The goal of this experimental cycle was to establish a reliable method for delivering siRNAs into MEC-1 cells and to evaluate whether this approach could achieve functional gene silencing of BCL-2. Given that MEC-1 is a suspension cell line with known sensitivity to lipid-based transfection reagents, particular attention was placed on identifying conditions that balance efficient uptake with minimal cytotoxicity.

• MEC-1 cells were cultured until viability reached ≥90% and then were seeded into 24-well plates.
• siRNAs were prepared at working concentrations.
• siRNAs complexed with INTERFERin in Opti-MEM and incubated to allow complex formation.
• Experimental layout included:
• Concentration gradient: 5–50 nM.
• Controls: untreated, INTERFERin-only, scrambled siRNA, siRNA-only.
• Plates were incubated at 37 °C, 5% CO₂ for up to 96 h.

Further procedural details are provided in the Lab Notebook.

• Samples were collected at 24, 48, 72, and 96 h for:
• Cell viability and cell proliferation capacity assays (Trypan Blue)
• Cy3 fluorescence uptake
• Microscopic observation of morphology and culture homogeneity

Transfection of MEC-1 cells with BCL-2 siRNAs revealed an inhibitory effect on cellular proliferation. At early timepoints (24 h), both proliferation and viability were reduced across all conditions, with the strongest inhibition observed at higher concentrations (50 nM) and in the INTERFERin-only group. However, at later timepoints, cells exhibited partial recovery of growth and viability, indicating an adaptive response following initial treatment stress, which appeared to be primarily associated with the transfection reagent rather than the siRNA itself.

These effects were largely cytostatic, reflecting transient stress responses rather than overt cytotoxicity. Over time, cells demonstrated progressive recovery: by 72 h, morphology and viability had stabilized, and by 96 h, proliferation in all groups, including high-concentration and INTERFERin-only conditions, surpassed that of the control. Simultaneously, indicates that MEC-1 cells adapt to INTERFERin exposure and restore normal growth dynamics following the initial stress phase. From this cycle, we established that concentrations in the 5–20 nM range provide a balance between effective transfection and minimal long-term impact on cell viability, forming the basis for subsequent functional silencing studies, with 20 nM showing the most favorable results. However, the cytotoxicity observed at higher INTERFERin volumes prompted a dedicated titration experiment to optimize reagent-to-cell and siRNA ratios and enhance overall cellular tolerance.

This cycle aimed to determine the optimal concentration range of INTERFERin that allows efficient siRNA delivery while minimizing cytotoxicity. MEC-1 cells were exposed to gradient volumes of INTERFERin. Cell viability, growth, and morphology were assessed over time to distinguish transient stress effects from sustained toxicity.

• MEC-1 cells were counted and plated in 24-well plates at uniform density.
• Master mixes are prepared by combining INTERFERin with Opti-MEM and incubating for 15 minutes to allow complex formation.
• Gradient of INTERFERin concentrations (μL) tested in parallel across technical replicates.
• 1
• 2
• 3
• 5
• Controls: untreated cells.
• Plates were incubated at 37 °C, 5% CO₂ for 48 h.

Further procedural details are provided in the Lab Notebook.

• Cell viability was measured at 24 h and 48 h across all INTERFERin concentrations.
• Cell proliferation trends were recorded relative to controls.
• Microscopic assessment for stress indicators (aggregation, clustering, reduced viable cell numbers).

Treatment of MEC-1 cells with INTERFERin demonstrated a dose-dependent inhibitory effect on cellular proliferation. Smaller volumes were associated primarily with cytostatic outcomes consistent with a temporary arrest of cell division, but higher volumes (5 μL) showed cytotoxicity. When comparing cultures seeded at different densities, the inhibitory effect was slightly less pronounced at 300,000 cells/mL than at 150,000 cells/mL, likely due to reduced effective dose per cell and the presence of survival-supporting microenvironmental factors.

INTERFERin exhibits cytotoxic/cytostatic effects at high doses. Its main negative impact at larger volumes (3–5 µL) is an inducible and reversible arrest of the cell cycle, rather than induction of cell death. This suggests that the observed effects are likely non-permanent and that MEC-1 cells can recover once the stress is alleviated.

From this cycle, we established an optimal working range of INTERFERin volumes that balances efficient transfection with minimal cytotoxicity and an optimal cell seeding setup, providing a reliable basis for downstream siRNA delivery experiments.

The objective of this second repetition was to validate and refine the conditions established in the initial siRNA transfection experiment. As MEC-1 cells are highly sensitive to lipid-based transfection reagents, we aimed to confirm the reproducibility of siRNA uptake and BCL2 gene silencing under the optimized parameters identified previously. Repeating the experiment allowed us to assess consistency across independent runs, strengthen the reliability of our observations, and further define the balance between efficient delivery and minimal off-target cytotoxic effects.

• MEC-1 cells were seeded into 24-well plates (200 µL/well).
• Eight experimental conditions were prepared with Master Mixes (Opti-MEM, siRNA at varying concentrations, including Cy3-labeled constructs, and INTERFERin).
• Complexes were incubated for 15 min at room temperature before addition to cells.
• Two plates were prepared for each time point (6 h, 24 h, 48 h):
• Plate A: RNA extraction + cellular uptake
• Plate B: protein extraction
• Medium was supplemented at 6 h post-transfection to reach a final volume of 1 mL/well.

Further procedural details are provided in the Lab Notebook.

• Cell viability & proliferation were measured at 6 h, 24 h, and 48 h using Trypan Blue staining and Neubauer chamber counting.
• Cell morphology was assessed by microscopy at each time point, with representative images captured.
• siRNA uptake was quantified using Cy3 fluorescence from one-third of the samples at each time point.
• RNA analysis: remaining cells from RNA plates were pelleted for RNA extraction, and later processed for qPCR to measure BCL2 expression relative to reference genes.
• Protein analysis: 300,000 cells per condition collected from protein plates, pelleted for downstream protein extraction and BCL2 assessment.

All conditions were tested in replicates to ensure reproducibility.

Transfection of MEC-1 cells with BCL2 siRNAs revealed a concentration-dependent pattern of uptake and processing. At 20 nM, a distinct fluorescence peak was observed at 6 h, followed by a gradual decline and stabilization by 24–48 h. This dynamic profile indicates efficient internalization and intracellular processing of the siRNA. In contrast, at 50 nM, fluorescence increased continuously over time, suggesting overloading of the delivery system and sequestration of siRNAs in vesicles, accompanied by mild growth inhibition. Lower concentrations (5–10 nM), scrambled siRNA, and INTERFERin-only controls exhibited no significant fluorescence signal. Importantly, cell viability remained consistently high (>85–90%) across all conditions, demonstrating that the treatments influenced proliferation dynamics rather than inducing cytotoxicity.

Our observations demonstrated that higher siRNA concentrations correlated with increased uptake efficiency, confirming successful delivery into MEC-1 cells. Among the range of concentrations tested, the 20 nM and 50 nM treatments produced similar cellular outcomes, including slightly reduced viability and evident cell cycle arrest. Notably, the 20 nM concentration, representing a milder treatment, achieved comparable functional results to 50 nM, suggesting it may serve as a more optimal and less cytotoxic condition for subsequent experiments. These findings underscore the importance of balancing delivery efficiency with cellular tolerance, guiding further optimization of transfection parameters.

Based on the findings of the previous engineering cycle, we firstly moved on to conduct literature review on the features that affect siRNA silencing efficacy, with the goal of applying feature engineering. We hypothesized that the addition of engineered (pre-calculated) sequence features to the model input would result in the improvement of the model performance, since the model would be able to make predictions based on more features that previously could not calculate on its own.

Through extensive literature review, we identified 100 thermodynamic and structural features to enrich our training dataset. All feature values are rounded to three decimal places. The dropdown button below lists these features with descriptions of their biological significance.

For the model fitting, we used 90% of the training dataset for model training and the rest 10% for the validation set.

Afterwards, we proceeded to search for more advanced machine learning models to employ for our task. We deemed that by using a model that adapts better to the characteristics of our task (multiple calculated features, medium-sized dataset, potential non-linear relationships), the model performance would be enhanced. After literature review, we selected LightGBM (Light Gradient Boosting Machine), an efficient implementation of gradient boosting decision trees.

Unlike linear models, LightGBM captures non-linear feature interactions, where silencing depends on the complex interplay of thermodynamic stability, accessibility, and positional effects rather than simple linear rules. LightGBM builds an ensemble of decision trees sequentially, with each tree correcting errors from previous ones. Its leaf-wise growth strategy expands branches that maximize error reduction, enabling the model to learn highly specific patterns. Efficiency optimizations including histogram-based feature binning, gradient-based sampling, and exclusive feature bundling make it faster and more memory-efficient than other gradient boosting implementations.

The advantages for our dataset are:

• Efficiently handles medium-sized datasets (~4,100 siRNAs) without requiring the massive training sets needed for deep learning.
• Naturally processes correlated and position-specific features (DG_pos1..18, DH_pos1..18, duplex energies) to identify non-linear patterns that simpler models cannot capture.

After completing the “Design” step of this engineering cycle, we moved on to write code for both feature engineering and LightGBM model development. Full documentation of these aspects of our modeling are available in our GitLab Repository.

Full documentation for the final training dataset creation can be found in our GitLab Repository.

Full documentation for LightGBM Model development can be found in our GitLab Repository.

The next step after we built our enriched training dataset and trained the model was to evaluate the model performance. We used the Mean Squared Error (MSE) metric, and added the Pearson Correlation Coefficient (R) metric, in order to measure the linear correlation between the predicted and the actual silencing efficacies. The table below shows the metrics values.

Compared to the previous models we implemented, we observed a clear improvement in predictive accuracy, as the Mean Squared Error (MSE) decreased to 0,0529. A lower MSE indicates that the predicted siRNA efficacies are, on average, closer to the experimental values, meaning the model has become more precise.

Additionally, the Pearson correlation coefficient (R = 0.5454) shows that there is a moderate positive linear correlation between our predictions and the true values. In practice, this means that the model successfully captures the trend of siRNA activity: when experimental efficacy is high, predicted efficacy also tends to be high, and vice versa. Although not perfect, this correlation confirms that the model is learning meaningful biological patterns rather than random noise.

At this stage, our LightGBM model already provided satisfactory results, showing both a reduced prediction error and a meaningful correlation with experimental values. We could have stopped here and considered this as our final predictive tool. However, we wanted to go the extra mile and explore whether a different machine learning approach, based on a distinct underlying mechanism, could further improve performance and capture additional patterns in the data.

To achieve this, we recognized the need for a deeper literature review on alternative architectures and modeling strategies, which would guide us in selecting and implementing the next model in our pipeline. Therefore, we implemented one more engineering cycle for the model development.

After implementing and evaluating LightGBM, we expanded our modeling strategy by exploring a fundamentally different machine learning approach. Our goal was to determine whether an alternative learning mechanism could capture additional patterns and potentially improve predictive performance.

Given our tabular dataset with complex non-linear dependencies, we focused our literature review on models specifically designed for tabular data prediction. We selected the TabPFN (Tabular Prior-Data Fitted Network) regressor, a transformer-based model pre-trained on millions of synthetic tasks, that enables accurate predictions with minimal hyperparameter tuning.

TabPFN leverages transformer architecture and extensive pre-training to capture complex, non-linear feature interactions in tabular data without requiring dataset-specific training. This fundamentally different mechanism from gradient boosting allowed us to explore whether alternative model architectures could enhance siRNA efficacy prediction.

Regarding our training dataset, we did not make any changes to it for this iteration of the engineering cycle.

We moved on to write code for the TabPFN Regressor Model. Full documentation for TabPFN model development is available in our GitLab Repository.

For the model performance evaluation, we used the Mean Squared Error (MSE) metric, and the Pearson Correlation Coefficient (R) metric. The table below shows the metrics values.

From the evaluation of the TabPFN regressor, we obtained a Mean Squared Error (MSE) of 0,0665 and a Pearson correlation coefficient (R) of 0,5576 on the test set. Although this MSE value is slightly higher than the one achieved by LightGBM (0,0529), it still reflects a reasonable prediction error. On the other hand, the Pearson correlation coefficient is slightly higher compared to LightGBM (0,5454), showing that TabPFN was able to capture the overall trend between predicted and true efficacies slightly better.

This comparison highlights the trade-off between the two models:

• LightGBM achieved lower prediction error, meaning its predictions were closer to the experimental values on average.
• TabPFN showed a marginally stronger correlation with the true values, suggesting it better preserved the global ranking and direction of siRNA efficacy trends.

Considering these complementary strengths, we propose both models as part of our modeling strategy, since each offers distinct advantages for understanding and predicting siRNA silencing efficacy.

Finally, we recognize that further engineering cycles could involve the construction of deep learning architectures tailored to RNA data. These may include Convolutional Neural Networks (CNNs) to capture local sequence motifs, LSTMs and attention layers to learn sequential dependencies, and the integration of pre-trained RNA language models such as RNA-FM. Such approaches could provide even more powerful modeling capabilities in future iterations of our work.