Machine learning is transforming synthetic biology by enabling predictive design to replace traditional trial-and-error approaches. Advanced and efficient machine learning models can generate accurate predictions that accelerate laboratory workflows, advance research, and reduce experimental costs. However, model performance critically depends on training data quality, making extensive preprocessing essential to ensure models learn from rich, high-quality datasets.

In siREN, our machine learning models provided initial predictions of candidate siRNA silencing efficacy before experimental validation. To support model development, we created siRBench, a harmonized dataset containing more than 4000 siRNAs with experimentally measured silencing efficiencies and complementary thermodynamic and structural features.

Creating a high-quality dataset was essential before developing machine learning models. While multiple public siRNA sequence datasets exist, direct comparison reveals significant challenges:

• Inconsistent efficacy reporting: Silencing efficiencies are experimentally measured using different assays, protocols, and cell lines, and are often reported on different scales.
• Heterogeneous sequence formats: siRNAs vary in length (commonly 19 nt or 21 nt) and orientation (5’-3’ or 3’-5’), complicating direct integration.
• Data inconsistencies: Several datasets contain minor errors that undermine reproducibility.

As iGEM Thessaloniki 2025, we created siRBench, the first harmonized training and benchmarking dataset integrating siRNA data at scale. The siRBench dataset is available on our GitLab Repository.

We established clear specifications for siRBench to guide our dataset creation pipeline. For each siRNA sequence, we defined the following required fields:

• The siRNA antisense strand: 19 nt length, without overhangs, in 5'-3' orientation
• Exact target mRNA site: 19 nt sequence
• Extended target mRNA site: 57 nt sequence containing the 19 nt target site with 19 nt flanking regions on each side
• Efficacy score: Normalized value from 0 (ineffective) to 1 (highly effective) representing silencing performance
• Data source: The originating dataset source publication (from 11 public sources)
• Cell line: The experimental cell system used for efficacy measurements
• Engineered features: Computed properties for sequence analysis

We enforced the following standardization rules across all sequences in the dataset:

• Sequence uniformity: All siRNA antisense strands are 19 nt in length, oriented 5'→3', and reverse-complement to their corresponding exact target mRNA sites
• Efficacy normalization: All efficacy values are normalized to a 0–1 scale, representing the percentage of target mRNA knockdown (0 = no silencing, 1 = complete silencing)
• Uniqueness: Each siRNA-target mRNA pair appears only once in the dataset

The final siRBench dataset comprises 4098 siRNA antisense sequences from 11 different source publications: Huesken ¹, Takayuki ², Shabalina ³,Simone ⁴, Amarzguioui ⁵, Harborth ⁶, Hsieh ⁷, Khvorova ⁸, Reynolds ⁹, Vickers ¹⁰, Ui-Tei ¹¹. Note that Mixset is an aggregated dataset containing sequences from seven of these sources (Amarzguioui, Harborth, Hsieh, Khvorova, Reynolds, Vickers, and Ui-Tei), which we separated back into their original publications for proper cell line annotation.

The dropdown buttons below include the documentation of the preprocessing of each different dataset we found.

This comprehensive preprocessing procedure yielded a balanced, normalized and high-quality training dataset maximizing the number of unique, reliable siRNA sequences.

Feature engineering transforms raw data into numerical representations that machine learning models can process. For siRBench, we converted each siRNA-target mRNA pair into a feature set capturing sequence composition, structural properties, and thermodynamic characteristics.

Machine learning algorithms cannot directly interpret nucleotide sequences, which are categorical rather than numerical data. Instead, models learn from quantitative features such as nucleotide frequencies, position-specific k-mers, predicted secondary structures, duplex stability (ΔG), and other sequence-derived properties. By pre-calculating these features, we enabled our models to indentify motifs that distinguish highly effective siRNAs from less effective ones.

Feature engineering was particularly critical for integrating heterogeneous data from multiple public datasets. It provided a unified numerical framework for comparing siRNAs measured under different experimental conditions and cell lines. The quality and diversity of the engineered features directly impact model accuracy and generalizability, making feature engineering a central pillar of reliable siRNA efficacy prediction.

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.

Before model development, we strategically split siRBench into training and test sets based on experimental cell line rather than using random splitting methods like scikit-learn's train_test_split function. This cell line-based approach ensures the model is trained and evaluated on different biological contexts, preventing information leakage and better reflecting real-world scenarios where predictions must generalize to new cellular environments. The training set comprises 3,408 siRNAs with efficacy measurements from H1299 and HeLa cell lines, while the test set contains 690 siRNAs from all other cell lines. The training set was further randomly divided into training (90%) and validation (10%) subsets.

With siRBench finalized, we developed machine learning models to predict siRNA silencing efficacy. Our model development followed iterative engineering cycles using the Design-Build-Test-Learn framework (see our Engineering Page).

We explored multiple machine learning approaches and ultimately implemented two models with distinct architectures, both fully documented in our GitLab Repository.

We sought an advanced machine learning architecture suited to our task's characteristics: multiple engineered features, a medium-sized dataset, and potentially complex non-linear relationships. After literature review, we selected LightGBM (Light Gradient Boosting Machine) ¹³, 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.

We evaluated performance using Mean Squared Error (MSE) and Pearson correlation coefficient (R) to measure alignment between predicted and actual silencing efficacies.

We applied this model to predict silencing efficacy for all siRNAs targeting BCL2 and BTK genes, including both our own designs and experimentally validated sequences from the literature. Predictions are available on our Model Page.

Full documentation for the LightGBM model can be found in our GitLab Repository.

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 ¹⁴, 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.

We evaluated performance using Mean Squared Error (MSE) and Pearson correlation coefficient (R) to measure alignment between predicted and actual silencing efficacies.

We applied this model to predict silencing efficacy for all siRNAs targeting BCL2 and BTK genes, including both our own designs and experimentally validated sequences from the literature. Predictions are available on our Model Page.

Full documentation for TabPFN model development is available in our GitLab Repository.

Our software currently comprises two machine learning models for siRNA silencing efficacy prediction: a gradient-boosting framework (LightGBM) and a probabilistic foundation model (TabPFN). Both provide strong baselines for predicting silencing efficacy in our project. However, we envision several future advancements that could significantly enhance the scope and accuracy of our tool:

• Deep Learning Architectures: Integration of Convolutional Neural Networks (CNNs) for motif extraction, Long Short-Term Memory networks (LSTMs) for sequential dependencies, and Attention-based mechanisms (e.g., Transformers) for long-range contextual interactions.
• Pre-Trained Biological Models: Incorporation of RNA foundation models (such as RNA-FM) to leverage embeddings trained on large-scale RNA datasets, enabling generalization beyond our experimental training data.
• Expanded Feature Engineering: Addition of features beyond thermodynamic and sequence-based properties, such as secondary structure accessibility metrics, to further improve model performance.
• Interactive User Interface: Development of a user-friendly web application allowing researchers to input candidate siRNA sequences and receive real-time predictive scores in a user-friendly manner.

By pursuing these advancements, our software can evolve from a predictive tool into a comprehensive decision-support platform for rational design of therapeutic siRNAs.