• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Monday, November 16, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Biology

Generative Model Feedback Prioritizes RNA Aptamers That Bind Isolated Mitochondria

by
November 16, 2026
in Biology
Reading Time: 5 mins read
0
Generative Model Feedback Prioritizes RNA Aptamers That Bind Isolated Mitochondria

Generative Model Feedback Prioritizes RNA Aptamers That Bind Isolated Mitochondria

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Computational biologists have developed a feedback-guided generative framework capable of learning from enriched experimental sequence pools to design and prioritize synthetic RNA aptamers targeting mitochondrial membrane preparations. In laboratory evaluations detailed by researchers at Hunan University, the generative architecture produced candidate RNA sequences that achieved nanomolar-scale apparent binding affinities against crude mitochondrial membrane-protein preparations while maintaining distinct sequence nonidentity relative to the initial training pool. RNA aptamers function by folding into sequence-dependent three-dimensional architectures that bind specific biochemical targets. While systematic evolution of ligands by exponential enrichment, commonly known as SELEX, serves as the standard experimental pathway for identifying functional nucleic acid binders from massive stochastic libraries, exploring sequence space beyond the physical sequences recovered in initial selection rounds remains an active bioengineering challenge.

To navigate this design space, the investigators constructed Feedback-MitAptamerGAN, an integrated computational platform coupling a Wasserstein generative adversarial network with gradient penalty to a dedicated external classification module. The generative pipeline operates on training data derived from an empirical starting library of ten million distinct RNA chains that underwent three progressive enrichment rounds against intact mitochondria isolated from HEK293 human embryonic kidney cells. High-throughput sequencing and deduplication of the third-round selection material yielded twenty thousand unique RNA sequences. To provide characterized structural reference points alongside this newly enriched pool, the team incorporated three published mitochondrial aptamers identified in earlier literature, comprising Mitomer1, Mitomer2, and a shortened Mitomer2 derivative originally isolated against rat-liver mitochondria using fluorinated RNA. Together, these sources provided an initial positive training resource of 20,003 sequence entries before cross-source deduplication.

The architecture guides sequence exploration through a multi-stage evaluation mechanism that integrates deep sequence representations with biophysical thermodynamic features. The evaluation network combines structural sequence embeddings extracted from a fine-tuned ERNIE-RNA language model with explicit secondary structural descriptors and minimum free energy calculations computed via RNAfold from ViennaRNA package 2.6.4. Within the evaluator pipeline, the first six layers of ERNIE-RNA were held frozen while layers seven through twelve were fine-tuned, utilizing the 768-dimensional layer-twelve CLS representation. Structural characterization executed at 37 degrees Celsius supplied 5,119 thermodynamic and base-pairing features, yielding a combined 5,887-dimensional feature vector fed into a multilayer dense classification head. The classifier was trained against continuous source-informed targets, where established literature aptamers received an assigned weight of 1.00 and filtered sequences from mitochondrial SELEX rounds six, seven, and eight contributed continuous targets of 0.85, 0.90, and 0.95, respectively.

Adversarial training balances candidate generation with directed selection through scheduled feedback loops that update the generator training pool during optimization. The generator processes a 100-dimensional Gaussian latent vector through three dense layers of 256, 512, and 1,024 units with rectified linear activation and batch normalization, producing a 70 by 5 output array representing the nucleotide alphabet alongside an end-of-sequence padding symbol. The corresponding critic network processes 350-dimensional flattened sequence representations across dense layers of 1,024, 512, and 256 units with leaky rectified linear units. Feedback introduces synthetic candidates scoring above an operational 0.8 evaluator threshold back into the generative training distribution. System parameter tuning determined that initiating feedback after seventy-five percent of preset training iterations, generating 4,000 synthetic candidates per step, and selecting the top three percent to replace up to seventy-five percent of the positive training resource yielded an optimized candidate pool exhibiting an average evaluator score of 0.91.

Sequence analysis of 205 generated candidates confirmed that the computational system proposed structural innovations rather than simple memorization of experimental inputs. Normalized pairwise alignment similarity against known mitochondrial references averaged 0.68 with a standard deviation of 0.12, contrasted with 0.41 plus or minus 0.09 against computational background RNA pools. Predicted secondary structure folding analyses revealed that inspected generated sequences exhibited minimum free energies ranging from minus 18.5 to minus 24.3 kilocalories per mole, closely mirroring the calculated folding values of minus 21.2 kilocalories per mole for Mitomer2 and minus 19.8 kilocalories per mole for its shortened derivative.

Experimental validation of the generative design strategy was conducted through a rigorous six-round SELEX campaign comparing physical libraries synthesized from different model variations. Researchers tested five distinct libraries across three matched experimental batches, with each independent selection run measured in technical qPCR triplicates. The tested libraries comprised the complete feedback model, three systematic ablations lacking feedback updates, ERNIE-RNA representations, or RNAfold features, and a 40-nucleotide random RNA control library. Starting libraries contained 10,000 sequences at a concentration of 1 nanomolar incubated with isolated HEK293 mitochondria, followed by stringent low-salt washing and high-salt elution. Amplification monitored via SYBR Green quantitative PCR targeted shared linker sequences to track the cycle threshold across successive selection cycles.

The complete-model library consistently outperformed all ablated variants and control pools, sustaining the lowest mean cycle threshold across all six selection rounds. From an initial round-one mean cycle threshold of 28.60 plus or minus 0.12, the complete-model library decreased to 14.37 plus or minus 0.20 at round six, representing an overall reduction of 14.23 cycles across the three independent experimental batches. In contrast, the ablated library lacking feedback achieved a cycle threshold decrease of 12.76 cycles, the variant lacking RNAfold dropped by 12.69 cycles, and the ablation omitting ERNIE-RNA representations decreased by 11.44 cycles, while the random library shifted by only 3.16 cycles. At round six, the complete-model library remained 2.58 cycles below the no-feedback condition, 2.55 cycles below the no-RNAfold condition, and 4.08 cycles below the model lacking ERNIE-RNA, reflecting a full-model trajectory slope of minus 2.877 cycle units per round.

Individual binding affinities were characterized using bio-layer interferometry on an automated Octet platform using crude mitochondrial membrane-protein fractions extracted from HEK293 cells. Twenty generated sequences were stratified into two score-selected panels of ten candidates each, screened to avoid sequences exceeding ninety-five percent mutual identity or falling within the marginal 0.78 to 0.82 score transition band. Five-prime biotinylated aptamers were immobilized onto streptavidin biosensors and exposed to six analyte concentrations referenced to a mixture-level nominal molecular weight convention of 42 kilodaltons. The ten high-scoring candidates demonstrated mean apparent dissociation constants spanning 50.87 to 71.70 nominal nanomolar, achieving a group mean of 59.06 plus or minus 6.42 nominal nanomolar. Conversely, the ten low-scoring candidates displayed apparent dissociation constants ranging from 2,269.80 to 15,682.13 nominal nanomolar, averaging 8,330.37 nominal nanomolar.

Reference control measurements placed these candidate affinities alongside established benchmarks, with literature controls Mitomer1 and Mitomer2 registering apparent dissociation constants of 50.13 plus or minus 1.42 and 59.63 plus or minus 0.12 nominal nanomolar under identical experimental conditions, while a negative control sequence yielded no measurable kinetic fit. Across the twenty evaluated synthetic sequences, model scores exhibited a strong inverse correlation with measured apparent dissociation constants, yielding a Spearman rank correlation coefficient of minus 0.9955 with a bootstrap 95 percent percentile interval between minus 0.9955 and minus 0.9910. Biochemical pull-down experiments and mass spectrometry confirmed functional protein capture from mitochondrial total lysates, identifying voltage-dependent anion channel 1, or VDAC1, with seven unique peptides and 30.4 percent sequence coverage in material recovered by representative candidate MitApt-A02.

While these experimental findings demonstrate the utility of coupling structural evaluators with generative networks, the authors note clear boundaries regarding study scope and interpretation. The observed kinetic parameters reflect apparent binding behavior against complex, detergent-solubilized membrane-protein mixtures under an operational 42-kilodalton nominal concentration model rather than absolute single-target dissociation constants. Furthermore, candidate prioritization relied on supervised reference weighting and checkpoint selection, meaning that broader generalization across unrelated RNA families or distinct mitochondrial import mechanisms requires subsequent target-specific verification.

Subject of Research: Machine learning generative design of mitochondrial RNA aptamers using SELEX feedback

Article Title: Feedback-MitAptamerGAN for SELEX-informed generation and prioritization of mitochondrial RNA aptamer candidates

Article References: Chen, Z., Long, Y., & Ji, Z. (2026). Feedback-MitAptamerGAN for SELEX-informed generation and prioritization of mitochondrial RNA aptamer candidates. Frontiers in Bioengineering and Biotechnology, 14, 2016105. https://doi.org/10.3389/fbioe.2026.2016105

Image Credits: AI Generated

DOI: 10.3389/fbioe.2026.2016105

Keywords: Feedback-MitAptamerGAN, RNA aptamers, mitochondria, generative adversarial networks, SELEX, ERNIE-RNA, ViennaRNA, RNAfold, bio-layer interferometry, VDAC1, HEK293, synthetic biology, apparent dissociation constant, cycle threshold, aptamer discovery, biomolecular design, computational biology

News Source: Drew Townsend. (November 16, 2026). Generative Model Feedback Prioritizes RNA Aptamers That Bind Isolated Mitochondria. Scienmag.

Tags: apparent dissociation constantaptamer discoverybio-layer interferometrybiomolecular designcomputational biologycycle thresholdERNIE-RNAFeedback-MitAptamerGANgenerative adversarial networksHEK293mitochondriaRNA aptamersRNAfoldSELEXsynthetic biologyVDAC1ViennaRNA
Share12Tweet7Share2ShareShareShare1

Related Posts

Automated Platform Maps Genomic Hotspots in CHO Cells for Antibody Production

Automated Platform Maps Genomic Hotspots in CHO Cells for Antibody Production

November 16, 2026
Complete Pig Genomes Uncover Y-Chromosome Centromere Shifting and Functional Structural Variants

Complete Pig Genomes Uncover Y-Chromosome Centromere Shifting and Functional Structural Variants

November 16, 2026

Early Life Obesity Programs Mouse T Cells to Impair Vaccine Immunity

November 16, 2026

Engineered Adeno-Associated Virus Enables Efficient In Vivo Immune Cell Programming

November 16, 2026

POPULAR NEWS

  • Interferon-Induced LAP3 Promotes Mouse Embryo Resorption by Inhibiting Trophoblast Mitophagy

    Interferon-Induced LAP3 Promotes Mouse Embryo Resorption by Inhibiting Trophoblast Mitophagy

    29 shares
    Share 12 Tweet 7
  • Systematic Review Outlines Artificial Intelligence Strategies for Designing Functional Taste Peptides

    29 shares
    Share 12 Tweet 7
  • Biliary Fascioliasis Mimics Bile Duct Stones: Insights from Four Clinical Cases

    29 shares
    Share 12 Tweet 7
  • Physical Fitness Declined Among Belgian Children Between 2004 and 2022

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Children With Nephrotic Syndrome Exhibit Severe Disruption Across Gut Microbial Kingdoms

Automated Platform Maps Genomic Hotspots in CHO Cells for Antibody Production

Low-Molecular-Weight Heparin Linked to Lower Clot Odds in Blunt Trauma

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.