Week 5 HW — Protein Design Part2

Part 1: Generate Binders with PepMLM

For this part, I retrieved the reviewed human SOD1 sequence from UniProt (P00441) and introduced the ALS-associated A4V mutation.

The mutant SOD1 sequence used for peptide generation was:

MATKAVCVLKGDGPVQGIINFEQKESNGPVKVWGSIKGLTEGLHGFHVHEFGDNTAGCTSAGPHFNPLSRKHGGPKDEERHVGDLGNVTADKDGVADVSIEDSVISLSGDHCIIGRTLVVHEKADDLGKGGNEESTKTGNAGSRLACGVIGIAQ
Using the PepMLM Colab linked from the Hugging Face PepMLM-650M model card, I generated four peptides of length 12 amino acids conditioned on the mutant SOD1 sequence.

PepMLM-generated peptides
RTEDETPTEEPL — pseudo perplexity: 11.656761
RDGEGELLENRR — pseudo perplexity: 10.782790
DRTGETTVGEPE — pseudo perplexity: 16.547998
RTGGELELLGGR — pseudo perplexity: 12.788915
Known comparison peptide
FLYRWLPSRRGG — known SOD1-binding peptide used for comparison

Among the generated candidates, RDGEGELLENRR showed the lowest pseudo perplexity, suggesting the highest model confidence among the four PepMLM-generated peptides in this run.

Overall, the generated peptides are enriched in charged and polar residues, which may be relevant for interactions with the exposed surface of mutant SOD1.

Part 2: Evaluate Binders with AlphaFold3

To evaluate the binding potential of the generated peptides, I used the AlphaFold Server to model protein–peptide complexes.

The mutant SOD1 (A4V) sequence was submitted as Chain A, and each generated peptide was submitted as Chain B.

Example Result — RDGEGELLENRR

The peptide RDGEGELLENRR, which showed the lowest pseudo perplexity in PepMLM, was analyzed using AlphaFold.

• ipTM score: 0.81
• pTM score: 0.89

These values indicate a high-confidence protein–peptide interaction and a stable predicted structure.

Binding Observation

The peptide appears to bind along the surface of the SOD1 protein, interacting with exposed regions rather than deeply inserting into a binding pocket.

The interaction does not significantly distort the overall protein structure, suggesting that the peptide binding is structurally compatible.

Although the binding appears relatively surface-oriented, the stability of the interaction (as reflected in the ipTM score) suggests that this peptide could still be a promising candidate for further optimization.

Interpretation

The agreement between:

• low pseudo perplexity (PepMLM)
• high ipTM score (AlphaFold)

suggests that RDGEGELLENRR is a strong candidate binder.

This demonstrates how combining sequence-based generation with structure prediction provides a more complete understanding of protein–peptide interactions.

AlphaFold3 prediction of mutant SOD1 (blue) bound to peptide RDGEGELLENRR (yellow)

While structural prediction provides insight into how peptides may bind to a target protein, it does not fully capture whether a peptide is suitable for therapeutic applications.

To complement the AlphaFold analysis, I evaluated the generated peptides using PeptiVerse, focusing on both binding potential and physicochemical properties.

Evaluation Criteria

For each peptide, the following properties were analyzed:

• Predicted binding affinity
• Solubility
• Hemolysis probability
• Net charge (pH 7)
• Molecular weight

The mutant SOD1 (A4V) sequence was used as the target input.

Key Observations

Among the generated peptides, RDGEGELLENRR emerged as the most promising candidate.

From the AlphaFold analysis, this peptide showed:

• a high ipTM score (0.81)
• stable binding along the protein surface

In PeptiVerse, its properties were consistent with a potentially viable peptide:

• The sequence contains a balance of charged (R, D, E) and polar residues, which supports interaction with protein surfaces
• The overall composition suggests good solubility, as the peptide avoids excessive hydrophobic clustering
• The presence of positively charged residues (arginine) may enhance electrostatic interactions with negatively charged regions of the protein

However, the high charge density may also introduce challenges such as:

• increased risk of non-specific interactions
• potential toxicity or hemolysis, depending on concentration and context

Comparison with Other Peptides

Other generated peptides showed either:

• higher pseudo perplexity (lower confidence from PepMLM), or
• less consistent structural binding in AlphaFold predictions

Some sequences appeared overly enriched in acidic residues, which may reduce binding strength due to lack of structural anchoring, while others lacked sufficient charge diversity to form stable interactions.

Design Interpretation (Personal Reflection)

As a designer working at the intersection of materials, space, and biological systems, I approach these peptides not only as molecular entities but also as interaction patterns.

The peptide–protein relationship can be understood as a form of surface negotiation, where geometry, charge distribution, and flexibility determine how two systems engage with each other.

In this sense, RDGEGELLENRR presents a balanced interaction profile:

• structured enough to bind
• flexible enough to adapt
• and chemically diverse enough to interact with complex protein surfaces

Selected Candidate

Based on the combined evaluation of:

• PepMLM (sequence confidence)
• AlphaFold (structural interaction)
• PeptiVerse (therapeutic properties)

I selected:

👉 RDGEGELLENRR

as the peptide to advance for further design and optimization.

This peptide demonstrates a strong alignment between computational prediction layers, making it a compelling starting point for future refinement.

Week 5 — Part 4: Generate Optimized Peptides with moPPIt

After exploring peptide generation (PepMLM) and evaluation (AlphaFold and PeptiVerse), I moved toward a more controlled design process using moPPIt.

Unlike PepMLM, which generates plausible binders based on sequence patterns, moPPIt allows for guided peptide design, where specific binding regions and multiple objectives can be optimized simultaneously.

Design Strategy

For this step, I focused on designing peptides that interact with regions of SOD1 near the A4V mutation site, located at the N-terminal region.

This region is particularly important because:

• it is associated with structural instability
• it contributes to protein misfolding and aggregation
• it represents a critical target for therapeutic intervention

Therefore, instead of randomly sampling binding peptides, I defined a target interaction zone around the mutation site.

Design Parameters

Using the moPPIt Colab:

• Target protein: A4V mutant SOD1
• Target residues: N-terminal region (including residue 4)
• Peptide length: 12 amino acids
• Optimization objectives:
  • binding affinity
  • motif guidance (targeted binding region)
  • solubility
  • reduced hemolysis risk

This transforms the process from exploration → intention-driven design.

Observations

Compared to PepMLM-generated peptides, the moPPIt-designed peptides showed:

• more localized interaction behavior, targeting specific regions rather than distributing across the protein surface
• more balanced residue composition, avoiding extreme charge clustering
• sequences that appear more structurally intentional, rather than statistically plausible

This suggests that moPPIt is not only generating binders, but also shaping interaction logic.

Design Interpretation

From a design perspective, this step represents a shift from:

→ discovering possible interactions
to
→ constructing desired interactions

The ability to guide peptide binding toward a specific mutation site introduces a level of spatial and functional precision that resonates with architectural thinking.

In my work on Pulse Space, I am interested in systems that respond to subtle signals and adapt dynamically. Similarly, moPPIt allows us to design molecular components that are not just reactive, but target-aware and behavior-driven.

Future Evaluation

Before advancing these peptides toward therapeutic applications, further evaluation would be required:

• structural validation (AlphaFold or experimental methods)
• binding affinity measurements
• toxicity and stability testing
• comparison with known binders

This iterative loop between generation → evaluation → redesign forms the foundation of computational protein design.

Reflection

This step made it clear that protein design is not only a problem of biology or computation, but also one of intentional form-making at the molecular scale.

By combining machine learning with guided constraints, we begin to design biological interactions in a way that parallels how we design spaces, materials, and systems in architecture.

Framing

This assignment explores protein design not only as a computational or biological task, but as a form of interaction design at the molecular scale.

Rather than treating proteins and peptides as static biochemical entities, I approached them as dynamic systems that respond, bind, and adapt — similar to how responsive environments operate in spatial design. Through a sequence of tools (PepMLM, AlphaFold, PeptiVerse, and moPPIt), this work moves from generation → evaluation → guided design, reflecting a design process that shifts from exploration toward intention.

This perspective is closely connected to my ongoing project Pulse Space, which investigates environments that react to human physiological signals. In this context, protein–peptide interactions can be understood as a micro-scale analogy of responsive systems, where structure, signal, and behavior are tightly intertwined.

Conclusion

This exercise revealed that protein design is not only about predicting biological function, but about constructing relationships between structure, behavior, and interaction.

By moving from sequence generation (PepMLM) to structural validation (AlphaFold), property evaluation (PeptiVerse), and finally guided design (moPPIt), I experienced a workflow that closely mirrors design processes in architecture and material systems — iterative, multi-scalar, and decision-driven.

What becomes particularly interesting is how control increases across these stages: from observing patterns, to evaluating performance, and ultimately to shaping outcomes intentionally. This shift transforms protein design into a form of design practice, where molecular interactions can be approached as spatial and responsive systems.

For my broader research, this opens up a new way of thinking about bio-responsive environments. Just as peptides can be designed to selectively bind and respond to specific protein states, future materials and spaces may be designed to sense and adapt to human physiological signals with similar precision.

In this sense, protein design becomes not only a biological tool, but a conceptual bridge between molecular systems and responsive spatial design.

Part C: Final Project — L-Protein Mutants

Objective

The goal of this assignment is to improve the stability and auto-folding properties of the MS2 bacteriophage lysis protein (L protein).

This protein plays a crucial role in the phage life cycle by inducing bacterial lysis through a mechanism that does not rely on enzymatic degradation of the cell wall, but rather through protein-mediated disruption.

Background

The MS2 L protein is a small membrane-associated protein (~75 amino acids) that functions as a single-gene lysis system.

Previous studies have shown that:

• The C-terminal domain is essential for lytic activity
• The protein forms oligomeric assemblies in the membrane
• Specific motifs such as the LS dipeptide are highly conserved and functionally important

Mutational studies indicate that many loss-of-function mutations cluster in structurally sensitive regions, suggesting that protein folding and stability are tightly linked to function.

Design Strategy

Rather than introducing random mutations, I approached this problem as a guided design task, focusing on improving structural robustness while preserving functional regions.

The following strategies were considered:

1. Stabilizing Secondary Structure

Mutations were selected to:

• promote alpha-helical stability
• reduce structural disorder
• improve folding energetics

2. Reducing Aggregation

To minimize aggregation:

• hydrophobic clustering was reduced
• polar and charged residues were introduced at surface-exposed positions

3. Preserving Functional Domains

Critical regions such as:

• the C-terminal domain
• conserved motifs (e.g., LS motif)

were kept intact to avoid disrupting lytic function.

Proposed Mutations

Based on these principles, the following mutation strategies were proposed:

• Substitution of flexible residues with helix-promoting residues (e.g., Ala, Leu)
• Introduction of charged residues to improve solubility
• Avoidance of mutations in conserved functional motifs

These mutations aim to improve folding efficiency while maintaining membrane interaction capability.

Interpretation

This approach reflects a shift from mutation screening → rational design.

The L protein can be understood as a minimal biological system where:

• structure
• membrane interaction
• and oligomerization

must be finely balanced.

Improving stability without disrupting function requires precise control over local structural features, similar to tuning material behavior in architectural systems.

Future Work

To validate these designs, the following steps would be required:

• structural prediction (AlphaFold / ESMFold)
• expression and folding assays
• membrane insertion studies
• functional lysis assays

• Design Logic Summary

Proposed Design Principle

The main principle I would follow is:

keep the functional core intact, stabilize flexible regions conservatively, and only adjust surface properties where aggregation risk appears higher than functional benefit.

This is important because the MS2 L protein is extremely small, so even a single mutation may have a disproportionately large effect on folding, membrane insertion, oligomerization, or lytic activity.

From a design perspective, this resembles working with a minimal structural system: when the system is very compact, every intervention must be precise and justified.

This iterative loop between design → prediction → validation is essential for advancing protein engineering toward therapeutic applications.

Referances

https://doi.org/10.1099/mic.0.000485

https://www.oaepublish.com/articles/mrr.2023.28

https://en.wikipedia.org/wiki/Bacteriophage_MS2?utm

Toward Rational L-Protein Mutants

Because I do not come from a molecular biology background, I approached this part less as a mutation-screening exercise and more as a design problem.

Instead of trying to guess many highly specific biochemical mutations, I focused on a small set of design principles that could improve the stability and folding behavior of the MS2 L protein while preserving its lytic function.

From the literature, three constraints seem especially important:

1. Do not disrupt the C-terminal functional region
   Mutational studies show that many loss-of-function mutations cluster in the C-terminal domain, suggesting that this region is critical for activity.

2. Preserve conserved functional motifs
   In particular, the LS motif has been reported as functionally important and should not be altered.

3. Respect membrane-associated behavior
   MS2-L is a membrane-associated lysis protein and forms oligomeric assemblies after membrane insertion, so mutations should avoid disrupting the membrane-interacting character of the protein.

Based on this, I would not propose dramatic redesigns. Instead, I would advance small, conservative mutation strategies:

• replace some highly flexible residues in non-critical regions with alanine to slightly stabilize local structure
• reduce aggregation risk by replacing selected exposed hydrophobic residues with more polar residues
• introduce mild surface charge balancing only in regions that are likely solvent-exposed, not in membrane-facing segments

In other words, my design strategy is:

preserve the functional core, stabilize the unstable edges, and avoid over-editing the membrane-active region.

Example mutation logic

Rather than claiming exact validated therapeutic mutants, I would prioritize these mutation types for testing:

• G → A in flexible, non-conserved positions
  to reduce local conformational freedom and improve folding stability

• S/T → A or L in helix-compatible non-critical regions
  to modestly favor secondary-structure stability

• I/V/L → Q/E/K only at predicted exposed positions
  to reduce aggregation tendency and improve solubility without disrupting membrane insertion

Why this matters

For me, the interesting part of this exercise is that protein engineering starts to look similar to architectural or material design:

• some regions behave like load-bearing structure
• some regions behave like interface surfaces
• some regions tolerate modification
• some regions must remain intact for the whole system to function

So the goal is not maximum change, but targeted intervention with minimum disruption.

References

• Chamakura et al. Mutational analysis of the MS2 lysis protein L
  https://pmc.ncbi.nlm.nih.gov/articles/PMC5775895/

• Goessens et al. A synthetic peptide corresponding to the C-terminal 25 residues of phage MS2 coded lysis protein dissipates the proton motive force in E. coli membrane vesicles
  https://europepmc.org/article/pmc/pmc454404

• Mezhyrova et al. In vitro characterization of the phage lysis protein MS2-L
  https://www.oaepublish.com/articles/mrr.2023.28