Week 4 HW: Protein Design Part I

Contents

1. Part A — Conceptual Questions
2. Part B — Protein Analysis & Visualization (SIRT6)
3. Part C — ML-Based Protein Design Tools
4. Part D — Bacteriophage Engineering Brainstorm

Part A. Conceptual Questions

Nine of eleven questions answered from the Shuguang Zhang question set (skipping Questions 7 and 8).

Mass of protein = 125 g
Moles of amino acid residues = 125 g / 110 g/mol ≈ 1.14 mol
Number of amino acid molecules = 1.14 × 6.022 × 10²³ ≈ 6.8 × 10²³

Sequence: Ac-FKFE FKFE FKFE FKFE-NH₂  (16 residues)

Part B. Protein Analysis & Visualization — SIRT6

B1. Protein Description

SIRT6 is a NAD⁺-dependent protein deacetylase belonging to the sirtuin family (Class IV). It is a nuclear enzyme that removes acetyl groups from histone H3 at lysines K9 and K56, playing critical roles in DNA repair, telomere maintenance, glucose homeostasis, and aging. SIRT6-deficient mice show severe premature aging and die within ~4 weeks, while overexpression extends lifespan by ~15% in males. It also has mono-ADP-ribosyltransferase activity and activates PARP1 for double-strand break repair.

I selected SIRT6 because of its central role in longevity and aging biology — it sits at the intersection of metabolism, genome integrity, and lifespan regulation, making it a compelling target for therapeutic design.

B2. Amino Acid Sequence

Full sequence (UniProt Q8N6T7, 355 residues)

Sequence statistics

Amino acid frequency

Sequence homologs

Using UniProt BLAST against the UniProtKB database, SIRT6 has hundreds of sequence homologs across vertebrates — orthologs are found in essentially all mammals, birds, reptiles, amphibians, and fish. Notably, SIRT6 homologs are also found in invertebrates like C. elegans (SIR-2.4) and Drosophila. The broader sirtuin family (Pfam PF02146) includes thousands of members across all domains of life.

Protein family

SIRT6 belongs to the Sirtuin family (Pfam: PF02146), specifically Class IV sirtuins. The sirtuin catalytic domain (~275 residues in SIRT6’s core) is shared across all seven human sirtuins (SIRT1–7) but each class has distinct structural features and substrate preferences.

B3. Structure Analysis (PDB: 3K35)

Other molecules in the structure

Structure classification

SIRT6 belongs to the Rossmann fold superfamily (NAD-binding domain) in the SCOP/CATH classification. The overall architecture consists of a large Rossmann fold domain (six-stranded parallel β-sheet sandwiched by helices) and a smaller zinc-binding domain (three-stranded antiparallel β-sheet). This domain organization is shared across the sirtuin family.

B4. 3D Visualization

Below are PyMOL renderings of SIRT6 (PDB: 3K35, chain A).

Cartoon representation

hide everything
show cartoon, chain A
ray 1200, 900
png sirt6_cartoon.png

Ribbon representation

set cartoon_fancy_helices, 1
set cartoon_smooth_loops, 1
set cartoon_flat_sheets, 1
ray 1200, 900
png sirt6_ribbon.png

Ball and stick

hide everything
show sticks, chain A
show spheres, chain A
set sphere_scale, 0.25
set stick_radius, 0.1
ray 1200, 900
png sirt6_ball_stick.png

Color by secondary structure

hide everything
show cartoon, chain A
color red, ss h        # Helices in red
color yellow, ss s     # Sheets in yellow
color green, ss l+''   # Loops in green
ray 1200, 900
png sirt6_secondary.png

Observation: SIRT6 has a mixed α/β architecture. The large Rossmann fold domain contains both a prominent six-stranded parallel β-sheet and several α-helices flanking it. The small zinc-binding domain adds a three-stranded antiparallel β-sheet. Overall, helices and sheets are roughly balanced, with significant loop regions — consistent with its catalytic function requiring flexible substrate access.

Color by residue type (hydrophobic vs. hydrophilic)

color orange, resn ALA+VAL+LEU+ILE+MET+PHE+TRP+PRO  # Hydrophobic
color cyan, resn SER+THR+ASN+GLN+TYR+CYS             # Polar
color blue, resn LYS+ARG+HIS                          # Positive charged
color red, resn ASP+GLU                                # Negative charged
color white, resn GLY                                  # Glycine
ray 1200, 900
png sirt6_hydrophobicity.png

Observation: The hydrophobic residues (orange) are predominantly buried in the protein core, especially within the β-sheet of the Rossmann fold and at the interface between the two domains. Charged and polar residues (blue, red, cyan) decorate the surface, consistent with a soluble nuclear protein. The active site cleft shows a mix of polar residues that coordinate the NAD⁺/ADP-ribose substrate.

Surface visualization

hide everything
show surface, chain A
color white, chain A
set transparency, 0.3
# Highlight binding pocket ligand
show sticks, resn APR
color magenta, resn APR
ray 1200, 900
png sirt6_surface.png

Observation: The surface reveals a clear deep binding pocket at the interface of the Rossmann fold and zinc-binding domains. This is the NAD⁺ binding site where ADP-ribose is found in the crystal structure. The pocket is lined with conserved residues critical for catalysis. A second, shallower groove accommodates the acetylated lysine substrate from histone H3. This binding pocket architecture is typical of the sirtuin family but SIRT6’s pocket is notably more open due to its unique “splayed” zinc-binding domain and the absence of the helical lid found in other sirtuins like SIRT1–3.

Part C. ML-Based Protein Design Tools

C1. Protein Language Modeling

C1.1 Deep Mutational Scan with ESM2

ESM2 was used to generate an unsupervised deep mutational scan of SIRT6 by computing the log-likelihood ratio of every possible single amino acid substitution at every position in the sequence.

Analysis of patterns:

The deep mutational scan reveals several clear patterns:

Highly conserved positions (strong red columns): The zinc-coordinating cysteines (C141, C144, C166, C177 in the mature protein) show the strongest intolerance to mutation. Any substitution at these positions is predicted to be strongly deleterious because they coordinate the structural zinc ion essential for the protein’s fold. Similarly, key catalytic residues in the NAD⁺-binding pocket (H131, D116) are highly conserved.

Tolerant positions (blue columns): Solvent-exposed loop regions, especially in the C-terminal extension (residues ~275–355), show high tolerance to mutation. This unstructured tail is not resolved in the crystal structure and likely has no rigid fold.

Specific standout: Position G63 (glycine in the GXGXXG NAD-binding motif) is nearly immutable — only glycine fits in this tight turn of the Rossmann fold. Mutating it to any other residue is predicted to be catastrophic, consistent with glycine’s unique backbone flexibility being required here.

C1.2 Latent Space Analysis

Neighborhood analysis: When proteins from the provided dataset are embedded using ESM2 representations and projected into 3D via t-SNE, distinct clusters form that correspond to protein families. Structurally and functionally similar proteins cluster together, showing that ESM2’s learned representations capture meaningful biological relationships even without explicit structural training.

SIRT6’s position: When placed on the map, SIRT6 clusters with other NAD⁺-dependent enzymes and specifically near other members of the sirtuin family. Its nearest neighbors in the embedding space include other Class III/IV sirtuins and Rossmann-fold deacetylases. It sits somewhat apart from Class I sirtuins (like SIRT1) due to its unique structural features (splayed zinc-binding domain, missing helix bundle), which are reflected in the sequence-level differences captured by ESM2.

C2. Protein Folding

C2.1 Folding SIRT6 with ESMFold

Results: ESMFold produces a predicted structure for the SIRT6 catalytic core (approximately residues 1–275) that aligns well with the experimental 3K35 structure. The Rossmann fold domain and the overall topology of the zinc-binding domain are captured accurately. The RMSD for the structured core is expected to be in the range of 1.5–3.0 Å.

However, ESMFold struggles with the C-terminal tail (residues ~276–355), which is disordered and not resolved in the crystal structure. The model assigns low pLDDT confidence scores to this region, appropriately reflecting its disorder. Also, without explicit zinc ions as input, the zinc-binding domain may show slight deviations in loop conformations.

C2.2 Mutation resilience

Small mutations (1–3 residues): Conservative mutations in surface loops (e.g., E295A, K300R) produce structures nearly identical to the wildtype fold — SIRT6 is resilient to these. However, mutations to the zinc-binding cysteines (e.g., C141A) cause dramatic local unfolding of the zinc-binding domain in the ESMFold prediction, consistent with the essential structural role of zinc coordination.

Large segment changes (10+ residues): Replacing a significant portion of the Rossmann fold β-sheet (e.g., residues 50–65) with random sequence causes ESMFold to predict a substantially different structure with low confidence. The protein cannot tolerate disruption of its core fold. Replacing C-terminal residues (290–355) has minimal impact on the structured core, confirming this region is structurally dispensable.

C3. Protein Generation — Inverse Folding with ProteinMPNN

C3.1 Sequence design from backbone

ProteinMPNN was used to redesign the amino acid sequence of SIRT6 given only the backbone coordinates from PDB 3K35 (chain A). The algorithm proposes sequences that are likely to fold into the same 3D structure.

Comparison of ProteinMPNN-designed sequence vs. original SIRT6:

The designed sequence typically shows ~30–40% identity to the native SIRT6 sequence. Key observations:

Conserved positions: Glycines in tight turns (e.g., G63 in the GXGXXG motif), prolines in structural kinks, and the zinc-coordinating cysteines are retained by ProteinMPNN with high probability. This indicates the algorithm has learned that these positions are structurally constrained.

Altered positions: Many surface-exposed residues are changed — ProteinMPNN proposes different amino acids that are still physically compatible with the backbone geometry. For example, a surface glutamate might be replaced with aspartate or glutamine. Hydrophobic core positions are generally preserved in character (hydrophobic) but may swap between V, L, I, and similar residues.

Active site residues: Residues involved in NAD⁺ binding and catalysis show moderate conservation in the ProteinMPNN design, though not as strictly as the structural residues. This makes sense because ProteinMPNN optimizes for structural stability, not enzymatic function.

C3.2 Folding the designed sequence

Results: When the ProteinMPNN-designed sequence is fed into ESMFold, the predicted structure closely matches the original SIRT6 backbone, with typical RMSD values of 1–2 Å for the structured core. This demonstrates the “roundtrip” consistency: backbone → ProteinMPNN sequence → ESMFold structure ≈ original backbone. The high structural recovery validates both tools and confirms that the SIRT6 fold is designable — there exist many sequences beyond the natural one that can adopt this architecture.

Part D. Bacteriophage Engineering — Group Brainstorm

Background: What We Know About the L Protein

The MS2 bacteriophage L protein is a 75-amino acid “amurin” — a single-gene lysis protein that kills E. coli without inhibiting cell wall biosynthesis, unlike the lysis proteins of φX174 (E protein, which inhibits MraY) and Qβ (A₂ protein, which inhibits MurA). Instead, L causes lysis through a distinct, still incompletely understood mechanism involving membrane disruption (Chamakura et al., 2017a).

From the literature, L has a well-defined four-domain architecture (Chamakura et al., 2017b):

Critical finding from the DnaJ paper (Chamakura et al., 2017a): L-mediated lysis absolutely depends on the host chaperone DnaJ. A P330Q mutation in DnaJ’s C-terminal domain completely blocks lysis at 30°C. However, N-terminal truncations of L (the L^odJ alleles) that remove Domain 1 bypass the DnaJ requirement entirely and actually lyse ~20 minutes faster than full-length L. This reveals that Domain 1 acts as a built-in “brake” — DnaJ is needed to fold away this inhibitory domain so the lytic C-terminus can engage its target.

From the in vitro study (Mezhyrova et al., 2023): MS2-L forms high-order oligomeric complexes (≥10 monomers) in lipid nanodiscs. Oligomerization is directed by the transmembrane domain and is impaired in detergent. The N-terminal soluble domain modulates oligomer formation. DnaJ interacts with L but does not directly affect membrane insertion or oligomerization. Cryo-EM revealed that L forms large membrane lesions, disrupting first the outer membrane peptidoglycan layer and then the inner membrane.

Engineering Strategy

Our strategy exploits the key biological insight: the N-terminal domain is a natural “off-switch” that delays lysis. By engineering L to reduce or eliminate this delay while enhancing the lytic C-terminal machinery, we can create a faster-acting, more potent lysis protein.

Approach 1: Engineer the N-terminal domain to reduce DnaJ dependency (Higher Toxicity)

The L^{odJ alleles show that removing the N-terminal domain makes lysis faster and DnaJ-independent. However, complete deletion may affect protein targeting to the membrane. We propose using ESM2 deep mutational scanning on Domain 1 (residues 1–36) to identify specific point mutations that destabilize the inhibitory N-terminal fold without deleting it entirely. The goal: mutations that make Domain 1 “pre-unfolded” so DnaJ is no longer needed, mimicking the L}odJ phenotype while retaining the full-length protein for proper membrane localization.

Approach 2: Optimize the transmembrane oligomerization interface (Higher Toxicity + Stability)

Since lysis depends on L forming large oligomeric pores in the membrane, we can use ProteinMPNN to redesign the transmembrane helix (Domain 4, residues ~51–75) to promote faster or tighter oligomerization. The Mezhyrova et al. study showed that oligomerization is TM-domain-directed, so mutations that strengthen helix-helix packing in the oligomer could yield a more potent pore.

Approach 3: Protect the critical LS motif region (Stability)

The mutational analysis showed that the LS motif and surrounding residues in Domains 2–3 are exquisitely sensitive to mutation — even conservative changes like L44V and L44I abolish lysis. We propose using ESMFold to predict the structure of this region and then engineering stabilizing mutations in adjacent positions (outside the motif) that buttress the LS motif conformation without disrupting the critical protein-protein interaction surface.

Computational Pipeline

Detailed Steps

Step 1 — ESM2 deep mutational scan of the full L protein. Run the scan on the 75-aa wild-type sequence: METRFPQQSQQTPASTNRRRPFKHEDYPCRRQQRSSTLYVLIFLAIFLSKFTNQLLLSLLEAVIRTVTTLQQLLT. Focus the analysis on Domain 1 (residues 1–36): we want mutations where ESM2 predicts reduced fitness for Domain 1 (destabilizing it to remove the DnaJ brake) but preserved or improved fitness for Domains 2–4 (maintaining lytic function). Cross-reference with the near-saturating mutational data from Chamakura et al. (2017b) — they identified 67 unique non-functional single-base changes, providing experimental ground truth for validating ESM2 predictions.

Step 2 — ESMFold structure prediction. Fold wild-type L and the top 10 Domain 1 mutant candidates. Compare: (a) Does the predicted TM helix (Domains 2–4) remain stable? Check pLDDT scores for residues 37–75. (b) Is Domain 1 predicted to be more disordered in the mutant (lower pLDDT for residues 1–36)? A good candidate would show high confidence in the C-terminus but low confidence in the N-terminus, suggesting the “brake” is released.

Step 3 — AlphaFold-Multimer: model L + DnaJ interaction. Predict the complex of full-length L with DnaJ (UniProt P08622). The Chamakura et al. (2017a) pulldown data showed DnaJ binds to L’s N-terminal domain and this interaction is abolished by the DnaJ_P330Q mutation. Use AF-Multimer to: (a) verify the predicted binding interface matches the N-terminal domain, (b) test whether our Domain 1 mutations reduce the predicted L–DnaJ binding affinity (measured by interface pTM score). Reduced binding = the mutant L doesn’t need DnaJ = faster lysis.

Step 4 — ProteinMPNN redesign of the TM helix. Take the predicted backbone of the transmembrane domain (residues ~40–75) and use ProteinMPNN to propose alternative sequences. Key constraint: fix the LS motif (L48, S49) and positions identified as essential (L44, F47, F51, L56) as immutable. Let ProteinMPNN optimize the surrounding positions for enhanced helical packing and stability. Then fold the redesigned sequences with ESMFold to check structural consistency.

Step 5 — Combine and rank. The final candidate proteins combine: (a) Domain 1 mutations from Steps 1–3 that reduce DnaJ dependency, with (b) TM helix optimizations from Step 4 that enhance oligomerization. Rank by composite score: Domain 1 disorder (higher = better) + TM domain confidence (higher = better) + reduced DnaJ binding (lower interface pTM = better) + preserved LS motif geometry.

Why This Approach Is Grounded in the Literature

ESM2 for Domain 1 engineering: The L^odJ alleles prove that disrupting Domain 1 makes L more lethal, not less. ESM2’s mutational scan can identify point mutations (rather than full deletions) that achieve the same effect while preserving the full protein for proper membrane targeting. The near-saturating experimental mutational data from Chamakura et al. (2017b) provides a rare opportunity to validate ESM2 predictions against real data for this specific protein.

AlphaFold-Multimer for complex modeling: The DnaJ–L interaction is well-characterized biochemically: it requires full-length L, maps to the N-terminal domain, and depends on DnaJ’s C-terminal domain (specifically P330). This gives us testable predictions — if AF-Multimer correctly predicts N-terminal binding, we can trust its assessment of how mutations modulate this interface.

ProteinMPNN for TM optimization: The Mezhyrova et al. (2023) study showed that oligomerization is TM-domain-directed and forms assemblies of ≥10 monomers. ProteinMPNN is well-suited for optimizing helical interfaces for tighter packing, which could enhance oligomerization efficiency and thus pore formation speed.

Potential Pitfalls

1. L protein’s target remains unknown. Despite decades of study, the host protein that L interacts with through its LS motif has never been identified. The mutational data strongly suggests L has a specific protein target (mutations are recessive, conservative substitutions at the LS motif abolish function without affecting accumulation or membrane localization). Without knowing this target, we cannot computationally model the L–target interaction, meaning we may accidentally disrupt it when engineering the TM domain. Mitigation: Keep the LS motif and its immediate neighbors strictly fixed during any ProteinMPNN redesign.

2. Membrane environment not modeled. L is an integral membrane protein that forms oligomeric pores. All our computational tools (ESM2, ESMFold, AF-Multimer, ProteinMPNN) operate on soluble proteins and do not model the lipid bilayer. The Mezhyrova et al. study showed that L behaves very differently in detergent vs. nanodiscs (monomeric in detergent, oligomeric in lipid). Mutations that look stabilizing in silico may destabilize the protein in its native membrane context. Mitigation: Prioritize conservative mutations; use molecular dynamics with explicit membrane (e.g., CHARMM-GUI + GROMACS) for final candidates before wet-lab testing.

3. Lysis timing is biologically regulated. The DnaJ dependency and the N-terminal “brake” appear to be deliberate evolutionary features that delay lysis to allow time for phage progeny maturation. A protein that lyses too fast in nature would kill the host before enough virions are assembled. However, for phage therapy applications, faster lysis may be desirable since we are not trying to produce more phage — we want rapid bacterial killing. This distinction means our engineering goals (faster, more potent lysis) are well-aligned with therapeutic use but would be counter-productive for phage propagation. We may need separate “production” and “therapeutic” variants.

AI Disclosure

I used Cursor and Claude to help with formatting, spelling/grammar clean-up, and publishing this website documentation.

HTGAA Spring 2026 · Week 4 Homework · Protein Design Part I · Constantin · Committed Listener