Sequence-based interaction prediction asks a simple question: can a model that reads only the linear amino-acid strings of two proteins decide whether, and where, they bind? The answer increasingly looks like yes — masked language models now generate peptide binders conditioned only on a target's sequence, and paired sequence models predict inter-protein residue contacts better than some structure-based systems. Meanwhile, Alzheimer's research has moved past the amyloid-only view of the disease: two of the most experimentally validated non-amyloid drivers are the alarmin protein S100A9, which nucleates amyloid-beta fibrils while amplifying neuroinflammation, and the Keap1–Nrf2 axis, whose collapse switches microglia into a neurotoxic state. Both are protein–protein interfaces — broad, flat, and poorly suited to small molecules — and both are, in principle, addressable by computationally designed peptides.
The obstacle is not the biology but the machine learning. When sequence-based interaction predictors have been stress-tested rather than benchmarked on random splits, they collapse: deep models trained to near-perfect accuracy drop to coin-flip performance once test peptides are excluded from training, because they memorize single-chain sequence frequencies instead of learning recognition; and when screened against a full-proteome decoy background, deep predictors assigned near-zero probabilities to real, FDA-approved peptide–target pairs. Nobody has checked whether these failures also doom the tools we would want to use against Alzheimer's targets. This proposal builds the missing evaluation regime, then builds the model that passes it: a paired, contact-grounded peptide design and screening system, validated on S100A9 and the Keap1–Nrf2 interface, with wet-lab readouts for binding and for disruption of the two pathogenic processes themselves.
Core Claim
The claim is twofold. First, off-the-shelf sequence-based peptide design tools will fail on Alzheimer's inflammatory targets in exactly the ways the generalization literature predicts — uncalibrated scores, no discrimination against proteome-wide decoys, silent memorization of training interactors — and this failure is measurable before any wet-lab work. Second, a model that scores candidate binders through a predicted inter-protein residue contact map, computed jointly from both sequences, restores generalization behind the "memorization wall," because a contact map is a physical claim about the interface that cannot be satisfied by recalling one partner's background statistics. The application payoff is a ranked list of designed peptides that block S100A9 self-assembly and Aβ cross-seeding, or that competitively disrupt Keap1–Nrf2 to restore microglial antioxidant defenses.
The Keap1–Nrf2 case is the cleanest mechanistic test. You Jung Kang and colleagues showed in 3D human Alzheimer's mini-brains that interferon-gamma drives Keap1 upregulation, Nrf2 loss, microglial reactive oxygen species buildup, and tau hyperphosphorylation, and that restoring Nrf2 reverses all of it (Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease). Genetic overexpression is not a drug; a competitive peptide that occupies Keap1's Nrf2-binding groove would be a drug-like stand-in, and its interface is a short, well-defined linear motif — precisely the regime where sequence-conditioned generation should shine.
Research Program
The work has three parts: a benchmark, a model, and a target application.
The benchmark, which we call AD-Pep-Hard, assembles positive protein–peptide and protein–protein pairs around four Alzheimer's-relevant targets: S100A9 (self-association and Aβ cross-seeding, per the EDB study showing S100A9 phosphorylation seeds Aβ fibrils, Edaravone-Dexborneol slows down pathological progression and cognitive decline via inhibiting S100A9 in APPswe/PS1dE9 mice), the Keap1–Nrf2 interface, the tau microtubule-binding region, and Aβ42 with its familial-disease point mutants. Negatives come from two sources: shuffled-interface controls and a decoy screen over the human surfaceome of 2,886 proteins, the protocol already established for therapeutic peptide evaluation (Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering). We define hard splits in the spirit of the TCR work that exposed the memorization wall (On TCR binding predictors failing to generalize to unseen peptides): test peptides, or test targets, must be absent from training, and the headline metric is the gap between random-split and hard-split performance plus top-1% target retrieval against decoys. The random-versus-hard gap is the single number that separates genuine recognition from database lookup.
The model is a generator–verifier pair. The generator follows the span-masking recipe of PepMLM, which concatenates a target sequence with a fully masked peptide segment and fine-tunes a protein language model to reconstruct the binder conditioned on the target (Target sequence-conditioned design of peptide binders using masked language modeling) — we reuse that technique rather than reinvent it. The new component is the verifier. Instead of scoring a candidate pair with pooled single-chain embeddings, the verifier processes both sequences jointly with cross-attention, in the manner of the paired sequence language model PPLM (A paired sequence language model for protein-protein interaction modeling), and produces an inter-protein contact map before any binding decision, in the manner of D-SCRIPT's structure-aware intermediate (Sequence-based prediction of protein-protein interactions: a structure-aware interpretable deep learning model). Its scoring rule is deliberately simple:
$$p(\text{bind} \mid a, b) \;=\; \sigma\!\left( w^\top \mathrm{pool}\big(\hat{C}(a,b)\big) + h(a,b) \right)$$
where $a$ and $b$ are the two sequences, $\hat{C}(a,b)$ is the predicted residue–residue contact map, $\mathrm{pool}$ is a sparsity-weighted aggregation of its top entries, $h$ is a small paired-embedding term, and $\sigma$ is a logistic link. The design forces the score to travel through the contact map: if the model cannot say which residues touch, it cannot say the pair binds. Generated peptides are then filtered by the verifier under proteome-wide decoys, so that the pipeline outputs both a sequence and an off-target profile — the piece PepMLM lacks, since it samples binders but never scores affinity or specificity.
Two Alzheimer's-specific extensions address known weaknesses. Phosphorylation drives S100A9 pathology, and sequence models are notoriously insensitive to single modifications; we will represent phospho-serine and phospho-threonine as augmented tokens and test whether the verifier's contact predictions shift at the modified sites, using the mutation-sensitivity findings of PALM — which struggled to resolve single substitutions in Aβ42 without large-scale retraining — as the diagnostic baseline (Predicting Peptide Aggregation with Protein Language Model Embeddings). And to place the targets in disease context, we will cross-reference candidates against the Alzheimer's Knowledge Base, whose graph correctly flags AD-associated genes with 96.2% balanced accuracy but contains no sequence or binding physics at all (The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research) — the graph prioritizes which targets matter, the sequence model decides what binds them, and neither capability exists today in one place.
Proposed Experiments
Experiment 1 is the stress test that motivates everything else. We run PepMLM, SaLT&PepPr, D-SCRIPT, PIPR, and SPRINT, unchanged, on the AD target panel with the surfaceome decoy screen. The prior from the literature is explicit: on FDA-approved peptides against the same 2,886-protein background, the k-mer tool SPRINT placed true targets in the top 1% for 16 of 21 pairs, while D-SCRIPT ranked none there and deep models emitted near-zero probabilities (Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering). If PepMLM-generated S100A9 binders and SaLT interface predictions survive the decoy screen unaided, the field's problem on these targets is smaller than we claim, and we will say so; the proposal then reduces to the wet-lab application, which is still unexecuted.
Experiment 2 trains the contact-grounded verifier on AD-Pep-Hard plus the broad PPI corpora these models already use, and reports the random-split versus hard-split AUROC gap, the top-1% decoy retrieval rate, and calibration (reliability of the predicted binding probability). The bar is concrete: on hard splits the TCR failure mode was a fall from 0.95 AUROC to roughly 0.55 (On TCR binding predictors failing to generalize to unseen peptides), and a model worth building should hold most of its random-split performance when the peptide is unseen. For generated binders, in silico docking quality is scored by AlphaFold-Multimer interface-predicted TM, the same evaluator under which PepMLM reached a 38% hit rate against RFdiffusion's 29% (Target sequence-conditioned design of peptide binders using masked language modeling); beating 38% on our Alzheimer's target panel is the generative target.
Experiment 3 is the wet-lab decisive arm. Top-ranked S100A9 binders go to surface plasmon resonance, where the reference affinity is the small-molecule inhibitor EDB's dissociation constant in the $10^{-8}$ to $10^{-9}$ M range (Edaravone-Dexborneol slows down pathological progression and cognitive decline via inhibiting S100A9 in APPswe/PS1dE9 mice); sub-micromolar peptide affinity would be competitive, especially since PepMLM-derived peptides reached 60 nM on other targets. Functional assays follow: a thioflavin-T cross-seeding assay testing whether designed peptides slow S100A9-accelerated Aβ fibrillization, and, for Keap1-disrupting peptides, the 3D human AD mini-brain microglial platform of Kang et al., measuring Nrf2 activity, reactive oxygen species, the neurotoxic markers CD86 and iNOS, and phospho-tau — the same readouts that genetically validated Nrf2 restoration (Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease). Specificity is checked by predicting, then experimentally sampling, off-target binding across the proteome, following the mass-spectrometry precedent in which peptide-guided degraders showed no detectable off-target depletion among thousands of quantified proteins (SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders).
Baselines
The naive reference point is random ranking: with 2,886 decoy proteins, a scorer with no signal places the true target in the top 1% for about 1 in 29 pairs, so any method must beat roughly 3–4% top-1% retrieval before it is doing anything at all. The strongest simple competitor is SPRINT, a k-mer similarity tool that needs no GPU and no training on our data; it is strongest precisely because it already won the only published head-to-head on this exact decoy protocol, ranking true targets in the top 1% for 16 of 21 FDA-approved peptide–target pairs while every deep predictor failed (Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering). We will run SPRINT on identical targets, decoys, and ranking metrics; it wins the optimistic setting but, by its authors' own result, fails completely in the de novo setting, which is the setting Alzheimer's targets actually occupy.
The strongest existing end-to-end method is PepMLM itself, run off-the-shelf against S100A9, Keap1, and tau, with its generated peptides scored by AlphaFold-Multimer; its published in silico hit rate of 38% (against RFdiffusion's 29%) and wet-lab 60 nM affinities make it the number to beat for any generative claim (Target sequence-conditioned design of peptide binders using masked language modeling). D-SCRIPT and PPLM serve as verifier baselines: D-SCRIPT for the contact-map idea with single-chain embeddings (Drosophila AUPR 0.552 versus PIPR's 0.231, but zero top-1% retrievals on the surfaceome screen), and PPLM for paired cross-attention without our contact-bottlenecked scoring head (81.3% top-$L$ contact precision on a challenging interface versus 16.7% without paired features, per A paired sequence language model for protein-protein interaction modeling). Fairness is enforced by giving every method the same training corpora where retraining is needed, the same single-GPU compute cap, the same random-search tuning budget, and identical decoy sets and split definitions. The key ablation removes the contact map from our verifier, leaving paired embeddings that score binding directly; if hard-split performance survives the removal, the contact map is decoration rather than the mechanism, and the core claim is wrong. A method beats a baseline when it significantly improves top-1% retrieval and hard-split AUROC at matched tuning budget, and — for the generative arm — when it exceeds the 38% AlphaFold-Multimer hit-rate reference on the Alzheimer's panel.
Why not just…?
Why not just run PepMLM or SaLT&PepPr on S100A9 tomorrow? This is the most serious objection, and Experiment 1 is designed to answer it rather than argue around it. The published evidence says both tools should be tried and both should wobble: deep sequence predictors failed on short therapeutic peptides and on de novo targets exactly like ours, assigning near-zero probabilities to validated pairs (Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering), and SaLT&PepPr cannot generate de novo binders at all — it cuts peptides from documented endogenous partners, which S100A9 largely lacks (SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders). PepMLM also generates without scoring, so it offers no off-target triage for a brain-delivered therapeutic. If the off-the-shelf run surprises us and passes the decoy screen, we will report that as a finding and pivot the contribution to the specificity and wet-lab layers — but the memorization literature gives strong grounds to expect it will not.
Why not just use 3D structure tools like RFdiffusion or AlphaFold-Multimer? Three reasons from the file, not from taste. More than 75% of the human proteome lacks experimental structures, and S100A9's pathogenic state is a phosphorylated, fibril-forming assembly that rigid-body methods handle poorly (SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders). On the shared evaluator, the sequence-only generator already beat the structure-based one, 38% versus 29% (Target sequence-conditioned design of peptide binders using masked language modeling). And paired sequence contact prediction has closed much of the structural gap, outperforming AlphaFold2.3, AlphaFold3, and DMFold by up to 8.1% in top-$L$ precision (A paired sequence language model for protein-protein interaction modeling). We are not hostile to structure: AlphaFold-Multimer is our independent judge, which keeps the evaluation honest precisely because it is not our method.
Why not just use SPRINT, since it beat the deep models? SPRINT wins by lookup: it ranks targets whose homologs appear in its database of known interactions, which is why it took 16 of 21 pairs in the optimistic setting and failed every pair in the de novo setting (Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering). S100A9's cross-seeding interface and a novel Keap1-blocking peptide have no training-database homologs to look up, and a similarity tool cannot output a new sequence. SPRINT stays in our baseline set as the honest upper bound for cheap methods; our claim is only that it cannot solve the de novo problem, which its own authors demonstrated.
Why not just use the small molecules that already work, like EDB? EDB is a genuine threat to the application half of this proposal: it binds S100A9 at nanomolar affinity and cut Y-maze performance decline in APP/PS1 mice from 80% to about 27% (Edaravone-Dexborneol slows down pathological progression and cognitive decline via inhibiting S100A9 in APPswe/PS1dE9 mice). But its own authors flag that the compound is pleiotropic and that S100A9-specific causality still needs knockout validation, and the Keap1–Nrf2 and S100A9 interfaces are exactly the flat, pocketless surfaces that small molecules struggle with (Target sequence-conditioned design of peptide binders using masked language modeling). Peptides buy sequence-programmable specificity and, ultimately, degradation via ubiquibody grafting, at the acknowledged cost of blood–brain-barrier delivery, which we treat as an explicit limitation and a downstream engineering problem rather than something this proposal solves. The comparison that matters here is mechanistic, not clinical: does a designed peptide block cross-seeding and restore Nrf2 in human mini-brains, which no existing molecule of any modality has shown for these designed sequences.
Decisive Checks
The proposal is falsifiable at three gates. Gate one is computational: if the contact-grounded verifier does not materially narrow the random-to-hard-split AUROC gap that felled the TCR models, the core mechanism is wrong and the ablation will show whether the contact map was ever load-bearing. Gate two is generative: if the pipeline's AlphaFold-Multimer hit rate on the Alzheimer's panel does not exceed PepMLM's published 38% reference under matched evaluation, the added machinery has not paid for itself. Gate three is mechanistic: SPR affinity for S100A9 binders in the nanomolar-to-sub-micromolar range, slowed S100A9-mediated Aβ fibrillization in the thioflavin-T assay, and reduced microglial CD86, iNOS, and phospho-tau in the 3D mini-brain platform would together establish that sequence-only design can be pointed at non-amyloid Alzheimer's pathology and hit. Each gate has a published number attached, so at every stage it will be plain whether the idea is working or whether the simpler alternatives should have been enough.
References
- Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease
- Edaravone-Dexborneol slows down pathological progression and cognitive decline via inhibiting S100A9 in APPswe/PS1dE9 mice
- Assessing sequence-based protein–protein interaction predictors for use in therapeutic peptide engineering
- On TCR binding predictors failing to generalize to unseen peptides
- Target sequence-conditioned design of peptide binders using masked language modeling
- A paired sequence language model for protein-protein interaction modeling
- Sequence-based prediction of protein-protein interactions: a structure-aware interpretable deep learning model
- Predicting Peptide Aggregation with Protein Language Model Embeddings
- The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research
- SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders