Search bioRxiv⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.10.31.763384

Parameter-Efficient Adaptation of Protein Language Models for Dual-Target Antimicrobial Peptide Discovery Under Strict Homology Constraints

Jat, R.·Ejaz, S. R.
View original source ↗Export metadata ↓

Abstract

The worldwide escalation of antimicrobial resistance (AMR) poses a severe challenge to modern public health, necessitating novel classes of therapeutic agents. Antimicrobial peptides (AMPs) represent highly promising alternatives to small-molecule antibiotics due to their co-evolutionary resilience and rapid membrane disruption dynamics. However, translating computational AMP leads into clinical drug candidates remains bottlenecked by off-target cytotoxicity, predominantly hemolytic activity against mammalian erythrocytes. Although pretrained foundational protein language models (PLMs) have rapidly advanced molecular property prediction, current architectures frequently isolate antibacterial potency and cytotoxicity into disconnected single-task pipelines or succumb to catastrophic performance drops under strict homology-controlled evaluations. Recently, the Quantitative Mapping of Antimicrobial Peptides (QMAP) benchmark proposed that frozen PLM embeddings show weak linearly accessible features for hemolytic activity (HC50) resulting in poor rank correlation ({rho} < 0.25) on out-of-distribution sequence splits. In this work, we introduce an end-to-end multi-task parameter-efficient transfer learning framework designed to simultaneously optimize antibacterial potency and minimize mammalian cytotoxicity without sequence leakage. Built upon ESM-2 (150M), our architecture incorporates Low-Rank Adaptation (LoRA) into self-attention projections, fine-tuning approx 0.70% of model weights. To restore gradient interference originating from disparate membrane lysis mechanisms, we implement an adaptive multi-task loss guided by Kendall homoscedastic uncertainty weighting. Evaluated across predefined QMAP benchmark splits enforcing a 60% maximal sequence identity threshold, our joint architecture remarkably outperforms conventional linear probes, tree-based regressors, and frozen foundational models. The framework achieves an E. coli potency rank correlation of {rho} = 0.6132 (r = 0.6816) and successfully elevates hemolytic toxicity correlation to {rho} = 0.5374 (r = 0.5541). To establish real world utility, we deploy the optimized joint network to screen a curated in silico library of 749 de-novo designed peptides, successfully isolating selective cationic-amphipathic candidates possessing Therapeutic Indices (TI = HC50/MIC) up to 144.03 while effectively deprioritizing broadly lytic motifs (TI < 2.0). These findings demonstrate that parameter-efficient adaptation unlocks fine-grained bioactive structural semantics in foundational PLMs, providing an accurate, scalable framework for therapeutic peptide engineering.

Explore related subjects

bioinformaticsPreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jat, R., Ejaz, S. R.. 2026-11-03. Parameter-Efficient Adaptation of Protein Language Models for Dual-Target Antimicrobial Peptide Discovery Under Strict Homology Constraints. https://doi.org/10.64898/2026.10.31.763384

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

PHATE-Derived Ancestry Coordinates Capture Continuous Population Structure in Human Genomic Data

Population genetic data are increasingly visualized using low-dimensional embeddings as ancestry coordinates, yet how these methods represent population structure can depend strongly on sampling and construction choices. Principal component analysis (PCA), when visualized using only leading components, can obscure fine-scale structure, while nonlinear neighbor-based methods such as uniform manifold approximation and projection (UMAP) often fragment geographic and admixture gradients into discrete clusters. Here, we introduce a framework for constructing ancestry coordinates using potential of heat-diffusion for affinity-based trajectory embedding (PHATE), a diffusion-based method that preserves relationships across scales. Across synthetic benchmarks and large genomic datasets, we show that appropriate scale tuning and population balancing allow PHATE to recover geographic organization and admixture gradients. We show that incorrectly tuned neighborhood scale and population imbalance can distort the resulting representations of population structure. These results therefore provide a methodological framework and practical guidance for constructing ancestry visualizations that better reflect continuous human genetic diversity.

bioinformatics↗
bioRxiv2026

FedEpitope: Parameter-Efficient Federated Protein Language Models for Linear B-Cell Epitope Prediction

Linear B-cell epitopes are antigenic regions recognised by antibodies that play crucial roles in vaccine development, antibody production, and other therapeutic applications.Their prediction can support peptide-based vaccine candidate screening and diagnostic assay development, but the experimental data required for model development may be distributed across institutions that cannot freely pool their records. We present FedEpitope, a federated framework that adapts the ESM-2 protein language model for peptide-level linear B-cell epitope prediction using Low-Rank Adaptation (LoRA), keeping client-partitioned data local and restricting the communicated trainable model state to approximately 1.11 MB. Simulated benchmarks of FedAvg and FedProx algorithms, evaluated using homology-aware data splits of five biologically heterogeneous, organism-partitioned clients, achieve promising global test performance without centralising training data, with FedAvg outperforming all single-client models in mean AP, F1, and MCC. Additional experiments characterise post-federation personalisation, record-level differential privacy, sensitivity to controlled client heterogeneity, LoRA rank, and generalisation to the independent BCPreds and NetBCE benchmarks, demonstrating the feasibility of parameter-efficient federated protein language model adaptation for data-localised B-cell epitope prediction and the associated trade-offs among predictive performance, personalisation, privacy, communication efficiency, and distribution shift.

bioinformatics↗
bioRxiv2026

A compact motif vocabulary is differentially interpreted across human cell types

How a common genome generates diverse human cell types may depend less on distinct cis-regulatory vocabularies than on context-dependent use of a shared one. We have developed CNN-CR, an interpretable neural network that learns regulatory motifs while predicting gene expression across 116 human biosamples. Promoter regulation proves highly compressible: dozens of recurrent motif families retain most of the predictive information of much larger cell-type-specific vocabularies and generalize to cellular contexts excluded from their discovery. These motifs are conserved at their binding sites and enriched at fine-mapped regulatory variants. Cellular context acts at two levels: chromatin state gates which motif instances are available for regulatory use, whereas reader identity determines how accessible motifs are interpreted. ChIP-seq data reveal frequent shifts in reader occupancy across cell types, and genome-scale perturbations support corresponding context-dependent reader functions. Thus, regulatory diversity can emerge through differential access to and interpretation of a compact shared DNA vocabulary.

bioinformatics↗