Engineering Success

Design

To systematically evaluate ASO efficiency, in the first iteration, we designed 18 ASOs targeting different regions of the GFP mRNA (Figure 1). This allows us to determine which region is most susceptible to degradation and provides insights for optimizing future designs based on our Design-Build-Test-Learn (DBL) cycle.

• for yeast, only CDS ASOs are relevant

At this early stage, our goal was to establish the computational foundation for ASO selection. We did not yet employ a machine-learning model; instead, we constructed a feature-based ranking framework grounded in well-established parameters from the literature on antisense oligonucleotide design.

Each ASO sequence was computationally evaluated across multiple features known to influence efficacy, including:

• Perfect Watson–Crick base pairing between ASO and on-target RNA
• Hybridization energy to the target, calculated using the RIsearch program
• Off-target hybridization energy and mismatch counts, assessed across the human transcriptome and yeast genome
• mRNA secondary structure and accessibility, predicted by ViennaRNA using rolling folding windows (40 nt, step size 15)
• GC content and sequence toxicity filters (excluding motifs such as “GGGG”)
• Self-folding> and ASO–ASO dimerization potential (melting temperature between two ASOs)
• Overall thermodynamic stability (melting temperature and total hybridization energy sum)

Each feature was standardized and used to generate a composite ranking score, reflecting expected potency and selectivity.

The top-scoring ASOs were then selected for synthesis and experimental validation, ensuring that multiple regions of the GFP transcript (5′, middle, degron-proximal, and 3′ ends) were represented - capturing diverse structural and accessibility contexts.

Build

All 18 antisense oligonucleotides (ASOs) were synthesized as unmodified DNA oligos (Integrated DNA Technologies, IDT) in the standard phosphodiester form. Each ASO was designed to be 15-20 nucleotides long, complementary to distinct regions across the GFP mRNA (see table below).

Oligos were resuspended in nuclease-free water to a final stock concentration of 100 μM and stored at –20°C until use. For experimental testing, ASOs were diluted to the desired working concentration and transfected into HEK293-GFP cells using Lipofectamine 2000 (Thermo Fisher Scientific), following the manufacturer’s protocol.

See our Notebook (6.5.2025) for the complete experimental workflow and our Results Page for the final results.

Test

No significant reduction in GFP expression was detected for any of the 18 unmodified ASOs tested. Likewise, the scrambled controls and the designated positive control failed to elicit significant silencing activity. These findings suggest that the lack of knockdown may reflect limitations in ASO uptake, intracellular stability, or effective concentration under the experimental conditions (Figure 2).

Learn

Conclusion and Next Steps

The lack of detectable GFP knockdown may be attributed to one or more of the following factors:

• Inefficient ASO transfection.
• Suboptimal ASO concentration insufficient to elicit measurable gene silencing.
• Rapid ASO degradation due to the absence of chemical modifications.

To investigate these possibilities, our next iteration focused on:

• Testing transfection efficiency directly using a fluorescently labeled oligonucleotide.
  (See Results - ASO - “Fluorophore-labeled DNA lipofection assay: Successful entry of oligo into cells”)
• Introducing chemical modifications to improve ASO stability and potency.

Following consultation and troubleshooting sessions with SKIP Therapeutics company (Dr. Ariel Feiglin & Dr. Maya David-Teitelbaum) and Dr. Nofar Mor from Sheba hospital (see our Human Practice page)we were advised to incorporate 2′-O-methoxyethyl (2′MOE) and phosphorothioate (PS) modifications. These adjustments are expected to enhance nuclease resistance, binding affinity, and intracellular half-life—key factors for achieving efficient gene silencing in mammalian cells.

In parallel, we also realized that feature-based ranking alone was insufficient for capturing the complex relationships among structural, thermodynamic, and accessibility parameters; therefore, our next design iteration focused on developing a more advanced machine-learning model to guide ASO prediction.

Design

Building on the insights from the first iteration, our second design cycle aimed to integrate both experimental feedback and data-driven modeling to refine ASO prediction and selection.
The feature-based ranking approach used earlier was replaced with a more advanced machine-learning model (XGBRanker) that learns complex, non-linear relationships between sequence, structure, and activity features.

To train this model, we compiled a curated dataset of 16,931 ASO entries (13,544 (85%) for training and 3,387 (20%) for testing) after filtering out sequences lacking inhibition data or derived from non-human cell lines.
Repetitions of the same experimental ASOs (based on ID) were averaged to avoid bias, and inhibition values were log-transformed and volume-normalized to account for experimental variation.

The key newly added or refined features included:

• Experimental NN PS hybridization - replaced the RIsearch energy term with experimentally validated nearest-neighbor thermodynamics for phosphorothioate backbones.
• Sequence-derived features - entropy, AT- and GC-skew, 3′-end GC content and homo-oligo count to capture local composition effects.
• Structural stability metrics - internal fold energy, self-energy, and hairpin score providing better differentiation between stable and unstable ASOs.
• Positional features - normalized start position within the mRNA and binary indicators for exon, intron, and UTR targeting to account for transcript context.
• Treatment period - integrated as a quantitative variable to reflect exposure time during inhibition assays.

The model was trained and validated using cross-cell-line grouping to test generalizability across biological contexts.

Overall predictive correlations reached 0.43 (Pearson) / 0.38 (Spearman) on the training set and 0.40 / 0.37 on the independent test set.

Feature-importance analysis consistently ranked RNA folding and hybridization as top predictors of efficacy, confirming the central role of accessibility and thermodynamic balance.

Using this model, we generated and scored all possible 20-mer 2′-MOE ASOs, selecting top-ranked candidates for both GFP and MALAT1 transcripts in human and yeast systems.
Preference regions predicted by the model guided the final selection for synthesis and validation.

This iteration marked the transition from rule-based design to a data-driven predictive pipeline, establishing the foundation for iterative optimization across future experimental cycles.

Build

The 6 antisense oligonucleotides (ASOs) were synthesized as 2′-O-methoxyethyl phosphorothioate (2′MOE-PS) modifications (Integrated DNA Technologies, IDT). Each ASO was designed to be 20 nucleotides long, complementary to distinct regions across the GFP mRNA and MALAT1 mRNA.

ASOs were resuspended in nuclease-free water to a final stock concentration of 100 μM and stored at –20°C until use. For experimental testing, ASOs were diluted to the desired working concentration and transfected into A549-GFP and A549 cells using Lipofectamine 2000 (Thermo Fisher Scientific), following the manufacturer’s protocol.

See our Notebook (26.8.2025) for the complete experimental workflow and our Results Page for the final results.

ASOs sequence are presented in the tables below.

GFP ASOs:

MALAT1 ASOs:

Test

We transfected A549 cells with the chemically modified MALAT1-targeting ASO . After incubation, total RNA was extracted, reverse-transcribed to cDNA, and MALAT1 expression was quantified by RT–qPCR relative to a housekeeping gene (β-actin/ACTB) (Figure 4).

Transfection of A549 cells with the chemically modified reference MALAT1 ASO resulted in a strong and consistent reduction of MALAT1 expression compared to untreated, Lipofectamine-only, and scrambled controls (Figure 5). Notably, ASO2 and ASO3, computationally designed using our modeling pipeline, achieved even greater knockdown efficiency than the commercial reference ASO.

We transfected A549-GFP cells and evaluate their knockdown efficiency relative to the validated control. To validate GFP reduction at both the mRNA and protein levels, we employed RT–qPCR and flow cytometry, as shown in Figure 6.

Flow cytometry and RT–qPCR analyses consistently demonstrated that GFP-targeting ASOs effectively reduced GFP expression in A549-GFP cells (Figure 7). Flow cytometry confirmed a decrease in GFP protein levels following transfection with the reference GFP ASO as well as our computationally designed candidates, with ASO2 and ASO3 producing the most pronounced knockdown compared to scrambled and Lipofectamine-only controls. Complementary RT–qPCR analysis of GFP mRNA levels at 24 h and 48 h post-transfection further validated these findings, showing significant reductions in GFP transcripts for all GFP-targeting ASOs and the reference ASO, relative to controls . Notably, the strongest knockdown was consistently observed with ASO2, ASO3, and the reference ASO, highlighting both the reliability of our system and the capacity of computationally designed ASOs to achieve potent silencing at both the mRNA and protein levels. Together, these results provide robust evidence for the sequence-specific activity of our GFP ASOs and establish a solid foundation for further optimization.

Learn

GFP-targeting ASOs designed with our computational pipeline and with chemical modifications achieved efficient and sequence-specific knockdown of GFP in A549 cells, as confirmed at both the mRNA and protein levels. ASO2 and ASO3 showed the strongest activity, performing comparably or better than the reference ASO, thereby validating the effectiveness of our design approach.

In addition, all of our computationally designed ASOs against MALAT1 demonstrated effective silencing, with two of them (ASO2 and ASO3) achieving even stronger silencing of MALAT1 than the commercial reference.

The experimental results highlighted the importance of RNA folding as one of the leading features in our predictive model, which was reflected in the strong performance of ASOs designed based on this parameter.

Furthermore, the strong performance of our top-ranked ASOs compared to the reference, supports the predictive accuracy of the XGBRanker model and demonstrates that the machine-learning–based scoring successfully prioritizes biologically potent candidates.

In future iterations, we plan to place greater emphasis on folding and accessibility features and explore additional structural parameters to further improve prediction accuracy.

Design

Following the performance of the second iteration, we revisited both our training data and model design to improve predictive precision and biological relevance.

The updated ML model (Model 2) introduced major corrections in data preprocessing, feature selection, and evaluation metrics, aimed at overcoming overfitting and increasing generalizability to new cell lines such as A549.

Model corrections and evaluation updates

In the previous version, model evaluation relied on the “average top” metric - a mean of inhibition values among top-ranked ASOs. While useful for an overview, it did not accurately represent the model’s ranking quality or its ability to prioritize truly effective sequences.

In this iteration, we adopted more informative ranking-based evaluation metrics: precision@K and NDCG (Normalized Discounted Cumulative Gain). These measures evaluate how well the model identifies the most effective ASOs within the top-ranked subset. We specifically focused on precision@50, precision@100, and NDCG@200, which better reflect the model’s ability to prioritize the strongest candidates at the very top of the ranking rather than evenly optimizing across all ASOs.

All evaluations were performed on unseen cell lines: data were grouped by cell line and we used a leave one cell line out - group-based validation scheme, always training on the remaining lines and testing on the held-out line. This assesses true generalization across biological contexts and avoids inflating performance via within-cell-line train/test splits.

To further improve reliability, the model was retrained on a carefully selected subset of four biologically relevant cell lines - A431, SK-MEL-28, KARPAS-229, and MM1.R.
These lines were chosen because they better represent cancer biology relevant to our system (melanoma and epithelial origins), while smaller or less consistent datasets, such as liver cell lines, were excluded to reduce noise and bias.

Data normalization and inhibition correction

To standardize inhibition values across experiments, we implemented a multi-stage correction that accounted for:

• Log transformation (reducing skew and scaling inhibition values).
• Volume normalization, correcting for experimental well size and transfection conditions.
• Treatment-period adjustment, modeling inhibition decay as a linear function over exposure time.

These corrections reduced experimental noise and allowed the model to capture subtler structure activity relationships.

Feature expansion and refined selection

Beyond the features from iteration #2, several new biologically meaningful features were introduced and validated through cross-cell-line feature selection:

• Modification minimum distance to 3′ end and modification skew index - capturing positional effects of 2′MOE chemical modifications.
• CAI_score_global_CDS - codon adaptation index across the target mRNA to reflect translational context.
• Sense_avg_accessibility - enhanced folding descriptors integrating accessibility.
• RNaseH1_Krel_score_R7_krel - proxy for RNase H1 cleavage propensity.
• off_target.top200.cutoff600.premRNA_TPM - refined off-target filtering based on expression-weighted transcripts.

Following cross-validation, feature selection prioritized combinations that maximized NDCG@200 on unseen cell lines, improving the model’s ability to generalize beyond the training set.

Performance overview

The revised model improved correlations across several lines, with Spearman correlations up to 0.48-0.55, demonstrating closer alignment between predicted and observed inhibition trends.
Although precision at small K values (e.g., top 50) remained moderate, the new configuration consistently detected top-performing ASOs without missing high-efficacy candidates entirely, a limitation of previous models.

These improvements provided a stronger computational foundation for the next ASO synthesis round.
The top predictions were used to design four new 2′-MOE–PS ASOs targeting distinct regions of MALAT1, selected for low off-target potential and optimal folding, GC content, and RNase H motif scores.

Build

The 4 antisense oligonucleotides (ASOs) were synthesized as 2′-O-methoxyethyl phosphorothioate (2′MOE-PS) modifications (Integrated DNA Technologies, IDT). Each ASO was designed to be 20 nucleotides long, complementary to distinct regions across the MALAT1 mRNA.

ASOs were resuspended in nuclease-free water to a final stock concentration of 100 μM and stored at –20°C until use. For experimental testing, ASOs were diluted to the desired working concentration and transfected into A549 cells using Lipofectamine 2000 (Thermo Fisher Scientific), following the manufacturer’s protocol.

See our Notebook (3.10.2025) for the complete experimental workflow and our Results Page for the final results.

ASOs sequence are presented in the table below.

Test

A549 cells were transfected with each ASO, and MALAT1 expression was quantified by RT–qPCR 24 hours post-transfection (Figure 8).

Results are shown below.

Learn

The third iteration provided valuable insights into the model’s behavior and the impact of newly introduced features.
While the four ASOs designed in this round did not show a clear improvement compared to previous versions, their overall performance remained comparable to or better than the reference ASO, reaffirming the robustness of our modeling approach and the consistency of our design principles.

Compared to the previous model, this version placed greater emphasis on features related to chemical modification and RNA accessibility. The results indicate that the model maintained its predictive capacity despite substantial adjustments to data preprocessing, feature selection, and evaluation criteria.
Although these additions helped preserve predictive stability, they also revealed that further refinement of feature weighting may be required.

These observations provided important feedback for balancing feature contributions and highlighted the importance of integrating a broader range of structural and biochemical contexts.
Moving forward, we plan to recalibrate these features to better capture modification- and structure-dependent effects across diverse cell types.
Overall, this iteration enhanced our understanding of the model’s strengths and limitations and provided clear directions for optimizing feature weighting and biological interpretation in the next design cycle.

Design

Our goal was to create a cross-species GFP reporter integrated into the yeast genome for rapid ASO testing. The construct included a constitutive promoter driving a human-codon-optimized GFP followed by a terminator sequence. Integration was designed to occur at the ADE1 locus of S. cerevisiae W303, using matching homology arms for precise genomic insertion. To mirror our mammalian model, the GFP coding sequence was amplified from a HEK293 GFP plasmid. The overall assembly plan consisted of Gibson-ready fragments containing the promoter, GFP, terminator, selection cassette, and ADE1 homology arms with overlapping ends suitable for seamless assembly.

Build

The construct was assembled using Gibson Assembly, combining a linearized plasmid backbone with PCR-amplified inserts. Following assembly, colonies were screened by PCR across junction sites, optionally verified by restriction digest, and further confirmed by Sanger sequencing of the GFP region and integration borders.

Test

After transforming the plasmid into yeast and selecting with hygromycin B, fluorescence was confirmed under a microscope. However, plate reader analysis revealed a very low GFP signal, similar to the negative control.

Learn

We concluded that despite correct integration and transcription, codon usage and promoter strength were likely limiting factors for translation efficiency and protein accumulation in yeast.

Design

We redesigned the construct with three major changes:

1. Codon optimization: The first 8 codons of the GFP gene were optimized for yeast expression.
2. Stronger promoter: We replaced the original promoter with GAL10, a strong galactose-inducible promoter shown to work even without induction in certain strains [3].
3. Degron addition: To enable controlled degradation in future ASO experiments, we added an auxin-inducible degron downstream of GFP.

Build

The new GFP-degron fusion was synthesized and assembled into a linearized vector backbone. Colony PCR, minipreps, and sequencing confirmed the correct construct. We linearized the plasmid with SrfI and integrated it into yeast as before.

Test

Four colonies were cultured in glucose and galactose media. In galactose, all showed substantially higher GFP expression, confirmed by plate reader analysis (see Figure 11).

Learn

Our results validated the importance of host-specific optimization (codon usage + regulatory elements) and demonstrated how the engineering cycle can be applied in synthetic biology to systematically troubleshoot and enhance construct performance.