🧬 biology

Shetti-AI: Precise prediction of protein sequence mutational landscapes to accelerate functional assays

Shetti-AI is a computational framework that integrates physicochemical properties, genetic distances, and generative adversarial networks to predict protein mutational landscapes and generate optimized motif-like variants, thereby accelerating functional assays and reducing experimental search spaces in synthetic biology and virology.

Original authors: Haitham Sobhy

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haitham Sobhy

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Proteins are the workhorses of life, tiny machines built from chains of amino acids that fold into complex shapes to perform specific tasks. Within these long chains, short sequences of just a few amino acids often act as critical switches or docking points, known as motifs. These small regions dictate how a protein interacts with other molecules, and a single change in one of these letters can alter the protein's entire behavior. Scientists have long known that to understand a protein, they must understand these motifs. However, testing every possible variation of these short sequences in a laboratory is a slow, expensive, and often impossible task. The space of potential changes is simply too vast to explore by hand, leaving researchers to guess which variations might work and which will fail.

To solve this, a researcher named Haitham Sobhy has developed a new computational tool called Shetti-AI. This system is designed to predict which changes to a protein's short motifs are likely to be successful before a scientist ever enters the lab. Instead of relying solely on how nature has evolved proteins in the past, Shetti-AI looks forward, simulating how a protein might change in the future. It starts with a known, working motif and generates a list of new, similar candidates. The system does this by balancing two main factors: how easy it is for a cell's genetic code to make the change, and how well the new amino acids preserve the physical properties needed for the protein to function. By filtering out unlikely options and highlighting the most promising ones, the tool aims to shrink the massive search space down to a manageable list of experiments.

The paper highlights a specific problem with current artificial intelligence tools used in biology. Many modern programs, often called protein language models, are excellent at spotting general patterns in vast amounts of genetic data. However, they struggle with the fine details of these short motifs. For instance, these tools might treat two very different sequences as nearly identical because they look similar on a broad scale, even if a tiny change in one of them would break the protein's ability to bind to its target. The author points out that these existing tools often miss the subtle physical differences between amino acids, such as swapping a charged particle for a neutral one, which can completely destroy a protein's function. Shetti-AI was built specifically to fix this blind spot by focusing on the precise physical and genetic rules that govern these small, critical regions.

The framework operates by first taking a validated motif, a short sequence known to work in nature, and mapping its properties into a detailed numerical profile. This profile considers eighteen different physical characteristics of the amino acids, such as their size, charge, and how they interact with water. The system then uses a series of mathematical models to generate new variations. One part of the system acts as a conservative filter, checking if a single change is genetically accessible and physically similar to the original. Another part uses a more advanced, generative approach to imagine entirely new combinations of amino acids that might not exist in nature yet but still follow the rules of physics. These generated candidates are then ranked based on a score that weighs their likelihood of working against the difficulty of creating them genetically.

In testing the system, the researchers used a known motif involved in regulating human proteins as a benchmark. The tool successfully identified variations that are known to function in nature, such as sequences that differ by just one or two letters but maintain the same shape and behavior. More importantly, the generative part of the system suggested new, novel sequences that shared the same physical properties as the original motif. When compared to a control sequence that was known to be different, the tool correctly identified that the new candidates were much closer to the working original. The results showed that while some models were better at finding small, safe changes, others were capable of exploring more distant possibilities, offering a range of options depending on what the researcher needed.

The paper suggests that this approach could be particularly useful for studying emerging viruses, where scientists often have very little time to test new variants. By using a known motif from a related virus as a starting point, the tool could quickly generate a list of likely candidates to test in the lab, speeding up the discovery process. It also offers a way to design custom proteins for synthetic biology, allowing researchers to create new tools with specific physical properties without having to synthesize thousands of failed versions. The author notes that the system is flexible; researchers can adjust the settings to be very strict, looking only for changes that are almost identical to the original, or more exploratory, looking for bold new designs.

Ultimately, Shetti-AI does not replace the need for laboratory experiments, but it changes the order in which they happen. Instead of testing random guesses, scientists can use the tool to prioritize the most promising candidates, saving time and resources. The system bridges the gap between the rigid rules of genetics and the fluid possibilities of protein design, offering a way to navigate the vast landscape of potential mutations with confidence. By focusing on the physical reality of how proteins work, rather than just statistical patterns, the tool provides a clearer path to understanding how life adapts and how we might engineer it for the future.

Technical Summary: Shetti-AI

Problem Statement
Protein motifs are critical for dictating molecular interactions, structural organization, and function. While pathogens often alter short motifs to acquire new host dynamics without changing entire protein structures, experimentally screening the vast space of sequence variants is slow, expensive, and resource-intensive. Current machine learning tools, particularly protein language models (pLLMs) like ProtT5, ProtBERT, and ESM-2, rely heavily on historical evolutionary conservation. The paper argues these tools fail at precise motif-level prediction because they cannot model codon-level DNA mutation accessibility, evaluate position-specific biophysical constraints, or distinguish between functionally distinct variants (e.g., failing to recognize that substituting a charged Arginine with a non-polar Leucine in an RGD motif disrupts integrin binding). Consequently, there is a need for a framework that moves beyond retrospective conservation to prospectively predict viable mutational landscapes.

Methodology
Shetti-AI is a computational framework designed to generate and prioritize motif-like variants starting from curated, experimentally validated biological references. The framework integrates three distinct architectures operating on an 18-dimensional physicochemical property matrix ( $P_{phys}$ ) and a codon substitution accessibility matrix ( $P_{genetic}$ ):

18-Dimensional Physicochemical Mapping: Amino acids are mapped to continuous numerical vectors combining normalized quantitative biophysical properties and ordinally encoded qualitative attributes. This facilitates gradient-based learning and generative sampling.
Codon-Aware Substitution Accessibility: A modular 20x20 matrix encodes genetic distance based on minimum codon substitution steps, assigning weights to accessibility (e.g., 1-step mutations are weighted higher than 3-step mutations). This allows for taxon-specific adjustments without altering the inference engine.
Three-Part Architecture:
- Architecture 1 (Convex Linear $\lambda$ -Baseline): A lightweight, non-generative model that evaluates point mutations via a convex combination of genetic accessibility and biophysical conservation ( $P_{final} = \lambda P_{DNA} + (1-\lambda)P_{phys}$ ). It serves as a baseline for single-variable trade-offs.
- Architecture 2 (Log-Space Bayesian Hybrid): A deterministic model that ranks single-residue substitutions using an empirical Bayesian posterior probability. It combines normalized codon accessibility with biophysical similarity scores in log-space to avoid numerical underflow while preserving candidate ranking.
- Architecture 3 (Generative GAN/WGAN-GP Engine): A continuous deep generative model (Conditional GAN and WGAN-GP) that expands the search space beyond single-residue variants. It generates novel candidate motifs from a reference using physicochemical representations and latent noise. Crucially, these candidates are post-filtered and ranked using the separate Bayesian constraint to ensure biological plausibility and mutation reachability.

Key Contributions

Prospective Prediction: Unlike tools relying on historical conservation, Shetti-AI predicts future variant fates and mutational landscapes before they appear in nature or the lab.
Hybrid Constraint System: The framework uniquely combines a Bayesian genetic prior with an 18-dimensional physicochemical likelihood, utilizing user-controlled $\lambda$ -weighting and WGAN-GP loss integration.
Modular Design: The separation of the Bayesian constraint from the neural networks allows biological assumptions to be inspected, modified, or replaced independently of the generative engine.
Open Source Implementation: The source code is available under the AGPL-3.0 license, implemented in Python using PyTorch for the deep learning components.

Results
Using the M-R-EL motif (a tubulin autoregulatory signature) as a benchmark, the framework demonstrated the following:

Conservative Prediction: The Bayesian and $\lambda$ -baseline models successfully predicted closely related variants such as MREI and MRDL, which are canonical sequences for the motif.
Generative Discovery: The GenAI components (GAN and WGAN) explored and generated new motifs closely related to the physicochemical characters of the reference MREL.
Differentiation: When compared to the TREL control (found in SARS-CoV), the framework correctly identified that the Threonine in TREL possesses different physicochemical characters, while generated variants remained similar to the reference MREL.
Model Performance: The WGAN generally outperformed the standard GAN, though both predicted the LKDM motif with similar Bayesian scores.
Parameter Tuning: The study showed that adjusting parameters (temperature, $\beta$ , and $\lambda$ ) allows the framework to shift between strict biophysical conservation (useful for viral scanning) and broad explorative generation (useful for synthetic biology).

Significance and Claims
The paper claims that Shetti-AI transforms sequence optimization from an "exhaustive guessing game" into a targeted, biophysically grounded pipeline. By bridging deterministic statistical constraints with continuous generative deep learning, the framework reduces the experimental search space for synthetic biology, emerging viral proteins, and functional assays.

The author emphasizes that Shetti-AI does not merely generate theoretical constructs but produces practical, experimentally tractable tools by enforcing genetic reachability and biophysical constraints. The framework is positioned as a high-throughput, cost-effective solution to accelerate functional assays, particularly for:

Synthetic Biology: Designing custom functional motifs with predictable physical properties.
Therapeutics: Aiding gene therapy (AAV vectors), immune neutralization, and CAR-T engineering.
Virology: Rapidly scanning emerging viral sequences (e.g., SARS-CoV-2) for motif-like regions when experimental data is limited.

The paper concludes that by narrowing the vast sequence space to a prioritized, biologically grounded candidate pool, Shetti-AI reduces the time and resources required for experimental validation while providing insights into evolutionary adaptation and novel functional interactions.

Technical Summary: Shetti-AI

More like this