Week 5 Review: Protein Design Part II

Week 5 — Protein Design II

AI-driven peptide and protein engineering, worked end-to-end on two targets.

TL;DR

• Tool stack for peptide design: PepMLM (generate) → AlphaFold3 (validate) → PeptiVerse (triage) → moPPIt (re-target). Each tool catches a failure the others miss.
• Target 1: SOD1-A4V (ALS). PepMLM alone produces mode-collapsed peptides that all dock at the wrong AF3 default surface. moPPIt with motif guidance produces target-aware chemistry. Advance: B3 PAEKWFVFWHPT (sub-µM predicted Kd, dimer-interface targeted).
• Target 2: MS2 L-protein. ESM-style saturation scan vs random vs experiment-led picks. Big finding: language-model preference and experimental lysis function have r = +0.007 correlation. The model’s top picks would have destroyed function.
• Meta-lesson: Unsupervised protein language models predict sequence plausibility, not function. On under-represented protein families they can be actively misleading.

Course: HTGAA Spring 2026 · Lecture (Mar 3): Gabriele Corso, Pranam Chatterjee — Protein Design Part II · Author: Fiona C (Committed Listener BioPunk Node)

The tool stack at a glance

flowchart LR
    T[Target sequence] --> P[PepMLM<br/>generate plausible binders<br/>perplexity score]
    P --> A[AlphaFold3<br/>co-fold target+peptide<br/>ipTM, PAE, pLDDT, pose]
    A --> V[PeptiVerse<br/>developability triage<br/>solubility, hemolysis, Kd]
    V --> D{Site OK?<br/>Developable?}
    D -- no, redirect --> M[moPPIt<br/>site-targeted re-generation<br/>multi-objective guided]
    M --> A
    D -- yes --> L[Lead candidate<br/>for wet-lab]
    style P fill:#e0f2fe
    style A fill:#fef3c7
    style V fill:#cfe9d4
    style M fill:#e9d5ff

Concept sidebars

Peptide vs small molecule — they’re different modalities, not just different sizes

A peptide can sit in the same MW range as a small-molecule drug, but in drug discovery the two are separate modality classes.

Rule of thumb: target topology dictates modality. Deep pocket → small molecule. Flat surface → peptide. Both modalities appear in Week 5 for exactly this reason.

Perplexity in one worked example

Perplexity = exponentiated mean negative log-likelihood. Read as: the effective number of equally-likely choices the model is hedging between at each position.

For a toy 5-mer PEPTI with per-position probabilities P=0.20, E=0.15, P=0.25, T=0.10, I=0.05:

So the model is hedging across the equivalent of ~7.7 amino acids per position. Versus the random baseline of 20, that’s real information; versus a strong binder (PPL 2–3), it’s mediocre.

Watch: ESM-2 / PepMLM use base-e. Some NLP literature uses base-2 (“bits per character”). Don’t mix.

Masked LMs like PepMLM report pseudo-perplexity — mask each position one at a time, predict the actual residue from the rest. Interpretation is the same.

AF3 metrics for protein-peptide complexes

Critical short-peptide caveat. Standard ipTM cutoffs were calibrated against protein-protein complexes. For a 12-mer vs a 154-aa target, ipTM is systematically biased downward (Stein & Dunbrack, bioRxiv 2025 — the ipSAE analysis). A 12-mer with ipTM 0.5 is not auto-junk. Use the literature benchmark’s ipTM as the calibration anchor, not a universal threshold.

moPPIt vs PepMLM

PepMLM samples binders plausible against the target as a whole. It can’t be told where to bind. moPPIt’s Multi-Objective Guided Discrete Flow Matching (MOG-DFM) adds:

1. Motif guidance — specify which target residues to engage.
2. Multi-objective optimization — affinity, solubility, motif specificity simultaneously, during sampling (not as a post-hoc filter).
3. Pareto-front output — multiple candidates at different trade-off points, not a single “best.”

Worked example 1 — SOD1-A4V peptide therapeutics

Why this target

flowchart TD
    WT[Native SOD1 homodimer<br/>Cu/Zn bound, Tm > 90°C] --> M[A4V monomer<br/>destabilized N-terminus]
    M --> R[Disulfide-reduced<br/>demetalated apo monomer]
    R --> O[Misfolded oligomers / trimers]
    O --> F[Insoluble fibrils]
    F --> D[Motor neuron death]
    style WT fill:#cfe9d4
    style F fill:#fecaca
    style D fill:#fecaca

A useful binder must engage the A4V site itself or the dimer interface. Anything else is therapeutically irrelevant, no matter how good the prediction looks.

Stage 1 — PepMLM generation

Default sampling on SOD1-A4V produced four 12-mer peptides plus the literature benchmark.

Three failure modes immediately:

Pitfall 1 — Mode collapse. All four peptides share residues at positions 1 (W), 3 (Y), 7 (A), and 11 (K, in three of four). Peptides 3 and 4 differ at only 3 of 12 positions. ESM-2’s training distribution under-represents this target; the model defaults to a generic aromatic-cationic anchor pattern.

Pitfall 2 — X ambiguity codes. X is the IUPAC “any/unknown amino acid” — not a real residue. ESM-2’s tokenizer carries X; when probability mass is spread, X wins the argmax. Substituted to G (matching benchmark’s GG terminus) for downstream use, with the substitution flagged transparently.

Pitfall 3 — Higher-than-textbook PPLs. Textbook “PPL ~ 2–5 = good binder” doesn’t apply here. The whole distribution sits at 10–20 for this target. Read PPL relative to benchmark, not against universal thresholds.

Stage 2 — AlphaFold3 validation

Five jobs on alphafoldserver.com (chain A = SOD1-A4V, chain B = peptide). Top model per job:

flowchart LR
    All[All 5 peptides<br/>including benchmark] --> Site[Top contacts:<br/>residues 138, 142, 143, 144<br/>= electrostatic loop, back face]
    All --> A4V[A4V site, residue 5<br/>max contact prob 0.00 - 0.01<br/>min PAE 11.7 - 14.8 Å]
    style Site fill:#fef3c7
    style A4V fill:#fecaca

Pitfall 4 — AF3 default-site convergence. When AF3 can’t find a strong anchor, it deposits poorly-anchored peptides on whichever surface is geometrically convenient. For SOD1-A4V that’s the C-terminal electrostatic loop. Five different peptides converging there — including the literature benchmark — means AF3 doesn’t have a confident pose for any of them and is showing us its default surface.

Stage 3 — PeptiVerse developability

All five passed developability (Soluble at 1.000 prob, hemolysis < 0.05). But the binding-affinity ranking inverts AF3:

Pitfall 5 — Cross-model disagreement. PepMLM says P2 best. AF3 says P1 best. PeptiVerse says P3 best. The “best” peptide depends on which model you trust most. Spearman correlation between AF3 ipTM and PeptiVerse pKd across the four generated peptides is strongly negative.

Stage 4 — moPPIt site-targeted re-generation

Two parallel runs with explicit motif guidance:

Run A — A4V cluster (5 samples):

Run B — Dimer interface (3 samples):

Key finding — target-aware chemistry. Run A (mixed basic/hydrophobic target) is compositionally diverse: charges span −2 to +3, frequent Cys, low aromatic content. Run B (flat hydrophobic interface, ~80% nonpolar in literature) is aromatic-rich, electrostatically neutral, with one candidate (B1) showing flanking Cys consistent with a macrocyclic design (though this is a hypothesis, not validated — see caveat below). moPPIt’s guidance produces chemistry appropriate to target-surface biology in a way PepMLM’s unconditional sampling cannot.

Caveat on B1. Flanking Cys at positions 1 and 12 could form an intramolecular disulfide and fold into a macrocycle. But it could also be coincidental (n=3 samples), a classifier-learned pattern without intentional macrocyclization, or a sampling accident. Validating the macrocyclic form requires Boltz-2 or RoseTTAFold All-Atom (not AF3 — the AlphaFold Server doesn’t support intramolecular peptide disulfides).

Caveat on B2. Predicted pKd 7.72 is the workspace headline number, but 7 of 12 residues are aromatic. PeptiVerse’s solubility classifier (1.000) is almost certainly out-of-distribution for this composition. In wet-lab practice, peptides this hydrophobic don’t dissolve, self-aggregate, and bind non-specifically. B2 is a teaching example of why predicted Kd is not the only metric.

Advance recommendations

flowchart LR
    Workspace[12 candidate peptides] --> P[Primary: B3<br/>PAEKWFVFWHPT<br/>balanced, sub-µM, dimer-interface]
    Workspace --> Alt[Alternate: B1<br/>CTAVLNVGLEWC<br/>macrocycle hypothesis]
    Workspace --> A4V[A4V site: A5<br/>QGKCKFKQFNPV<br/>best Run A profile]
    style P fill:#cfe9d4
    style Alt fill:#fef3c7
    style A4V fill:#e0f2fe

Worked example 2 — MS2 L-protein engineering

Why this target

Three pick strategies, head-to-head

We ran three strategies for picking 5 mutations each, then cross-scored against the Chamakura experimental lysis dataset (n=59 unique mutations with measured lysis 0/1).

flowchart TD
    Q[Pick mutations for engineering] --> R[1. Random mutagenesis<br/>option3 python script<br/>seed=42, 2-4 substitutions]
    Q --> L[2. LLR-informed only<br/>ESM saturation scan<br/>top 1% by LLR]
    Q --> E[3. Experiment-led<br/>filter for lysis=1<br/>rank by LLR within set]
    R --> RR[Mean LLR: -1.04<br/>One variant breaks initiator Met<br/>Lysis outcome: unknown for these]
    L --> LR[Mean LLR: +2.15<br/>All in top 1% of landscape<br/>~0 percent likely lysis preservation]
    E --> ER[Mean LLR: +0.18<br/>Lysis preserved: 100 percent<br/>by construction]
    style RR fill:#fef3c7
    style LR fill:#fecaca
    style ER fill:#cfe9d4

The big finding

ESM-2 LLR and experimental lysis function have essentially zero correlation for L-protein:

Statistical-power note: at n=59, the 95% CI on r is approximately ±0.26 (Fisher z). The data are consistent with anywhere from a small inverse correlation to a small positive correlation. The qualitative point — LLR is not informative enough for confident-pick selection on L-protein — holds regardless.

Quartile breakdown (59 unique mutations sorted by LLR):

The top-LLR quartile has the worst lysis preservation rate. Functionally essential residues are positions where the WT looks “unusual” to the model precisely because the unusual residue is doing necessary work. The model says “change it”; the experiment says “don’t.”

What our top LLR picks would have done

Cross-checking each LLR-informed pick against the experimental neighborhood:

The model’s highest-confidence picks land on residues whose unusual identity is doing functional work.

The meta-lesson

Unsupervised protein language models predict sequence plausibility, not biological function. For under-represented protein families (phage proteins, membrane proteins, anything outside UniRef50’s bulk training distribution), they can be actively misleading — and the misleading is worst at the top of their confidence distribution. Use them as one signal among many, weight experimental data heavily when available, and never trust the top picks blindly on an unfamiliar target.

Why ESM-2 fails on L-protein specifically

• Phage protein — under-represented in UniRef50
• Membrane-active — PLMs are notoriously weak on membrane proteins
• Many WT residues are “unusual” by general-protein statistics precisely because they’re doing functional work (single free Cys at 29, Lys mid-TM at 50, Arg-rich N-terminus)
• The measured function (E. coli lysis via DnaJ chaperone + membrane insertion) requires correct folding and downstream context the model can’t see

Pitfalls cheat sheet

Recommended reading

Course resources