State-Pair Prototypes: Breaking Few-Shot Saturation by Modeling Ligand-Induced Asymmetry

Many proteins work as pairs of identical chains, yet a single small molecule can force the two identical partners into two different shapes. Structural biology has documented this repeatedly: one leptin molecule bridges two identical receptor subunits and locks them into non-equivalent roles (Saxton et al., 2023); a single water molecule infiltrating one pocket of the SARS-CoV-2 main protease collapses one active site while the other stays active (Ansari et al., 2021); and leucine-rich-repeat receptor homodimers pack by a screw transformation — rotation plus translation — so one protomer physically blocks the other's binding site (Gong et al., 2025). In short, identical sequences in identical environments routinely break symmetry when a ligand arrives.

Meanwhile, few-shot learning — adapting a model to a new task from a handful of examples — has a stubborn empirical ceiling: performance jumps from 1 to 5 examples and then flattens, whether the task is cell segmentation (Dawoud et al., 2020), teachable object recognition (Massiceti et al., 2021), or kinase-peptide binding prediction, where MetaKinase reached an Area Under the Precision-Recall Curve (AUPRC) of 0.316 with just 2 samples and only 0.338 with 10 (Luo, Ma, et al., 2019). Our claim is that these two facts are connected in molecular machine learning: when the true states of a protein system are mutually exclusive asymmetric configurations of identical chains, the standard few-shot recipe — average the support examples into one class prototype — averages away exactly the information that additional examples should be adding. The plateau is not (only) a statistics problem; it is a representation problem caused by forcing an inherently asymmetric state space into a symmetric class description.

Core Claim

We propose to build and test state-pair prototypes: a few-shot representation in which each support class is encoded not as a single centroid but as an ordered pair of protomer-state embeddings $(p_c^{A}, p_c^{B})$, linked by a learned geometric operator and matched to queries with an asymmetric measure. Two ingredients come directly from the structural literature. First, the ordered-pair structure encodes ligand-induced asymmetry: the two entries are the non-equivalent conformational states that a single ligand stabilizes, as seen in leptin receptor activation (Saxton et al., 2023), HIV-1 Env stepped activation (Ma et al., 2018), and PAS-domain cross-regulation of a partner's catalytic domain (Stanton et al., 2025). Second, the asymmetric matching measure follows Li et al., 2020, who showed that comparing a single query to a whole support class is inherently directional and that symmetric distances lose information; we extend their intuition from "query vs. class" to "protomer state vs. paired protomer state," where the direction encodes which subunit the ligand occupies first.

Concretely, a class $c$ with support set $S_c$ is embedded as an ordered pair rather than a mean:

$$p_c^{A}, p_c^{B} = \arg\min_{u,v} \sum_{x \in S_c} \min\big(d(x,u),\ d(T_\theta x, v)\big),$$

where $T_\theta$ is a learned screw transformation (rotation composed with translation, the operator Gong et al., 2025 showed governs asymmetric homodimer packing) and $d$ is an asymmetric divergence — Kullback-Leibler (KL) divergence between Gaussian embeddings, following Li et al., 2020. A query is assigned to a state, then to a class; a conditional binding head then models how occupancy of state $A$ changes the binding geometry of state $B$, which is the conditional protein-protein interaction (C-PPI) structure that PromptMSP identified as essential but implemented only for rigid chains.

The falsifiable prediction is sharp: if few-shot saturation in molecular tasks is partly an artifact of symmetric prototype pooling, then replacing the prototype with an ordered state pair should convert the flat learning curve into a rising one — gains should continue past $k=10$ where MetaKinase gains only 0.022 AUPRC between $k=2$ and $k=10$ — while leaving the backbone, data, and tuning budget untouched.

Research Program

The program has three parts. First, we build the state-pair module as a drop-in replacement for prototype pooling on top of frozen protein language model embeddings, so that every comparison against prototype-based baselines isolates the representation change and nothing else. Second, we construct LigAsym-Bench, an evaluation set of homodimers with documented ligand-induced asymmetry, drawn from the systems resolved in the structural literature: the LGR screw-transform dimer (Gong et al., 2025), leptin receptor (Saxton et al., 2023), SARS-CoV-2 main protease (Ansari et al., 2021), HIV-1 Env (Ma et al., 2018), the FlrB PAS-domain dimer (Stanton et al., 2025), KAP1 (Fonti et al., 2019), and Rpd3S (Guan et al., 2023), paired with symmetric homodimer controls. Each entry carries the two protomer states and the conditional binding annotation. Third, we run the saturation analysis: learning curves over $k$ from 0 to 50 on both the molecular tasks and a vision control, testing whether the plateau is a general property of symmetric class descriptions or specific to bimodal molecular state spaces.

The connective hypothesis draws on two prior observations. Dernoncourt et al., 2021 found that as support sets grow, complex meta-adaptation machinery loses its edge over simple fine-tuning — machinery built for extreme scarcity becomes superfluous — and Johnander et al., 2022 showed prototype methods plateau after 2–3 support examples because averaging cannot represent multi-modal variation. We sharpen both: in ligand-induced asymmetric systems, the modes are not arbitrary appearance variation but a structurally constrained pair whose geometry (screw offset, conditional occupancy) is known. That is what lets us predict the fix, not just observe the plateau.

Baselines

The "do nothing" reference point is a plain prototypical network: support examples embedded by a frozen protein language model, averaged into one centroid per class, classified by nearest centroid. This is the standard recipe that Johnander et al., 2022 showed saturates after 2–3 examples, and on the kinase-peptide phosphorylation benchmark no reported result exists for this exact configuration on Lacuna; we will run it ourselves under our budget to establish the plateau in our setting. It matters because it is the minimal system our hypothesis says should saturate, and it shares our backbone, splits, and tuning budget so any difference is attributable to the state-pair representation alone.

The strongest existing method on the molecular few-shot question is MetaKinase (Luo, Ma, et al., 2019), which meta-trains across kinase families and few-shot fine-tunes to data-poor kinases; it reports AUPRC 0.316 at $k=2$ and 0.338 at $k=10$ on the kinase-peptide phosphorylation evaluation set, the exact plateau we target. We will rerun it on the same splits with the same support-set construction and the same embedding budget, extending the sweep to $k=50$. Beating it means two things jointly: equal or higher AUPRC at every $k$, and a learning curve that keeps rising where theirs flattens — specifically, a gain from $k=10$ to $k=50$ strictly larger than their observed $k=2$-to-$k=10$ gain of 0.022 AUPRC. On assembly, the strongest competitor is PromptMSP (ICLR 2024), which scores an average TM-Score of 0.67 on PDB-M multimers with 3–10 chains against 0.52 for the prior best (MoLPC); beating it means a higher TM-Score on a flexible, ligand-conditioned subset of PDB-M where its rigid-chain assumption should hurt, run with matched compute (its path prediction already costs only ~0.04 minutes per 30-chain multimer, so compute parity is easy).

The key ablation isolates the proposal's central ingredient: the same state-pair module but with the ordering and asymmetric measure removed — two unordered centroids from K-means on the support set, matched with a symmetric distance (Wasserstein, the symmetric control Li et al., 2020 used in their ablations) and no screw operator $T_\theta$. If unordered multi-prototypes match ordered state pairs, the structural story collapses to "use more centroids," and we would report that honestly.

Why not just…?

Why not just scale the backbone or fine-tune on more data? This is the most common dismissal of saturation as a real phenomenon, and the file contains direct counter-evidence. Metalic built on a tiny ESM2-8M backbone and outperformed models eighteen times larger on ProteinGym, showing raw parameter scaling does not fix adaptation; and Boudiaf et al., 2021 showed a standard pretrained backbone saturates inductively and only a change in inference, not capacity, broke the ceiling. We will still run a backbone-scaling control within our budget, but the prior evidence predicts it fails, and our hypothesis explains why: a bigger encoder of a symmetric class description still averages away the state pair.

Why not just cluster support features into a few prototypes (multi-prototype or mixture models)? This is the most serious alternative, and we treat it as a real threat rather than arguing around it. Two pieces of evidence suggest it falls short: Li et al., 2020 found that symmetric distances underperform asymmetric KL even in multi-prototype settings (54.26% versus DN4's baseline on miniImageNet 1-shot, and 56.01% on tieredImageNet), and Johnander et al., 2022 observed prototype-style methods plateau regardless of clustering. But the decisive test is ours to run: the K-means ablation above. If unordered two-centroid models break the saturation as well as ordered state pairs, the asymmetry story is wrong and the mixture story suffices; if they break it only partially, the residual gap measures how much the ordering and conditional geometry contribute.

Why not just use transductive inference at test time? RePRI (Boudiaf et al., 2021) broke saturation in vision with test-time optimization on the query set, gaining +5.1% at 5-shot and +6.1% at 10-shot over the then-best methods on PASCAL-$5^i$, and its code is public (RePRI codebase). This is a genuine competitor because it needs no structural prior at all. We will run a RePRI-style transductive variant on our molecular tasks with the same data and iteration budget. Our expectation is that transduction helps when unlabeled queries span both protomer states (it implicitly discovers the bimodality) but cannot label which state a query occupies or model the conditional binding geometry, so it should improve accuracy without recovering the state-level predictions LigAsym-Bench scores; if it matches state-pairs on every metric, the structural representation is unnecessary and we will say so.

Why not just use prompt learning or in-context learning? PromptMSP adapts multimer prediction with learnable prompts but explicitly requires pre-computed rigid dimer structures and flags flexible docking — precisely what ligand-induced symmetry breaking is — as out of scope. Metalic excels at single-mutant fitness (Spearman 0.484 zero-shot on ProteinGym) but degrades on multi-mutant landscapes because in-context conditioning alone cannot resolve higher-order structural coupling. Both constrain the design space: conditioning without an asymmetric state representation has already been shown to hit its own wall.

Proposed Experiments

Experiment 1 runs the saturation test on the kinase-peptide phosphorylation evaluation set from MetaKinase, extending the support sweep from its reported $k \in [0,10]$ to $k=50$, measuring AUPRC at each $k$, with MetaKinase (0.316 at $k=2$, 0.338 at $k=10$), the plain prototypical network, the unordered multi-prototype ablation, and the transductive variant all under identical splits and budget. The decisive readout is curve shape, not just endpoint: state-pair prototypes must show continued gains beyond $k=10$ where every symmetric baseline flattens.

Experiment 2 evaluates on LigAsym-Bench: given a few-shot support of labeled protomer states, predict for a held-out homodimer whether a ligand induces asymmetry, which protomer occupies which state, and the sign of the resulting negative cooperativity. Metrics are AUROC and AUPRC for the asymmetry call and state-assignment accuracy, again as $k$-curves. The structural papers supplying the positive entries — Gong et al., 2025, Saxton et al., 2023, Ansari et al., 2021, Ma et al., 2018 — defined these states experimentally, so the ground truth is structural, not model-derived. No reported result exists on Lacuna for any model on this task, which is exactly why the benchmark is worth building.

Experiment 3 tests assembly scoring on a ligand-conditioned subset of PDB-M against PromptMSP (TM-Score 0.67 on 3–10 chain multimers; MoLPC at 0.52), checking whether the conditional, geometry-aware state pair recovers assembly paths that rigid prompting misses. Experiment 4 is the vision control: the same ordered-pair, asymmetric-KL matching module transplanted into few-shot segmentation on PASCAL-$5^i$ using the public RePRI codebase, compared against RePRI's reported +5.1%/+6.1% margins at 5/10-shot and DGPNet's continued scaling to 10 shots (it reports +8.4 mIoU on COCO-$20^i$ over prior methods). This checks whether the mechanism generalizes beyond molecules or is genuinely a structural-biology story — either answer is informative, and the ablations above ensure we can tell the difference.