PaperlayerView exact revision on arXiv

Source-linked AI summary

RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation

Liuzhenghao Lv, Yuyang Liu, Yuyang Gao, Li Yuan, Yonghong Tian

arXiv:2610.01891v1cs.CEq-bio.BM

TL;DR

Mutation effect generation must describe a mutation’s functional consequence in natural language, but existing representations do not explicitly organize structural and biochemical evidence. RipplePLM uses contact-guided structural pathways and latent property supervision; it improves over mutation-specific baselines on temporal and structural splits, with expert evaluation also favoring its biological relevance.

  • Problem

    Mutation effect generation requires fluent descriptions grounded in mutation-specific structural and property evidence, which existing undifferentiated representations do not explicitly organize.

    Source: §1 · pp. 1–2

  • Method

    RipplePLM constructs a residue-level Mutation Perturbation Field and uses contact-guided DDCA for direct and distal context alongside PLChain latent supervision for biochemical property changes.

    Source: §1 · p. 2, §2.1 · p. 3

  • Results

    +13.42 absolute average ROUGE-L gain over matched-backbone MutaPLM* on the structural split; RipplePLM also improves on the temporal split and receives stronger expert ratings for biological accuracy or relevance.

    Source: §1 · p. 2, §3.2 · p. 6

  • Takeaways & Limitations

    On MutaDescribe, RipplePLM’s structured mutation representations are associated with improved generation over evaluated mutation-specific baselines across temporal and structural splits.

    Source: §5 · p. 10

    RipplePLM focuses on single-point mutations, and DDCA’s structural signal may depend on the quality of ESM-2-predicted contact maps.

    Source: §5 · p. 10

Abstract

from arXiv · show

Protein mutation effect generation asks a model to describe the functional consequence of a point mutation in natural language. Existing protein-to-text systems typically encode mutation information into undifferentiated representations, overlooking the organization of mutation-induced evidence across structural and biochemical factors. We propose RipplePLM, a mutation-aware generation framework centered on Direct-Distal Cross-Attention (DDCA). By constructing a residue-level Mutation Perturbation Field from pre-trained protein language models, DDCA leverages predicted contact maps to organize mutation representations into two pathways: the mutation site's immediate contact neighborhood and its multi-hop distal context. To complement this structural decomposition, we further introduce the Property Latent Chain (PLChain), which injects expert-guided supervision of biochemical property changes (e.g., thermostability and optimal pH) into the LLM hidden-state pathway through latent property tokens. On MutaDescribe, RipplePLM improves over mutation-specific baselines on temporal and structural splits; under a matched-backbone comparison, average structural-split ROUGE-L increases from {22.23} to {35.65}. Expert evaluation further shows a higher proportion of biologically accurate or relevant descriptions than the mutation-specific baseline. Additional ablations, representation diagnostics, and low-$N$ fitness regression experiments further support the effectiveness of the learned mutation-aware representations. Code: https://github.com/Lyu6PosHao/RipplePLM.

1 Introduction

Mutation effect generation aims to describe a mutation’s functional consequence in natural language, but useful descriptions must account for structural and biochemical evidence. RipplePLM organizes that evidence with direct and distal structural pathways and property-aware supervision, and improves over evaluated mutation-specific baselines.

1 Introduction

  • Mutation effect generation seeks natural-language descriptions of functional consequences, which must ground fluent text in mutation-specific evidence from the residue, structural neighborhood, distal context, and biochemical properties.The paper frames these dependencies as mutation perturbations propagating through contact networks and giving rise to measurable biochemical changes.Source: §1 · pp. 1–2
  • Existing protein–language systems align protein encoders with LLMs, while MutaPLM models protein deltas; neither explicitly separates local from distal contact evidence, and prompt text leaves property supervision disconnected from generation representations.Expert models provide mutation-level signals such as thermostability and optimal pH, but ordinary prompts reduce these to coarse symbolic cues.Source: §1 · p. 2
  • RipplePLM combines contact-guided Direct-Distal Cross-Attention (DDCA) with Property Latent Chain (PLChain) to organize structural evidence and inject biochemical property supervision into generation.DDCA separates immediate contact-neighborhood evidence from distal multi-hop context; PLChain uses expert-guided latent tokens for thermostability and optimal-pH changes.Source: §1 · p. 2
  • RipplePLM improves over evaluated mutation-specific baselines on temporal and structural splits, with expert evaluation finding more biologically accurate or relevant descriptions than for MutaPLM.The evaluation also includes ablations and representation analyses of the structural decomposition and property-aware supervision.Source: §1 · p. 2
  • DDCA separates direct contact-neighborhood evidence from distal contact-graph context, while PLChain integrates expert property signals into the LLM hidden-state pathway.The study combines lexical and expert-based biological evaluations with ablations and representation diagnostics.Source: §1 · pp. 2–3

2 Method

RipplePLM turns a point mutation and optional context into a natural-language description by preserving residue-level mutation changes and organizing them into structural and biochemical conditioning signals. Its method combines contact-guided direct and distal pathways with expert-supervised property tokens, trained in two stages.

2.1 Mutation Perturbation Field (MPF)

  • The Mutation Perturbation Field (MPF) is a residue-indexed field of mutant-minus-wild-type frozen ESM-2 representations, retaining changes beyond the mutated position for contact-guided aggregation.The method extracts final-layer representations for both sequences and constructs their difference.Source: §2.1 · p. 3
  • Direct-Distal Cross-Attention (DDCA) separates first-shell contact evidence from distal evidence reached through multiple steps on the predicted contact graph.The wild-type contact map supplies a shared pre-mutation topology; per-head decay allows different effective distal radii.Source: §2.1 · p. 3
  • DDCA adds structural weights as learnable attention-logit biases rather than hard masks, so attention is guided by but not restricted to predicted contacts.This choice addresses noisy predicted contacts and allows content-based attention to adjust residue weighting.Source: §2.1 · p. 4

2.2 Property Latent Chain (PLChain)

  • PLChain uses expert pseudo-labels for thermostability and pH-optimum changes to supervise property-token hidden states that also feed autoregressive generation.Auxiliary heads project the property-token states to expert label spaces, and the auxiliary loss regularizes their representations.Source: §2.2 · p. 5

2.3 Training Procedure

  • RipplePLM trains in two stages with ESM-2 frozen: first MPF under language modeling loss, then PLChain and LoRA with an annealed property-loss weight.The staged approach first learns a structural interface before adding property supervision; the property weight declines from 0.4 to 0.2.Source: §2.3 · p. 5

3 Experiments

Experiments report gains on MutaDescribe’s temporal and structural splits, with ablations and representation analyses examining the contributions and learned signals of RipplePLM’s components. Low-N fitness regression also evaluates the mutation representations beyond text generation.

3.1 Experimental Setup

  • The evaluation uses MutaDescribe’s temporal and structural splits, a matched-backbone MutaPLM* comparison, and BLEU, ROUGE, and METEOR generation metrics.RipplePLM uses frozen ESM-2 with a DeepSeek-R1-Distill-Llama-8B backbone adapted via LoRA.Source: §3.1 · pp. 5–6

3.2 Main Results

  • On the temporal split, RipplePLM outperforms evaluated baselines across six metrics; expert evaluation reports 48.56% biologically accurate or relevant descriptions versus MutaPLM’s 40.06%.RipplePLM also remains ahead after output-length truncation; one-shot general-purpose LLMs remain below it on the structural split.Source: §3.2 · pp. 6–7
  • A +13.42 average ROUGE-L gain over matched-backbone MutaPLM* accompanies a Hard-subset score of 31.27, above MutaPLM’s Easy-subset 25.80.The structural split tests performance across varying degrees of structural novelty.Source: §3.2 · p. 6, Table 2 · p. 7
  • Case studies show more mutation-specific detail in some outputs, including the 100-fold GTPase-activity decrease for N226A, while also documenting examples where both models remain imperfect.For G93A, RipplePLM retains TDIF–TDR interaction context that MutaPLM* omits.Source: §3.2 · p. 7

3.3 Ablation Studies

  • Removing DDCA’s distal pathway or PLChain degrades temporal-split metrics; removing function context lowers structural ROUGE-L, while a 650M ESM-2 encoder raises it to 36.43.Distal-pathway removal drops BLEU-2 by 1.39 and ROUGE-L by 3.18; PLChain removal drops them by 2.06 and 3.90. Shuffled property labels recover negligibly over no PLChain, and discrete prompt labels underperform the full model.Source: Table 3 · p. 7, §3.3 · pp. 7–8

3.4 Representation Analysis

  • Perturbation remains 9.1× above background at two contact-graph hops; direct weights concentrate at one hop, while distal weights peak at two hops.The direct pathway assigns 70.9% of normalized weight to 1-hop residues; the distal pathway assigns 58.7% to 2–4-hop residues. Direct and distal tokens have a silhouette score of 0.74.Source: §3.4 · pp. 8–9, Figure 3 · p. 8
  • PLChain output tokens achieve 90.2% probe accuracy for thermostability and 88.3% for optimal pH, whereas pre-LLM MPF queries are near chance.The output tokens form property-specific clusters after LLM processing.Source: §3.4 · p. 9

3.5 Low-N Fitness Regression

  • With 192 labeled mutations, RipplePLM has the highest mean Spearman correlation on both Spike-ACE2 and avGFP, though standard deviations overlap those of the strongest baselines.The regression uses frozen protein features without textual supervision.Source: Table 4 · p. 9, §3.5 · p. 9

4 Related Work

Prior protein–language systems connect protein representations to text generation and instruction-following, while many mutation-effect predictors output scores or labels rather than explanations. RipplePLM is distinguished by structuring mutation-specific evidence into direct and distal pathways and adding property-aware latent supervision.

Protein–language alignment

  • Protein–language alignment spans contrastive or knowledge-guided pretraining, protein-conditioned description generation, and LLM-based protein dialogue and instruction following.These systems establish the feasibility of protein-conditioned generation, but most treat protein input as a static object.Source: §4 · p. 9

Mutation effect prediction and generation

  • Many mutation-effect systems predict scores or labels; MutaPLM [10] instead formalizes natural-language generation with a protein delta network and literature-mined reasoning supervision.Specialized systems also predict biochemical properties such as stability, solubility, and optimal pH.Source: §4 · p. 10
  • RipplePLM differs from prior generation approaches by separating direct contact-neighborhood evidence from distal contact-graph context and adding property-aware latent supervision.The related-work discussion notes that AugmentedESM and ESM-based generation baselines do not explicitly separate these structural contexts.Source: §4 · pp. 9–10

5 Conclusion

RipplePLM organizes mutation evidence with structural pathways and property-aware latent supervision. Experiments support its performance on the evaluated splits and suggest its mutation representations are useful for low-N fitness prediction, while the authors identify scope and validation limits.

Conclusion

  • RipplePLM combines DDCA’s direct and distal structural evidence pathways with PLChain’s property-aware latent supervision.The conclusion describes this as structured conditioning rather than undifferentiated protein representations.Source: §5 · p. 10
  • Experiments show improvements over evaluated mutation-specific baselines on both MutaDescribe splits, and low-N regression suggests MPF features are useful for fitness prediction.The authors propose extending the framework to multi-mutation settings and larger community-curated benchmarks with richer experimental contexts.Source: §5 · p. 10

Limitations

  • The study focuses on single-point mutations; DDCA’s structural signal may depend on predicted-contact quality, and linking learned signals to biophysical mechanisms requires experimental validation.Performance also depends in part on wild-type function annotation availability and quality.Source: §5 · p. 10

Broader Impacts

  • The authors suggest RipplePLM may assist protein engineering, variant interpretation, and literature-driven discovery, but position outputs as decision support rather than substitutes for validation or clinical expertise.Source: §5 · p. 10

A Implementation Details

RipplePLM uses a frozen ESM-2 encoder and staged mutation-conditioning training, with structural and property pathways supplying evidence for generation. In low-N fitness regression, it has the highest mean Spearman correlation on both datasets, but variability overlaps with the strongest baselines.

A Implementation Details

  • Training proceeds in two stages: MPF-only adaptation for one epoch, followed by PLChain-enabled LoRA adaptation for 10 temporal-split epochs or two structural-split epochs.Stage 1 uses no property supervision; Stage 2 applies LoRA to attention and feed-forward projections.Source: A Implementation Details · p. 15
  • The complete training objectives include overlap regularization between normalized direct and distal pathway weights, with λoverlap fixed at 0.05.The penalty acts on contact-derived weights used to construct pathway seeds and attention biases.Source: A Implementation Details · p. 15
  • The structural diagnostic analyzes 110 temporal-test examples, summing pathway weights over reachable contact-graph hops 1–10 before averaging across examples.Examples have protein lengths of 100–300 residues and at least three predicted contacts above 0.5 at the mutation site; the contact graph is thresholded at 0.3.Source: A Implementation Details · pp. 15–16
  • At inference, PLChain auxiliary losses are disabled while learned property tokens remain in the conditioning sequence; average end-to-end latency is approximately 1s per mutation.The reported inference setup uses greedy decoding with a 1024-token maximum output length and batch size 1 on one A100 GPU.Source: A Implementation Details · p. 16
  • MutaDescribe evaluation follows official temporal and structural splits, with PLChain using PPT-Stab and EpHod pseudo-labels for thermostability and pH optimum.The temporal split tests descriptions from later literature; the structural split groups proteins by maximum sequence homology to training proteins.Source: A Implementation Details · p. 16
  • The prompt supplies wild-type function context and replaces protein and property placeholders with learned MPF/DDCA structural tokens and PLChain property queries.The matched-backbone MutaPLM* comparison also receives ground-truth function context to focus the comparison on mutation conditioning.Source: A Implementation Details · pp. 16–17, Table 5 · p. 16

B Low-N Fitness Regression

  • The low-N protocol trains a lightweight regression head on 192 labeled mutations and evaluates Spearman correlation over five random runs.RipplePLM augments the shared frozen ESM-2 feature views with MPF features extracted without updating the generator.Source: B Low-N Fitness Regression · p. 17
  • RipplePLM has the highest mean Spearman correlation on both datasets, although standard deviations overlap with the strongest baselines.The authors treat this as auxiliary evidence that MPF retains fitness-relevant information, not as definitive superiority in quantitative prediction.Source: B Low-N Fitness Regression · p. 17

C Full Structural Split Results

Table 6 reports RipplePLM's full structural-split results across six metrics.

C Full Structural Split Results

D Case Studies

The case studies highlight settings where mutation-specific structural or mechanistic evidence is relevant, alongside examples of remaining description errors. They also include a competitive counterexample.

D Case Studies

  • The qualitative comparison selects representative Easy, Medium, and Hard structural-split cases covering partial matches, quantitative effects, interaction-specific descriptions, and competitive counterexamples.RipplePLM is compared with MutaPLM* on these examples.Source: D Case Studies · p. 17, Table 7 · p. 18
  • RipplePLM is most effective in cases where reference descriptions depend on mutation-specific structural or mechanistic evidence.Examples include GTPase reduction, TDIF–TDR interaction, and SEC22B cargo packaging.Source: D Case Studies · p. 17
  • The examples reveal errors including omitted secondary clauses, substituted downstream outcomes, and confusion between related biochemical activities.The Q49A counterexample indicates that structured conditioning does not eliminate all enzyme-activity errors.Source: D Case Studies · p. 17

E.1 Biology-Aware Evaluation

An expert assessed all RipplePLM outputs on the structural test set for agreement with reference descriptions in biological function or process and effect direction. The review reports higher Accurate-or-Relevant frequency and lower Opposite frequency than MutaPLM.

E.1 Biology-Aware Evaluation

  • The Opposite rating rate fell from 7.13% for MutaPLM to 6.41% for RipplePLM.The evaluation complements lexical metrics with expert assessment of biological fact agreement.Source: §E.1 · p. 18
  • A PhD-level biology expert assigned four categories to all 1,248 RipplePLM structural-test outputs using agreement in biological function or process and mutation-effect direction.The categories were Accurate, Relevant, Opposite, and Irrelevant.Source: §E.1 · p. 18, Table 8 · p. 18
  • 48.56% of RipplePLM outputs were rated Accurate or Relevant, an 8.50-percentage-point increase over MutaPLM.Table 8 presents the expert-reviewed distribution alongside published MutaPLM results.Source: Table 8 · p. 18, §E.1 · p. 18

E.2 Response Length

Truncation raises temporal-split ROUGE-L, while output length is negatively correlated with per-example ROUGE-L on both splits. Bootstrap intervals exceed published MutaPLM estimates, and RipplePLM retains an advantage over recent general-purpose LLMs on the structural split.

E.2 Response Length

  • Output length is negatively correlated with per-example ROUGE-L on the temporal and structural splits, indicating length is unlikely to be the main source of generation gains.RipplePLM averages 30.65 words versus 28.03 in temporal references, and 21.73 versus 31.87 words on the structural split.Source: §E.2 · pp. 18–19
  • ROUGE-L rises from 28.56 to 32.68 after truncation; METEOR, ROUGE-1, and ROUGE-2 improve, while BLEU-2 and BLEU-4 decrease slightly.Table 9 compares temporal-split performance before and after truncating over-length outputs to reference lengths.Source: Table 9 · p. 19, §E.2 · p. 19

E.3 Statistical Uncertainty

  • 95% bootstrap interval lower bounds exceed corresponding published MutaPLM point estimates for BLEU-2 and ROUGE-L on both splits.Intervals resample test examples 10,000 times; the temporal and structural sample sizes are 1,611 and 1,248, respectively.Source: Table 10 · p. 19, §E.3 · p. 19

E.4 Recent General-Purpose LLMs

  • Recent LLMs score broadly comparably to GPT-4-0613, while RipplePLM retains a substantial advantage on both structural-split metrics.The comparison evaluates GPT-5.4 and DeepSeek-V4 variants on all 1,248 structural-test examples in a one-shot setting without external tools or retrieval.Source: Table 11 · p. 19, §E.4 · p. 19

F.1 Wild-Type Function Context

Function context improves RipplePLM’s ROUGE-L, but removing it leaves the model ahead of MutaPLM in the reported comparisons. A larger encoder improves four of six structural-split metrics, and judged relevance increases with predicted-contact confidence.

F.1 Wild-Type Function Context

  • Removing function context cuts RipplePLM ROUGE-L by 7.10 points (19.9%), yet it exceeds no-function MutaPLM by 10.01 points and with-function MutaPLM by 4.75.Both models receive the same ground-truth wild-type function description in the matched-backbone comparison; function descriptions have ROUGE-L overlap of 9.76 with references, and no reference is an exact substring.Source: §F.1 · pp. 19–20, Table 12 · p. 20

F.2 Protein Encoder Size

  • ESM-2 650M improves four of six structural-split metrics, including METEOR from 29.30 to 30.45 and ROUGE-L from 35.65 to 36.43, with small BLEU-2 and BLEU-4 decreases.The comparison changes only the MPF input projection and retains other hyperparameters and training settings; the 150M encoder remains the main setting to balance predictive performance and computational cost.Source: Table 13 · p. 20, §F.2 · p. 20

F.3 Predicted-Contact Confidence

  • Contact confidence is the mean ESM-2 150M probability over top-L long-range residue pairs; 1,247 of 1,248 structural-test examples had eligible pairs for grouping.Examples were divided into three approximately equal-sized confidence groups.Source: §F.3 · p. 20
  • Accurate-or-Relevant rates rise from 44.23% in the lower-confidence group to 54.22% in the higher-confidence group, an association rather than a test of contact accuracy.The analysis uses expert-reviewed labels and treats predicted contacts as a structural prior alongside content-based attention.Source: Table 14 · p. 20, §F.3 · p. 20

Related papers

Loading 2610.01891v1…