathogenic variants (well-studied disease mutations). In these experiments, the top OPED designs led to successful edits at most targeted loci and outperformed existing pegRNA design tools by 2.1- to 47.5-fold across several prime editing systems, including PE2, PE3/PE3b, and enhanced prime editors (ePE). This demonstrates that OPED doesn’t just predict which designs might work, its optimized designs translate directly into measurably better lab results.

Taken together, the feasibility of this approach rests on three validated components: well-characterized resistance mutations, a genome editing method capable of precise correction, and an AI-driven design framework that has been shown experimentally to dramatically improve editing outcomes. While practical challenges remain, including TB’s slow growth rate and the need for high-biosafety laboratory conditions to work safely with Mycobacterium tuberculosis — these affect experimental logistics rather than technical plausibility. As a result, this idea represents a realistic and testable direction for proof-of-concept studies, rather than a speculative extension of unproven tools.

Limitations and Challenges

While promising, the approach has limitations:

• Delivery: Introducing prime editing machinery into TB bacteria efficiently is still technically difficult.
• Complex mutations: Some drug resistance may involve multiple interacting mutations or unknown regulatory mechanisms that are harder to correct.
• Biosafety and ethics: Working with TB requires high-level containment, and genome editing raises ethical considerations.
• Translation to patients: Applying this in clinical settings would require robust systems for safely editing bacteria inside the human body, a major hurdle for future research.

These challenges do not make the idea impossible, but they do highlight the gap between lab feasibility and clinical application, and the areas where innovation will be needed.

Broader Impact and Closing Reflection

If successful, this approach could transform TB treatment worldwide. Restoring the effectiveness of existing drugs could shorten treatments, reduce side effects, and lower the chance of multidrug resistance. Beyond TB, the same AI-guided prime editing framework could be applied to other bacterial infections or even complex genetic diseases, where multiple mutations interact in networks.

More broadly, this project represents a shift in medicine: from reacting to disease toward controlling it at the genetic level, with AI serving as the link between data, prediction, and action. By integrating sequencing, AI, and precise gene editing, we move closer to a world where we can anticipate resistance before it happens, design interventions that target the source, and potentially prevent treatment failures before they occur.

Sources

Liu, Feng, et al. “Design of Prime-Editing Guide RNAs with Deep Transfer Learning.” Nature Machine Intelligence, vol. 5, no. 11, 1 Nov. 2023, pp. 1261–1274, www.nature.com/articles/s42256-023-00739-w, https://doi.org/10.1038/s42256-023-00739-w.

Abstract

Prime editing can make precise DNA changes without double-strand breaks, but its success depends heavily on designing the right pegRNA, and existing tools often perform poorly because they rely on small datasets and simple design rules. OPED (Optimized Prime Editing Design) addresses this by learning from large experimental datasets and modelling the complex sequence features that influence editing efficiency. It predicts and ranks pegRNA designs and their paired nicking sgRNAs, outperforming alternative tools across PE2, PE3, and PE3b systems. Experimental validation in human cells shows that OPED-optimized designs consistently yield higher editing efficiencies. OPED also powers OPEDVar, a database of optimized designs for 77,738 ClinVar pathogenic variants, and a web application that generates customized pegRNA/sgRNA sets from user inputs. Overall, OPED turns prime-edit design from guesswork into a predictable, scalable engineering process.

Table of Contents

1. Introduction

2. Background: Prime Editing & Why pegRNA Design Is Hard1. Introduction

• 2.1 What pegRNAs actually do
• 2.2 Why pegRNA design is so hard
• 2.3 Why current design tools weren’t enough

3. Existing Approaches (and Their Limitations)

4 . Introduction to OPED

5. Technical Deep Dive

• 5.1 How OPED Works
• 5.2 Performance Evaluation of OPED
• 5.3 Experimental Validation
• 5.4 Comparison With Alternative Methods
• 5.5 Extending OPED to Optimize Nicking sgRNA

6. OPEDVar Database

7. OPED Web Application

8. Discussion and Applications

• 8.1 Importance for Research
• 8.2 Therapeutic Potential
• 8.3 Advantages Over Alternative Methods
• 8.4 Broader Implications

9. Limitations and Future Directions

10. Conclusion

11. Sources

1. Introduction

Every genetic disease, from sickle cell anemia to Tay-Sachs, starts with something unbelievably small: a single change in the 3.2 billion letters of human DNA. For years, our ability to correct these mutations has fallen far behind our ability to identify them. CRISPR brought gene editing into the mainstream by giving scientists a programmable way to cut DNA, but cutting isn’t always the safest or most efficient strategy.

CRISPR does something called double-strand breaks, where both strands of the DNA helix are severed at once, forcing the cell into emergency repair mode. Instead of carefully fixing the sequence, the cell often rushes to patch the break, leading to random insertions or deletions, unwanted mutations, and sometimes even large-scale chromosomal rearrangements. These repair mistakes can introduce new mutations that the scientist never intended and increase the risk of genomic instability, which is especially problematic for therapeutic applications.

Prime editing was designed to solve this. Often described as a “DNA search-and-replace tool,” prime editing can precisely rewrite letters in the genome without cutting both DNA strands. It’s capable of fixing the majority of known mutations, at least in theory.

But in the real world, prime editing still struggles.

Why?
Because the success of a prime editing experiment depends almost entirely on the design of the prime editing guide RNA (pegRNA).

Because the success of a prime editing experiment depends almost entirely on the design of the prime editing guide RNA (pegRNA). Choosing the best and most effective pegRNA is incredibly hard; there are dozens of possible PBS lengths, RTT lengths, PAM positions, distances, alternative configurations, and biological constraints (PBS is the short “landing pad” the polymerase binds to; RTT is the template containing the edit; PAM is the short DNA sequence Cas9 needs to latch onto). Small changes in pegRNA design can dramatically reduce editing efficiency, and for researchers, designing pegRNAs often feels like guesswork.

That is where OPED comes in. OPED is a machine-learning model trained on thousands of real prime-editing outcomes. It predicts which pegRNA designs will work best, ranks all possible designs, and even optimizes the secondary nicking single-guide RNA (sgRNA) used in PE3 systems. It is built to handle the complexity that human intuition and existing rule-based tools cannot. OPED doesn’t just make prime editing easier; it makes it smarter, more accurate, and more scalable.

In a field where one wrong nucleotide can be the difference between curing a disease and creating a new mutation, OPED offers something rare: precision.
And precision is what pushes gene editing from a lab to a clinical treatment.

2. Background: Prime Editing & Why pegRNA Design Is Hard

Prime editing is one of the most precise genome-editing technologies developed to date. Instead of cutting both DNA strands like CRISPR-Cas9, prime editors use a more controlled, flexible architecture that can install substitutions, insertions, and deletions without relying on double-strand breaks. Prime editing has evolved into multiple versions, but PE2 and PE3 remain foundational in research and therapeutic development.

PE2 consists of a Cas9 nickase fused to a reverse transcriptase (RT). It uses a single engineered pegRNA to both target the genomic locus and encode the desired edit. PE3 builds on this by adding a second guide RNA, known as the nicking sgRNA, which creates an additional nick on the non-edited DNA strand. This second nick shifts the DNA repair balance toward incorporating the edited sequence, often increasing editing efficiency.

2.1 What pegRNAs actually do

A pegRNA is not just a standard CRISPR guide RNA. It contains:

• Spacer sequence (targets Cas9 to the genomic site)
• Scaffold sequence
• PBS (primer binding site)
• RTT (reverse transcriptase template) containing the desired edit

These extra elements make pegRNAs highly flexible, but also difficult to design optimally. Even small changes in the PBS or RTT lengths can dramatically alter editing success. Reverse transcriptase has to properly bind, extend, and integrate the edited DNA patch, which varies widely depending on pegRNA structure.

2.2 Why pegRNA design is so hard

Researchers quickly discovered that pegRNA design is the bottleneck of prime editing. The editing efficiency can crash if:

• The PBS is too short (unstable primer binding) or too long (secondary structures)
• The RTT is too short (incomplete edits) or too long (reduced efficiency)
• The distance between the edit site and the nick isn’t ideal
• The PAM near the site restricts the spacer sequence
• The secondary nicking guide (PE3) doesn’t align properly

These design choices interact in various ways, meaning that there is no simple rulebook to follow, and each edit requires an almost ‘unique’ pegRNA design.

Additionally, even if the edit type is the same, a pegRNA that works perfectly at one site can fail completely at another. Because of this variability, researchers began to rely on:

• intuition
• trial-and-error
• partial rules from previous papers
• limited computational tools with narrow applicability

But no one method could reliably predict which pegRNA would work best for a new pathogenic mutation.

2.3 Why current design tools weren’t enough

Before OPED, pegRNA design mostly came from:

• classical rule-based calculators
• small training datasets
• heuristics (e.g., “start with an 8–12 nt PBS”)
• models restricted to a few cell types or PE versions

These approaches often failed when:

• switching from PE2 → PE3 or PE3b
• using different Cas9 variants with flexible PAMs (NGG, NG, SpG, SpRY)
• trying to correct a large number of clinical variants
• predicting efficiency across diverse genomic contexts

Because editing efficiency can change a lot depending on the pegRNA design and the billions of possible combinations, manual design would be practically impossible at such a scale.

3. Existing Approaches (and Their Limitations)

Several computational and heuristic tools have tried to design effective pegRNAs, but most have fundamental limitations. Early models often assume linear effects, treating features like PBS length, RTT length, melting temperature, or edit-to-nick distance as independent. In reality, pegRNA efficiency depends on complex, non-linear interactions between sequence context, secondary structure, and target locus constraints. As a result, traditional tools often fail to perform well on new sequences.

Furthermore, most existing methods are limited by small or low-diversity training datasets, which restricts their ability to predict pegRNA outcomes across different cell types, edit classes, or genomic environments. As a result, their predictions often break down when applied to large experimental libraries or novel edit types. This becomes especially problematic in prime editing, where small changes in design, such as a two-nucleotide change in RTT, dramatically shift editing efficiency.

Another major limitation is scalability. Many current tools were not designed to evaluate or compare thousands of pegRNA variants for a single edit. Their algorithms rely on slow, sequential scoring, making them impractical for large-scale applications such as testing thousands of pegRNAs across many genes or the whole genome, for research or therapy purposes.

Most critically, no existing tool prior to OPED was able to optimize both the pegRNA and the secondary nicking sgRNA used in PE3 systems. PE3 requires coordination between the pegRNA-encoded edit and the nicking sgRNA. Previous tools treated these components independently or ignored the second guide entirely, leading to suboptimal predictions and lower editing performance.

OPED addresses these limitations by learning non-linear sequence context, scaling to extremely large design spaces, and simultaneously optimizing both pegRNA and nicking sgRNA elements, something no earlier method could do.

4 . Introduction to OPED

Prime-editing performance depends heavily on the design of the pegRNA, but most existing tools struggle to predict which designs will actually work in cells. OPED, Optimized Prime Editing Design, was developed specifically to solve this problem. OPED is an interpretable nucleotide language model trained on thousands of experimentally measured prime-editing outcomes across many datasets. Because of this large and diverse training base, it learns patterns in pegRNA performance that earlier tools missed.

Instead of relying on simple rules like “longer PBS is better” or “keep the edit close to the nick,” OPED analyzes multiple layers of information at once: the full sequence context around the edit, the relationships between the PBS, RTT, and edit site, and additional biological constraints that affect pegRNA stability and function. It also captures non-linear relationships within the sequence that influence efficiency.

Using all these features together, OPED predicts the efficiency score for every possible pegRNA design for a given edit. It then ranks all designs and outputs the most efficient ones. In PE3 systems, it can also optimize the secondary nicking sgRNA, making it the first method that provides coordinated design recommendations for both components.

5. Technical Deep Dive

5.1 How OPED Works

OPED takes the full sequence context of a prime-editing design and converts it into a set of features that determine how efficient a pegRNA will be. The model uses several inputs simultaneously, including the pegRNA sequence, PBS, RTT, the edit position, PAM availability, and the edit-to-nick distance. Because all possible pegRNA designs share the same intended edit, OPED’s job is to figure out which PBS/RTT combinations and positional arrangements are most likely to give high editing efficiency.

Under the hood, OPED uses topological and sequential embeddings to learn how nucleotides interact across the pegRNA components. Unlike traditional models that rely on simple handcrafted features or linear decision boundaries, OPED captures non-linear patterns that influence prime-editing success. The architecture processes the pegRNA and genomic context to learn an internal representation of the design, incorporating both local sequence motifs and broader positional relationships.

The output layer performs regression, generating a predicted “editing score’’ for each candidate pegRNA. These scores reflect the expected efficiency measured in real experiments. OPED then ranks all pegRNAs according to these predicted scores, enabling precise selection of the most promising design for the intended edit.

5.2 Performance Evaluation of OPED

OPED was tested against a variety of other computational methods to see how well it predicts which pegRNA designs will work best. These other methods included different types of machine-learning models, like basic neural networks, tree-based models, and methods that look at patterns in the data in different ways, as well as simpler rule-based approaches that rely on fixed “if-then” rules.

All the models were trained using the exact same datasets, so they had access to the same examples of successful and unsuccessful pegRNA designs. In every comparison, OPED performed better than all of these alternative methods, consistently making more accurate predictions and ranking the best pegRNA designs higher. This shows that OPED’s approach, which considers complex interactions in the sequence, is more effective than traditional or other machine-learning strategies.

OPED was evaluated on how well it could predict the success of different pegRNA designs. To do this, the researchers split their data into different groups for testing. Each test group focused on a slightly different challenge:

• HT-test (High-Throughput test): This set included thousands of pegRNA designs from large-scale experiments. It was used to see if OPED could predict efficiency across many designs at once.
• Type-test: This set focused on specific types of edits, like single-letter changes (SNPs), insertions, or deletions. It tested whether OPED could handle different kinds of edits accurately.
• Position-test: This set looked at edits occurring at different positions in the DNA sequence. It tested whether OPED could predict efficiency even when the target site within the gene changed.

Across all these tests, OPED consistently performed better than other models. It made more accurate predictions, handled new or unseen pegRNA designs well, and gave stable rankings of which designs were likely to work best.

In simpler terms, OPED could reliably tell which pegRNAs would give high editing efficiency, even when the editing situation changed (like the type of mutation or its position in the gene). Compared to other models, it improved predictive power roughly 2–3 times, especially in situations where editing outcomes were usually unpredictable or highly variable.

A key strength is that OPED maintains strong accuracy across different prime-editing systems. It reliably predicts pegRNA efficiency for PE2, PE3/PE3b, and ePE systems, whereas most prior methods were limited to PE2-only because of a lack of large PE3 datasets. OPED’s ability to generalize across systems and sequence contexts makes it significantly more robust than existing tools.

5.3 Experimental Validation

To validate OPED’s predictions in the lab, the researchers tested whether pegRNAs ranked highly by OPED actually produced higher prime-editing efficiencies. The experimental workflow involved plasmid construction, cell culture, transfection, and sequencing, closely following standard protocols.

Plasmids:
Plasmids are small, circular pieces of DNA that scientists can use to give instructions to cells. Think of them like USB drives carrying a program: when you put them into a cell, the cell can read the instructions and perform specific tasks.

pCMV-PE2 (main plasmid):
This plasmid contains the “prime editing machinery,” which is the molecular tool that actually makes precise edits in the DNA. Without this plasmid, the cells wouldn’t be able to carry out prime editing.

pegRNA plasmids:
These plasmids contain the “guide” molecules, which are the pegRNA. To make these plasmids, scientists started with an existing plasmid (pGL3-U6-sgRNA-EGFP) and modified it by adding synthesized scaffold and spacer sequences:

• Spacer sequence: Specifies the DNA sequence to target.
• Scaffold sequence: Helps the pegRNA fold correctly so it can interact with the prime editor.

Other plasmids (pCMV-PE2-P2A-hMLH1dn and pCMV-NG-PE):
In addition to the main prime editor plasmid and the pegRNA guide plasmids, the researchers also used these alternative plasmids to test different versions of the prime editor. These variants may work better in certain cells or improve editing efficiency, allowing the researchers to compare performance across different prime-editing systems.

Cell culture and transfection:
The researchers used HEK293T cells, a type of human cell commonly grown in labs. These cells were kept in a nutrient-rich liquid called DMEM with 10% serum, at body temperature (37°C) and in an environment with 5% carbon dioxide, conditions that help the cells stay healthy and grow.

For the experiments, the cells were placed in small wells of a 24-well plate, which acts like a mini test tube for each group of cells. The scientists then introduced a mixture of plasmids, the main prime editor (pCMV-PE2), the pegRNA plasmid, and a nicking sgRNA plasmid, into the cells using a chemical called a transfection reagent. This reagent helps the cells take in the plasmids, similar to how a delivery system brings instructions into the cell.

After about 72 hours, the researchers used a fluorescence-based method to identify which cells successfully took up the plasmids. Cells that were “lit up” were collected, ensuring that only the cells that received the editing instructions were analyzed further.

DNA extraction and sequencing:
After collecting the cells that successfully took up the plasmids, the researchers extracted DNA from them. This is like opening the cell and taking out its instruction manual.

Next, they focused on the parts of the DNA that were meant to be edited. Using a technique called PCR (polymerase chain reaction), they made many copies of these specific regions, so there was enough DNA to study. They used barcoded primers, which act like tiny labels, to keep track of which DNA came from which experiment.

The amplified DNA was then sequenced, meaning the exact order of letters (A, T, C, G) in the targeted regions was read. Finally, the sequencing data was analyzed with a software called CRISPResso2, which calculates how efficiently the prime editing worked and how often the desired edits occurred.

Results:
The lab experiments confirmed that pegRNAs ranked highly by OPED worked better than standard designs, giving higher editing efficiency. This showed that OPED’s predictions translate into real improvements in prime-editing experiments for both PE2 and PE3 systems.

5.4 Comparison With Alternative Methods

OPED distinguishes itself from existing prime-editing design strategies in several key ways.

Heuristic approaches: Traditional PE design methods rely on simple rules or heuristics, such as fixed PBS and RTT lengths or standard edit-to-nick distances. While these methods provide quick guidance, they often fail to account for the complex, non-linear interactions in prime editing, leading to suboptimal editing efficiencies. OPED overcomes these limitations by learning from thousands of real editing outcomes, enabling predictions that reflect biological context rather than relying on fixed assumptions.

Machine-learning alternatives: Existing ML models for prime-editing design generally depend on limited datasets or handcrafted features, which constrain their scalability and accuracy. OPED, by contrast, is trained on a much larger dataset and incorporates sequence context, positional information, and biological constraints. This allows it to generalize across different variants, including previously unseen mutations, and consistently outperform prior ML approaches.

Technical advantages:

• OPED efficiently handles long RTT sequences and variable edit-to-nick distances, scenarios where traditional methods struggle.
• It can simultaneously optimize pegRNAs and secondary nicking sgRNAs for PE3 systems, a capability unique among existing tools.
• The result is a strong, flexible system that improves prime-editing efficiency on many different types of targets. This makes OPED a major step forward compared with previous rule-based tools and machine-learning models.

5.5 Extending OPED to Optimize Nicking sgRNA

While OPED was initially developed to optimize pegRNAs, it has been extended to jointly optimize the secondary nicking sgRNA in PE3 systems. This extension addresses a limitation in many prior models, which focus solely on pegRNA design and ignore the impact of nicking sgRNAs on editing efficiency.

In this framework, for each candidate pegRNA, OPED identifies all potential sgRNAs on the non-edited strand within a user-defined range, set by the sgRNA-nick-to-pegRNA-nick distance constraints. This ensures that only sgRNAs positioned appropriately relative to the pegRNA are considered, which is critical for efficient PE3 editing.

The pegRNA score from OPED is combined with a step-by-step rule system to pick the best sgRNA for the PE3 setup. By optimizing both components, OPED significantly enhances the overall editing efficiency in PE3 systems, outperforming approaches that treat pegRNA and sgRNA design separately.

6. OPEDVar Database

The OPEDVar database was constructed to provide a comprehensive resource of optimized prime-editing (PE) designs for pathogenic human variants. It is built from 77,738 ClinVar pathogenic variants filtered for uniqueness and clinical significance. For each variant, the researchers generated every possible pegRNA and sgRNA design that fit the key biological rules, such as PAM type, edit-to-nick distance, and PBS/RTT lengths, so only realistic and potentially effective designs were considered.

OPED then ranked all candidate designs using predicted editing scores, producing a top-10 list of optimized pegRNAs and their corresponding sgRNAs for each variant. This ranking allows researchers to select the most promising candidates while retaining alternative designs in case experimental conditions require adjustments. The database covers single-nucleotide variants (SNVs), insertions, and deletions, providing broad coverage of mutation types.

To make the tool easy to use, OPEDVar can be searched by ClinVar allele ID, gene ID, gene symbol, or HGNC ID. This lets users quickly find the top 10 optimized prime-editing designs for any pathogenic mutation. By combining predicted editing efficiency with essential biological constraints and ranking all possible design options, OPEDVar provides a reliable and comprehensive resource for researchers who want to apply prime editing across large numbers of human disease-causing variants.

7. OPED Web Application

To make OPED accessible and user-friendly, a web application was developed that allows researchers to input either a DNA sequence or a genomic coordinate along with the desired edit. Users can then set key hyperparameters, including PAM type, PBS length, RTT length, and the distance from the edit site to the pegRNA nick site.

Once these parameters are provided, the application identifies all candidate pegRNAs and sgRNAs and ranks them according to predicted editing efficiency using the OPED model. The result is a streamlined, ranked list of optimized designs, enabling researchers to quickly select the most promising pegRNAs and corresponding sgRNAs for experimental testing.

This tool is useful for both basic research and therapeutic applications, as it simplifies the process of designing high-efficiency prime editing experiments and ensures that biological constraints are consistently met. By integrating prediction, ranking, and filtering in a single interface, the OPED web application helps bridge the gap between computational modelling and practical lab work.

8. Discussion and Applications

The development of OPED and its associated tools represents a significant advance in the field of prime editing. By integrating sequence context, topological embeddings, positional information, and biological constraints, OPED can reliably predict the efficiency of candidate pegRNAs and sgRNAs, ranking them to guide experimental design. This has several important implications for both research and therapeutic applications.

8.1 Importance for Research

OPED enables researchers to design prime editing experiments more efficiently and accurately. Traditional pegRNA design relied heavily on heuristics, manual optimization, or small datasets, which limited success rates and reproducibility. With OPED, researchers can rapidly generate candidate pegRNAs and sgRNAs for thousands of variants, confident that these designs meet necessary constraints. Experimental validation using HEK293T cells, plasmid transfection, cell selection, genomic DNA extraction, and targeted deep sequencing showed that OPED-optimized pegRNAs consistently outperformed baseline designs, demonstrating the practical value of the model in real lab conditions.

8.2 Therapeutic Potential

The OPEDVar database further extends OPED’s utility by providing pre-ranked top-10 pegRNA and sgRNA designs for over 77,000 ClinVar pathogenic variants, covering single-nucleotide variants (SNVs), insertions, and deletions. Researchers and clinicians can access optimized designs by gene, allele ID, or HGNC ID, enabling targeted editing for disease-associated variants. Combined with the web application, OPED allows rapid translation from variant identification to experimental or therapeutic design, potentially accelerating the development of gene therapies.

8.3 Advantages Over Alternative Methods

Compared to traditional heuristics and existing machine learning approaches, OPED offers superior predictive accuracy, robustness across different prime editing systems (PE2, PE3, PEmax), and the ability to jointly optimize secondary nicking sgRNAs. By handling variable RTT lengths, edit-to-nick distances, and other constraints, OPED is applicable to a wide range of experimental conditions, reducing trial-and-error in lab workflows and minimizing wasted resources.

8.4 Broader Implications

The accessibility of the OPED web tool means that even researchers without extensive computational backgrounds can leverage high-quality prime editing designs. This democratization of prime editing technology has the potential to accelerate both basic research and translational applications, bridging the gap between computational design and experimental execution. Additionally, the ability to optimize both pegRNAs and sgRNAs enhances the efficiency of PE3 systems, which is crucial for experiments targeting complex genomic edits or therapeutic interventions requiring high precision.

In summary, OPED and its associated database and web application provide a powerful, validated, and user-friendly framework for high-efficiency prime editing. By addressing limitations in existing design methods and providing experimentally validated guidance, OPED represents a major step forward in enabling precise and reliable genome editing.

9. Limitations and Future Directions

While OPED represents a substantial advance in prime editing design, several limitations remain. One key challenge is data bias. OPED was trained on existing datasets of editing outcomes, which may not fully capture the diversity of all genomic contexts or variant types. As a result, its predictions could be skewed toward the sequences and experimental conditions present in the training data, potentially limiting performance on less common variants or unusual genomic regions. Expanding the training data with more diverse sequences and experimental results could further improve the generalizability of the model.

Another limitation lies in cell-type specificity. Much of the experimental validation described used human embryonic kidney cells or HEK293T cells. While these cells provide a controlled system for testing, prime editing efficiencies can differ across cell types due to variations in the cellular environment. Future studies applying OPED-designed pegRNAs in a wider range of cell types, including primary cells or clinically relevant tissues, will be crucial for understanding how broadly its predictions translate to different biological systems.

Long RTT sequences also present a challenge. Although OPED can handle variable PBS and RTT lengths, efficiency tends to decrease as RTT length increases, especially in complex or repetitive genomic regions. Further refinement of the model to better account for long RTTs, potentially by incorporating additional sequence features or thermodynamic constraints, could enhance its predictive power.

Finally, there is room for improvement in off-target prediction and integration with newer editing systems. Currently, OPED focuses on optimizing pegRNA and nicking sgRNA design for PE2, PE3, PEb, and PEmax, but future iterations could integrate off-target prediction to reduce unintended edits. Additionally, emerging systems like PEmax2 or PASTE could benefit from OPED-style optimization, extending its utility to next-generation prime editing technologies. These directions suggest that while OPED already streamlines and improves prime editing, its evolution could further enhance precision, efficiency, and versatility across research and therapeutic applications.

10. Conclusion

Prime editing has emerged as a transformative tool for precise genome editing, offering the ability to install, correct, or remove specific variants without inducing double-strand breaks. Despite its promise, the technology has historically faced challenges due to the difficulty of designing efficient pegRNAs and predicting outcomes reliably. Poorly designed pegRNAs can lead to low editing efficiency, wasted experiments, and slower progress in both research and therapeutic applications.

OPED represents a major step forward in addressing these challenges. By learning from thousands of real editing outcomes, capturing sequence context, topological embeddings, positional information, and biological constraints, it converts prime editing from a process that once relied heavily on trial-and-error into a more systematic, engineering-driven approach. Its ability to rank optimized pegRNAs and nicking sgRNAs, validate them experimentally, and provide accessible tools via OPEDVar and the web application significantly accelerates the design process.

Beyond efficiency improvements, OPED lays the groundwork for broader applications of prime editing. By enabling reliable design across a wide range of pathogenic variants, it facilitates research into genetic disease correction and potentially paves the way for large-scale therapeutic interventions. While limitations remain, such as data bias, cell-type specificity, and challenges with long RTT sequences, OPED’s framework offers a scalable and generalizable approach that brings the vision of treating genetic diseases with precision editing closer to reality.

Sources

Liu, F., Huang, S., Hu, J. et al. Design of prime-editing guide RNAs with deep transfer learning. Nat Mach Intell 5, 1261–1274 (2023). https://doi.org/10.1038/s42256-023-00739-w

Nelson, J. W., Randolph, P. B., Shen, S. P., Everette, K. A., Chen, P. J., Anzalone, A. V., An, M., Newby, G. A., Chen, J. C., Hsu, A., & Liu, D. R. (2022). Engineered pegRNAs improve prime editing efficiency. Nature biotechnology, 40(3), 402–410. https://doi.org/10.1038/s41587-021-01039-7

Scholefield, J., & Harrison, P. T. (2021). Prime editing — an update on the field. Gene therapy, 28(7–8), 396–401. https://doi.org/10.1038/s41434-021-00263-9

Anzalone, A. V., Randolph, P. B., Davis, J. R., Sousa, A. A., Koblan, L. W., Levy, J. M., Chen, P. J., Wilson, C., Newby, G. A., Raguram, A., & Liu, D. R. (2019). Search-and-replace genome editing without double-strand breaks or donor DNA. Nature, 576, 149–157. https://doi.org/10.1038/s41586-019-1711-4

Imagine that the fight against genetic diseases and cancers could be won, not just in a lab, but on a computer screen. Imagine having the ability to edit the blueprint of human life with the certainty and precision of a surgical tool. That was the core of my project.

I used a powerful computational platform called Benchling to design a highly precise edit in the BRCA2 gene (Breast Cancer gene 2). This gene is crucial because it acts as the cell’s DNA repairman; when it fails, the risk of cancers like breast and prostate cancer increases drastically.

By using computational design to target and fix a single mutation in the BRCA2 blueprint, my work demonstrates how, on a larger scale, AI and synthetic biology are engineering the solutions to complex genetic diseases and mutations. We are moving beyond simply treating symptoms and creating the precise molecular tools required to address genetic flaws right at their source.

I. Step 1: The AI-Driven Quality Check (The High-Scoring GPS)

The first and most critical step in any CRISPR-based therapy is selecting the perfect location for the “surgical cut.” We utilized the CRISPR system’s Guide RNA (gRNA), which serves as the molecular GPS that directs the Cas9 enzyme to the target DNA. However, it is important to find a safe and effective gRNA.

The problem is that numerous potential gRNAs exist, but many are either too inefficient to make a clean cut or, worse, risky because they possess the potential to accidentally cut the wrong gene — an undesirable “off-target” effect.

I didn’t guess; I used AI-driven scoring to guarantee the quality of the edit. I selected a guide based on two primary metrics:

1. Efficiency: I chose a guide with a high On-Target Score of 68.7 (above the required 60 threshold) to ensure the cut was highly efficient and effective.
2. Safety: More critically, I selected a guide with an excellent Off-Target Score of 73.5 (above the required 50 threshold), guaranteeing maximum specificity.

This digital quality control is incredibly important as it ensures the cutting enzyme only operates at the designated spot on the BRCA2 gene, validating the tool before proceeding to the next step.

II. Step 2: The Precise Edit (Designing the DNA Patch)

With the cutting tool secured, the next challenge was to design the new genetic instructions, the Homologous Recombination (HR) Template, which acts as the clean DNA patch used by the cell to repair the cut site.

The goal was to model a precise therapeutic edit. I successfully designed the template to introduce a specific point mutation in the BRCA2 sequence. This change was modelled on the DNA level, where I changed the letter Adenine (A) to Thymine (T) at the precise location of 32,319,199. This single-base substitution results in a desirable amino acid change in the resulting protein, replacing Threonine (T) with Serine (S), a clear demonstration of the ability to enact a controlled, precise genetic modification. The synthetic HR Template carries these corrected, precise instructions for insertion.

III. Step 3: The Surgical Safety Lock (Preventing Recutting)

The final, and perhaps most sophisticated, design step was to ensure the edit was permanent. After the cell successfully uses the HR template to repair the cut, the active Cas9 enzyme would normally still be able to recognize and recut the newly repaired BRCA2 gene, undoing the work.

To prevent this destructive cycle, I engineered a surgical safety lock. The Cas9 enzyme needs a small adjacent sequence called the PAM sequence (Protospacer Adjacent Motif) as a handle to make its cut. I designed a silent mutation within this PAM sequence on the HR Template. This change in the template makes the new DNA invisible to the cutting enzyme, but it does so without altering the desired protein change. This final computational trick ensures the A 🠮 T edit is stable and permanent, protecting the fixed BRCA2 gene forever.

IV. Conclusion: My Mind-Blowing Introduction to the Future of Medicine

As a complete beginner to CRISPR design, finishing this project was truly mind-blowing. In just a few minutes, I finished designing the three main molecular parts needed for a stable, high-precision edit on the critical BRCA2 cancer gene. It’s important to remember that my project is very small compared to the immense complexity of real lab work and clinical trials. However, the core idea is clear: Synthetic Biology is changing science. We are no longer just watching life; we are now confidently engineering it.

The way computer tools (AI) work perfectly with advanced science gives us the precision we need to tackle terrible, complex diseases like cancer and other genetic disorders. This project solidified my belief in the incredible potential of this technology to completely change medicine. It shows that future researchers will have the speed and surgical accuracy required to truly fix genetic flaws right where they start.

Imagine a mother in rural Kenya, watching the light of her life, the child she held in her womb for 9 months, diminish with each passing day. This innocent soul, whose life had just barely begun, doomed to the tight grasp of Tuberculosis. She gives her child multiple antibiotics a day, holding onto a sliver of hope, but the doctors tell her the drugs they have might not work against this particular strain. By the time more effective medicines could be developed through traditional methods, her children might already be orphans. This isn’t a hypothetical scenario; it’s the daily reality for millions across the bottom billion of our global population.

But what if I told you Tuberculosis doesn’t have to take any more lives? That because of recent advancements in the fields of AI and synthetic biology, scientists and researchers now have the ability to build tools to predict protein structures at incredibly fast rates, never before seen, allowing them to use that information to design biological therapies that are more effective and affordable than current treatments. What does any of that mean, though and how does it work? That is what I will be exploring for the duration of this article: how these incredible breakthroughs in technology can offer hope for millions of people who have been let down by traditional drugs.

What is Tuberculosis

For those of you who don’t know, TB or Tuberculosis is a result of the bacterium Mycobacterium tuberculosis. When it enters the body, the immune system traps it inside macrophages, forming granulomas, tiny “cages” to stop it from spreading, but this doesn’t mean that it is eradicated from the body; it is simply “sleeping,” sometimes for years or decades.

Latent vs. Active TB

If the immune system weakens (from HIV, pregnancy, or other conditions), the bacteria can “wake up” again, turning into active TB, which damages the lungs and spreads to others. So:

• Latent TB: contained, but not killed.
• Active TB: spreads and needs antibiotics.

The Human Toll of TB

Because TB hides inside cells with a thick, waxy wall, treatment can take 6–12 months, and sometimes longer when the bacteria become drug-resistant.

One patient particularly describes “taking 22 pills a day”, often with harsh side effects like nausea, fatigue, and liver damage, which makes treatment a painful struggle. Many avoid or abandon therapy because of the toll it takes on their bodies and lives.

This tragedy kills around 1.3 million people annually (Partners In Health), which means that every 21 seconds, a wife is widowed, or a child is orphaned, left to brave the world without the comfort and guidance of their parents. Additionally, over 80% of affected people are in low and middle-income countries(World Health Organization). Change isn’t optional; it’s urgent.

Current Challenges in Drug Development

Traditional drug development is slow and expensive, with the top 20 pharma companies spending about $60 billion annually on R&D (McKinsey & Company). The cost to bring one drug to the market has risen to $2.6 billion, a 140% increase (McKinsey & Company) over the last decade. Bringing these drugs to the people who need them takes an average of 10–15 years (National Institutes of Health.

Introduction to Synthetic Biology

Synthetic Biology offers a cost-effective and efficient solution to that. It is a discipline that combines concepts from:

• Biology🧬
• Engineering 🏗️
• Chemistry ⚗️
• Computer Science 💻
• Mathematics 📐

And related fields to build organisms with new functions or redesign existing ones for useful purposes, creating biologically-based solutions.

Engineering with Biological “Lego Bricks”

One of the biggest advances was the ability to synthesize DNA from scratch, allowing scientists to build fully new bits and systems. The modularity of biological parts, such as enzymes, genes, and proteins, has allowed researchers to treat each of them like ‘Lego bricks.’ Each ‘brick’ has a specific function, but can be connected in different combinations to create entirely new biological systems. This approach has made it possible to design targeted treatments faster and more precisely, an important step for diseases like Tuberculosis, where, currently, traditional drug development is not sufficient.

But to truly engineer biology, we need to understand the smallest working pieces, especially proteins, which carry out most of the work inside cells.

Proteins and Their Importance

Proteins are incredibly complex macromolecules or polymers built from amino acid chains that execute various tasks throughout our body. These proteins go through stages as a result of a process called protein folding, which causes them to fold into different 3D structures.

• Primary structure: A protein is in a linear sequence of amino acids (it is like the blueprint for a house)
• Secondary Structure: Hydrogen bonds form between atoms, and the protein folds into alpha helices or beta-sheets (walls, staircases, etc.)
• Tertiary Structure: The alpha helices or beta sheets fold into each other, forming geometrical, 3D shapes (a complete house)
• Quaternary Structure: two or more separate protein chains (each with its own primary → secondary → tertiary structure) come together to form a larger functional complex. (neighbourhood or housing complex)

For TB, knowing how its proteins fold could reveal why some drugs fail and where new therapies could be improved.

Protein Folding: A Key Challenge

Protein folding isn’t a simple or predictable process; it’s an incredibly sensitive one. The way a protein folds determines its function, and even a tiny error like a single misplaced amino acid can cause the protein to misfold, which in turn can completely change how it behaves or potentially lead to serious diseases or failed drug interactions. Think of it like a puzzle; even one misplaced piece can throw off the entire picture.

For decades, scientists have tried to figure out protein structures using methods like X-ray crystallography or cryo-electron microscopy. These techniques were accurate but incredibly slow and expensive, sometimes taking years just to map a single protein.

This inefficiency made drug discovery difficult and expensive. To design a drug that works, scientists need to know the exact 3D shape of the protein they’re targeting, how it folds, where its active sites are, and what might bind to it. Trying to synthesize a drug without that information would be next to impossible.

AI and Protein Structure Prediction

That’s where AI changed everything. Models like AlphaFold, created by DeepMind, a deep learning system, can now predict a protein’s structure with minimal information and revolutionary accuracy. What used to take years can now be done in hours, if not minutes and currently the AlphaFold database provides open access to over 200 million protein structure predictions.

Now, researchers have access to the predicted structures of proteins like Mycobacterium tuberculosis, the bacteria behind TB. For the first time, researchers can see the exact molecular shapes they need to target, enabling the creation of faster, smarter, and more effective treatments.

This is where synthetic biology and AI truly allow for change in the medical industry. AlphaFold and other computer programs don’t just observe proteins; they give synthetic biologists precise blueprints that they then use to construct new ones. Synthetic biology, in turn, brings these designs to life, verifying and engineering them in living organisms and cells. The information generated by such experiments in real life subsequently feeds back into AI systems, teaching them to be even more precise. This forms a strong feedback loop: AI designs → testing biology → biology feeds back into AI, speeding up the discovery rate higher than either could on its own.

AlphaFold

AlphaFold is a revolutionary deep learning protein prediction model created by DeepMind. Like every deep learning system, it was trained on millions of pieces of protein structures that have already been experimentally solved. It requires you to input the sequence of amino acids in a protein, and from there, it applies complex neural network models to accurately generate the 3D structure. When AlphaFold’s predicted protein structures were compared to experimentally known structures, the average difference between the positions of backbone atoms was only 0.96 Å, which is less than the width of a single atom. This accuracy was unprecedented as it significantly reduced the time it took to map the structure from years to hours, if not minutes.

How It Works

AlphaFold’s breakthrough comes from its ability to detect patterns in evolution. Proteins often mutate over time, but if two parts of a sequence tend to change together across different species, that usually means they fold close together in the 3D structure. AlphaFold uses this idea to figure out which amino acids are likely to interact, even if the protein has never been studied in a lab before.

To do this, it first builds a multiple sequence alignment (MSA), basically a massive comparison of thousands of related protein sequences. By looking across these sequences, the model identifies co-evolving amino acids that probably contact each other in the folded protein.

Example MSA Diagram:

Seq 1: A P L K T V G L Y R Q N

Seq 2: A P L K T I G L Y R Q N

Seq 3: A P L K S O G L Y R Q N

Seq 4: A P L K T I G L Y R Q D

In this made-up example, notice positions 5 and 10:

• Position 5: changes from T → S,
• Position 10: changes from R → K.

Patterns like this suggest these amino acids might interact or fold close together in the protein’s 3D shape. AlphaFold looks for thousands or even millions of these co-evolving pairs across huge datasets of related sequences, using them as clues to predict how the protein folds, even if that exact protein has never been studied in a lab before.

Next, AlphaFold feeds that information into something called the Everformer network. It is essentially a deep learning system that predicts how the amino acids interact. Then it creates an initial ‘mock up’ of the protein’s structure and repeatedly refines it to make sure that the protein is actually physically and chemically possible. This is what has finally given researchers the ability to see the complex proteins that make up TB in detail, which is a huge step for designing more effective treatments.

Confidence Scores: pLDDT and PAE

Another feature is that AlphaFold provides confidence scores for each part of the protein, called pLDDT and PAE.

• The pLDDT score (Predicted Local Distance Difference Test) tells scientists how confident the model is about the position of each amino acid; higher scores mean the predicted location is very likely correct.
• The PAE score (Predicted Aligned Error), as its name suggests, shows how much AlphaFold expects there to be an error with the alignment of two parts of a protein. This helps to understand the overall reliability of larger structures and also plays a part in refining the protein structure.

These scores help scientists and researchers identify which parts of a prediction are accurate and which parts may need more confirmation. Implementing these systems and design choices in AlphaFold are what make the model so extremely accurate, and things like these confidence scores are especially valuable when dealing with proteins like TB, where even the smallest errors could cost lives.

Why it Matters for TB Treatment

A recent project by researchers at the University of Washington’s Institute for Protein Design showed how synthetic biology can create entirely new proteins that don’t exist in nature but can still fight disease. The team built small “minibinder” proteins that attach tightly to the spike protein of SARS-CoV-2, the virus that causes COVID-19. These synthetic proteins were engineered to block the virus from entering human cells. What made this breakthrough so powerful wasn’t just its success against COVID-19, but what it represented: the ability to rapidly design a new therapeutic protein in weeks rather than years.

That same approach could be applied to tuberculosis. Many of TB’s key enzymes and structural proteins have already been mapped using tools like AlphaFold, meaning scientists now know their shapes and weak points. By designing synthetic proteins that bind to and disable them, similar to how the minibinders blocked the coronavirus spike, researchers could one day build new treatments for drug-resistant TB.

Every experiment run in the lab, every protein that behaves differently than predicted, provides new data for machine learning systems to learn from. That feedback helps future AI models better understand the complex “rules” of protein folding and function, closing the loop between digital prediction and biological reality. With every iteration, both AI and synthetic biology get smarter, and drug development gets faster.

A similar method or approach could be used to develop more cost-effective and efficient treatments. As I mentioned earlier, traditional drug development takes years and involves a lot of trial and error, which is extremely expensive and exceptionally slow. Using AI and synthetic biology, researchers can create and test designs of proteins, drugs, etc., all virtually before trying to physically make them. This would reduce the number of failed experiments, which would speed up drug development and lower costs. If scaled, it could make treatments more affordable and accessible, potentially helping millions by providing new or better therapies.

Challenges that we might face in bringing this to the market

While this may all sound great, there are still obstacles that need to be overcome before these treatments can reach those who need them. Many of these projects are still in early or pre-clinical stages and haven’t been tested on humans just yet. In addition to that, like every other new drug in the market, these treatments would require regulatory review for safety and efficacy evaluation from the government, which often takes several months to a year.

Another potential issue could be cost and accessibility. Even with AI and synthetic biology making drug development faster and cheaper to design, equal global access can still be an issue. A major problem that could create a barrier lies in intellectual property rights, when large pharmaceutical companies hold patents on discoveries. It gives them the right to decide who can make or sell these treatments, and this often slows the process of getting the medicine to the people in communities that actually need it. While patents safeguard the rights of many people and also reward innovation, they can also unintentionally limit access.

Call to action

But these breakthroughs, without a human heart behind them, will do no good. For these cures to actually reach the people who need them, it requires collaboration between governments, global health organizations, and private companies alike to warrant fair licensing agreements. It would also require global collaboration to ensure these treatments are distributed fairly and are actually accessible to everyone.

That child who just left his parents behind? That wife who left behind a mourning family? Let them be the last to fall. Let the world see a new type of change. A change that lifts the weight off someone’s shoulders so they don’t have to wake every day, only to resent life, watching themselves become a shell of a person they were.

As the world moves forward, lest we forget every life lost to disease, but instead remember them and imagine their loss and let that push the world towards a change that isn’t just a necessity but vital for the survival of all humanity.

Sources

Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2

World Health Organization. (2024). Tuberculosis. https://www.who.int/news-room/fact-sheets/detail/tuberculosis#:~:text=Over%2080%25%20of%20cases%20and,every%20part%20of%20the%20world

Partners In Health. (2023). TB is the deadliest infectious disease — so why haven’t you heard of it? https://www.pih.org/article/tb-deadliest-infectious-disease-so-why-havent-you-heard-it

Jain, A., & Mondal, R. (2020). Current treatment of tuberculosis and pharmacological challenges. Journal of Family Medicine and Primary Care, 9(6), 2820–2826. https://pmc.ncbi.nlm.nih.gov/articles/PMC7342338/

University of Washington Medicine. (2022). Designed antiviral proteins inhibit SARS-CoV-2 in the lab. https://newsroom.uw.edu/news-releases/designed-antiviral-proteins-inhibit-sars-cov-2-lab

McKinsey & Company. (2023). The pursuit of excellence in new drug development. https://www.mckinsey.com/industries/life-sciences/our-insights/the-pursuit-of-excellence-in-new-drug-development