Group Final Project

Group Brainstorm on Bacteriophage Engineering

Computational Engineering of the MS2 Lysis Protein (L) Background. The MS2 L protein is a 75-amino-acid polypeptide that lyses E. coli by an incompletely understood mechanism. Its C-terminal transmembrane (TM) domain inserts into the cytoplasmic membrane and oligomerizes, causing depolarization that triggers host autolytic enzymes to degrade the murein layer. Recessive, conservative missense mutations clustered around a conserved LS dipeptide strongly implies L engages an unidentified host protein target rather than simply disrupting the bilayer. The dispensable N-terminal domain binds chaperone DnaJ (with solved PDB structures), modulating lysis timing. Its removal causes lysis ~20 min earlier. No experimental structure of L exists. Goals. (1) Stabilize L for more robust membrane accumulation. (2) Accelerate lysis by bypassing DnaJ-dependent regulatory timing and improving delivery of functional L to the membrane. Because the downstream lytic target is unknown, we do not attempt to enhance per-molecule toxicity at the point of target engagement; we focus on removing regulatory brakes and increasing the supply of functional protein. Pipeline: Three Tools, Each Non-Redundant

  1. Clustal Omega (Conservation Map). Align L homologs across Leviviridae (MS2, f2, R17, GA, PP7, AP205, PRR1, M12, KU1, JP34). Conserved C-terminal residues, especially the LS motif, are presumed to mediate the unknown heterotypic interaction and are excluded from mutation. This map constrains all downstream design.
  2. ESM2 + Deep Combinatorial Scanning (Fitness Oracle). Score every single-point mutation by log-likelihood change: increases at mutable positions indicate stabilizing substitutions (Goal 1). N-terminal scanning identifies mutations that disrupt DnaJ binding (Goal 2). A strict preservation rule applies near the LS motif: mutations are evaluated for maintenance of wild-type fitness, not improvement. The genetics show even conservative changes there cause recessive loss of function. Pairwise combinatorial scanning (about ~2M pairs) captures epistatic synergies at mutable positions. This could be potentially pushed further with enough compute.
  3. AlphaFold 3 (Structural Filter + Complex Model). Predicts variant structures as a sanity check (does the TM helix survive?) and models the L–DnaJ complex to verify that N-terminal truncations/mutations disrupt the regulatory interface. Used as a filter, not a design engine. PAE matrix identifies confident interface contacts. Ranking. Composite score: ESM2 log-likelihood gain (stability) + conservation preservation (all essential residues intact) + AF3-predicted DnaJ-binding disruption (for timing bypass). Top 10–20 variants advance to experimental validation.

Why Not More Tools? ProteinMPNN is excluded because it is trained on crystallized globular PDB proteins, not predicted structures of disordered membrane peptides. The compute is invested in combinatorial ESM2 depth. Pitfalls No experimental structure: All structural reasoning rests on AF3 predictions for a challenging target; mitigated by treating AF3 as a filter and cross-referencing against the conservation map. Unknown lytic target: The central limitation. We cannot optimize target-binding affinity for an unidentified partner; engineering is restricted to upstream properties (stability, membrane delivery, DnaJ bypass). Autolysin bottleneck: If lysis rate is limited by host autolytic enzyme activity rather than L accumulation, stabilization gains may show diminishing returns; the plaque assay will reveal this.