Biology subjects

Herron, L.

Publications and source records attributed to Herron, L..

6 recordsLinked to original sources

bioRxiv2026

Detecting and quantifying overparametrization in RNA language models with REDIAL

While RNA language models (LMs) have served as foundation models (FMs) to advanced structural prediction, their evaluation relies heavily on supervised downstream tasks. Such tasks can often mask FM inefficiencies and reflect downstream training set memorization. To address this, here we introduce REDIAL (RNA Embedding perturbation Diagnostics for Language models), a zero-shot, unsupervised framework designed to extract coevolutionary signals directly from the high-dimensional latent spaces of RNA language models. By applying REDIAL, we uncover stark, layer-wise disparities in how popular RNA language models (LMs) internalize structural constraints through a layer-wise dissection and ablation study. Our results showed how such layerwise behavior deviates from protein LMs and is related to design flaws in the architectures. Specifically, we show that current RNA LMs are severely overparameterized relative to the limited sequence diversity of available RNA databases, leading to profound parameter inefficiency and overfitting. Furthermore, we establish that structure-guided pretraining fundamentally improves the signal-to-noise ratio of learned coevolutionary couplings compared to sequence-only baselines. Ultimately, this unsupervised evaluation paradigm exposes critical flaws in current parameter scaling strategies and provides a rigorous diagnostic benchmark to guide the development of more efficient, generalizable foundation models for RNA therapeutics and de novo design.

bioinformatics↗

bioRxiv2026

Ab initio prediction of RNA structure ensembles with RNAnneal

RNA utilizes three-dimensional structure in addition to sequence to carry out diverse functions in gene expression and disease. Much like well-folded proteins, RNAs adopt specific three-dimensional structures to carry out their function. Yet comparatively few RNA structures have been solved by atomic resolution structural techniques, in part because unlike structured proteins, RNAs fold into heterogeneous ensembles of interconverting structures that pose a challenge for high-resolution structure probing methods. In this work, we introduce RNAnneal as a method for RNA structural ensemble prediction that seamlessly integrates generative deep learning with statistical physics and molecular dynamics modeling. Given the primary sequence, RNAnneal uses ab inito (i.e., first principles) modeling to sample an ensemble of 3D structures, which, in turn, are used to train an ensemble of unsupervised deep learning models. The RNAnneal score, representing the consensus of the deep learning models, is then used to evaluate the 3D structures. We evaluated RNAnneal structures against 16 experimentally-resolved conformations (ERCs) of riboswitch RNAs and found that pseudoknot-free (PK-free) ERCs were well-reproduced by RNAnneal, with clear avenues for improving performance even on PK-comprising structures. Furthermore, we found that the RNAnneal score outperforms the Rosetta score and a state-of-the-art RNA forcefield on the task of classifying ERCs from decoys. We then introduce the interaction entropy as a measure of conformational heterogeneity within an ensemble and use it to assess our predictions. RNAnneal thus provides a generalizable framework for predicting RNA structural ensembles that will accelerate RNA-targeted drug discovery and the design of functional RNA molecules.

biophysics↗

bioRxiv2025

Regulatable In Vivo Gene Expression via Adaptamers

Precise, reversible control of transgene expression is essential for safe and durable gene therapy, yet current inducible systems remain difficult to translate in vivo due to large size, limited induction duration, and dependence on immunogenic regulators. Here we present the adaptamer (ADAR modulatable aptamer), a compact (<120 bp) RNA switch that couples an FDA-approved small-molecule-responsive aptamer with endogenous ADAR-mediated RNA editing to regulate expression. Ligand binding stabilizes a double-stranded RNA structure that recruits ADAR to convert a stop codon into a sense codon, restoring downstream protein translation. This cleavage-free, post-transcriptional mechanism enables precise, small-molecule-dependent modulation without exogenous protein machinery. We demonstrate that adaptamers are highly functional across multiple cell lines, including human T-cells. In mice, AAV-delivered adaptamer-controlled FGF21 expression induced metabolic remodeling, significantly increasing energy expenditure and reversing obesity. This minimal, programmable system offers a clinically compatible approach for tunable genetic medicines and safe deployment of pleiotropic and dose-limited proteins.

synthetic biology↗

bioRxiv2025

DiffDock-Glide: a hybrid physics-based and data-driven approach to molecular docking

Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into protein binding sites. But DiffDock struggles to compete with conventional docking methods, especially for targets outside its training set. We develop a hybrid model called DiffDock-Glide which addresses some shortcomings of deep learning docking methods: it uses a modified generative process to generate samples within a binding pocket and the confidence model is replaced with Glides post-docking minimization pipeline. We evaluate DiffDock-Glide on the Posebusters dataset and show improved sampling of near-native poses, especially for sequences without homologues in the training set. We also evaluate DiffDock-Glides performance in virtual screening compounds from the DUD-E dataset against receptor structures generated by AlphaFold2 and report enrichment values that broadly surpass those from traditional Glide.

molecular biology↗

bioRxiv2023

Phenomenological modeling reveals the emergent simplicity of host-associated microbiomes

A key step towards rational microbiome engineering is in silico sampling of realistic microbial communities that correspond to desired host phenotypes, and vice versa. This remains challenging due to a lack of generative models that simultaneously capture compositions of host-associated microbiomes and host phenotypes. To that end, we present a generative model based on the mechanistic consumer/resource (C/R) framework. In the model, variation in microbial ecosystem composition arises due to differences in the availability of effective resources (inferred latent variables) while species resource preferences remain conserved. The same latent variables are used to model phenotypic states of hosts. In silico microbiomes generated by our model accurately reproduce universal and dataset-specific statistics of bacterial communities. The model allows us to address three salient questions in host-associated microbial ecologies: (1) which host phenotypes maximally constrain the composition of the host-associated microbiomes? (2) how context-specific are phenotype/microbiome associations, and (3) what are plausible microbiome compositions that correspond to desired host phenotypes? Our approach aids the analysis and design of microbial communities associated with host phenotypes of interest.

systems biology↗

bioRxiv2022

SMbiot: A Shared Latent Model for Microbiomes and their Hosts

The collective nature of the variation in host associated microbial communities suggest that they exhibit low dimensional characteristics. To identify these lower dimensional descriptors, we propose SMbiot (pronounced SIM BY OT): a Shared Latent Model for Microbiomes and their hosts. In SMbiot, latent variables embed host-specific microbial communities in a lower dimensional space and the corresponding features reflect controlling axes that dictate community compositions. Using data from different animal hosts, organ sites, and microbial kingdoms of life, we show that SMbiot identifies a small number of host-specific latent variables that accurately capture the compositional variation in host associated microbial communities. By using the same latents to describe hosts phenotypic states and the host-associated microbiomes, we show that the latent space embedding is informed by host physiology as well as the associated microbiomes. Importantly, SMbiot enables the quantification of host phenotypic differences associated with altered microbial community compositions in a host-specific manner, underscoring the context specificity of host-microbiome associations. SMbiot can also predict missing host metadata or microbial community compositions. This way, SMbiot is a concise quantitative method to understand the low dimensional collective behavior of host-associated microbiomes.

systems biology↗