Overview

Our 2025 modeling integrates two layers that together drive an engineering loop from sequence design to in-cell delivery: a sequence-level model that predicts siRNA–mRNA silencing efficacy and a system-level model that screens delivery ligands for lipid nanoparticles (LNPs). The two layers connect through our wet-lab pipeline and Human Practices (HP), turning modeling outputs into concrete design decisions for FeliSilence.

Layer 1 — siRNA efficacy (what to deliver).
We built siRNAttention, an attention-based heterogeneous graph neural network that models siRNA, mRNA and their interaction as distinct node types. Unlike rule-based heuristics (e.g., GC%), siRNAttention leverages attention-driven message passing, positional encoding, thermodynamics, base-pairing probabilities and AGO2-related features to capture long-range and contextual effects. Trained with siRNA-based 10-fold splits on the HUVK dataset, it achieved performance comparable to the current state of the art, supporting reliable in-silico prioritization of Fel d 1 siRNAs before synthesis and testing.

Layer 2 — LNP ligand docking (how to deliver).
To enable non-invasive delivery, we modeled oral/sublingual routes and focused on the feline M3 muscarinic acetylcholine receptor (M3R), which is abundantly expressed in salivary glands. We reconstructed the feline M3R structure via AlphaFold/ColabFold with active-state templates, relaxed it using AMBER, predicted binding pockets with PrankWeb, and performed ~30,000-compound docking using QuickVina2. The top ligands showed strong predicted affinities (around –16.5 to –15.7 kcal/mol), yielding promising candidates for LNP surface decoration to trigger receptor-mediated uptake.

Interface between the two layers.
Layer 1 ranks siRNA candidates for Fel d 1 silencing; Layer 2 proposes ligand handles to decorate LNPs that carry these siRNAs. Practically, this forms a predict → select → package workflow: (1) use siRNAttention to shortlist high-efficacy siRNAs; (2) encapsulate those sequences in LNPs; (3) decorate LNPs with top-ranked M3R ligands to enhance uptake via salivary-gland epithelia.

Integration with experiments and HP.
Model outputs guide what we synthesize/test (candidate siRNAs) and how we design delivery (ligand-decorated LNPs). In the lab, we plan qPCR/fluorescence readouts for knock-down ranking consistency (Layer 1), and receptor-binding/uptake assays in feline salivary cell lines for ligand hits (Layer 2). For HP and Entrepreneurship, the modeling quantifies feasibility, reduces screening cost, and supports a consumer-safe, scalable concept—evidence we communicated to stakeholders and advisors during interviews.

What’s next.
We will (i) expand hyperparameter tuning and out-of-distribution tests for siRNAttention, (ii) validate top ligands experimentally for binding and LNP uptake, and (iii) couple the two layers into an end-to-end automated design loop (siRNA ranking → LNP formulation → ligand selection), accelerating the path from in-silico prediction to in-vitro verification and ultimately informing product design.

siRNA Efficacy Prediction

Summary: We used a novel heterogenous graph attention network to predict siRNA-mRNA efficacy, achieving similar results to state of the art models

1. Introduction

One of the crucial aspects of RNA interference is the selection of a siRNA sequence that can silence the targeted mRNA effectively. For our project, we’ve designed numerous siRNA sequence capable of silencing Fel d 1 at differing efficacies. Many guidelines have been developed, namely the calculation G/C [1]. With the recent development of AI, the paradigm has shifted from manual rule-based prediction to machine learning [2, 3] to deep learning techniques [4, 5, 6]. Current state of the art models use a graph neural network approach, where mRNA and siRNA interactions are models using nodes and edges [4]. The use of graphs is naturally reasonable for modeling biological problems which involve components connected to each other, making it a good fit for this efficacy prediction problem. A graph neural network operates by passing information among a node and its neighbor, and predictions can be done on a graph by attaching a predictive head onto it. A good introduction to graph neural networks can be found at [7].

There are numerous weaknesses to existing graph neural network-based approaches that can be summarized as follows:

1. Existing models use simple aggregation, mean, and pooling methods when calculating message propagation, which limits the expressiveness of the model.

2. Many models also lack long-range dependency and global contextual awareness, which is crucial for accurately predicting efficacy as nucleotides far from the base pairing site may still influence the result.

3. Existing models rely on arbitrary predefined rules to improve performance, which increases model complexity with marginal benefits [4].

4. Models tend to significantly lose accuracy when tested on new datasets and struggle generalizing out of distribution (i.e. model struggle to learn the underlying patterns outside of their seen training data)

To facilitate the design of a highly effective siRNA sequence, we developed a novel attention-based heterogenous graph network. Attention refers to the weighing of each message the node receives, which allows for the model to choose what message should be prioritized over another. siRNAttention avoids using manual rule-based features, which are largely arbitrary.

2. Objectives

● O1: Create a model capable of accurately predicting siRNA efficacy.

● O2: Incorporate attention-based mechanisms to enhance model expressiveness.

● O3: Reduce the use of manually defined rule-based features.

3. Methods

3.1 Data Preparation

For our experiment, we use the compiled dataset which is collectively called ‘Dataset_HUVK’ from [4]. The dataset consists of 2816 siRNA-mRNA pair and their corresponding efficacy. The dataset includes repeated siRNA and mRNA sequences. This means that if the same siRNA or mRNA appear in both the training and testing dataset, the model will not accurately reflect its results on unseen data. To prevent this data leakage, we follow the siRNA based 10-fold data split proposed in [4].

3.2 Architecture

Figure 1. High-level overview of model's architecture. Part A shows a graphical representation of the relationship between the siRNA, mRNA, and the interaction node. Part B shows the neighboring message attention scoring mechanism for the interaction node. Part C shows a overview of the message passing layer architecture.

We define a heterogeneous graph (i.e. a graph with numerous node and edge types) to represent the siRNA to mRNA interaction. Following existing literature, we created three types of nodes representing siRNA, mRNA, and an interaction node. We propagate two bidirectional relationships between siRNA-interaction and mRNA-interaction. Each relationship (e.g. siRNA to interaction) is parametrized by its own set of weights to maximize the expressivity and allow for differentiation of different node types. The final efficacy value is predicted using a linear layer which has access to the interaction node.

3.3 Features

The siRNA and mRNA sequences are one-hot encoded to represent the mRNA and siRNA sequence. That is, the sequence is represented by a vector where the nucleotides A, C, G, T are encoded with [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1], respectively.

Many deep learning models lack positional context, which is crucial for siRNA-mRNA interaction tasks as a small change in the ordering may drastically alter efficacy. We follow existing work and use a sinusoidal embedding to encode positional context.

We also provide model features regarding the thermodynamic stability profile of the siRNA, the siRNA-mRNA base pairing probability, and RNA-protein interaction probability. The enlisted features crucially model the interaction’s free energy, its ability to base pair, and the mRNA and siRNA sequence’s probability of interaction with the AGO2 protein, which is part of the RISC assembly process.

Regional context is provided to the model by calculating nucleotide frequencies around each node. Each node is given its 1-mer (e.g. A) to 5-mer (e.g. AUUCG) regional context. This has shown to greatly improve performance as it allows the model to learn common motifs.

Our model avoids defining any features which may be arbitrary, such as rule-based scores and G/C percentages. The features can be summarized as the following:

Node Type	Features
siRNA	Sequence Embedding Positional Embedding Nucleotide Frequencies siRNA-protein Interaction Probabilities Base-pairing probabilities
mRNA	Sequence Embedding mRNA-protein Interaction Probabilities Base-pairing probabilities
Interaction	Thermodynamics Profile siRNA-mRNA Interaction Probabilities

Table 1. Node features for siRNA, mRNA, and the interaction node

Hyperparameters
Batch size: 64 Epochs: 16 Layer Size: [32, 32] Dropout: 0.2 Learning Rate: 0.0005 Attention Heads: 2 Attention Dropout: 0.1

Table 2. Hyperparameter used to train the final model

3.4 Message Passing

Each graph neural network consists of a message passing layer, where information is transferred between nodes. For this design, we use two message passing layers of dimension each. The message passing layer consists of three parts, an interaction-specific attention scoring, a level-specific attention scoring, and a transformer-like propagation. For each layer, we compute the message each node will send (e.g. siRNA nodes will send messages to the interaction node) using a multi-head GATv2 convolution layer. By using attention, the model can learn to weigh the importance of each message propagation. The feature of each node is also transformed using a self-projection linear layer. To calculate the message the interaction node should receive, we pass through an attention network the siRNA and mRNA message as well as the interaction node’s self-projection, returning a weighted sum of the siRNA and mRNA message. To finally update the nodes, we concatenate the neighbor message and self-projection and pass through a transformer inspired block consisting of layer normalization and feed forward network. We use the output as the new node feature.

3.5 Loss & Evaluation

The model is trained using the Adam optimizer with a learning rate of 5e-4 and uses the MSE loss. We’ve also experimented with incorporating a loss that jointly optimizes MSE and PCC. Note that although PCC is scale invariant, it is still practical for situations involving ranking potential siRNA candidates. For evaluation, we calculated the model’s test set performance in terms of PCC, SPCC, MSE, and AUC defined at the threshold of >0.7.

4. Results & Discussion

Figure 2. Pearson correlation coefficient on the HUVK dataset between baselines and siRNAttention. The result of siRNADiscovery is reproduced using the author’s official codebase, while the PCC of the other baselines are extracted from literature.

On the HUVK dataset, our model performs similarly to siRNADiscovery, the current state of the art model for siRNA-mRNA efficacy prediction. This shows that the use of attention-based mechanisms is highly promising, and match SAGE-based inductive methods. This importantly shows that a well parametrized attention-based model can match out of distribution generalization of methods designed for it, namely inductive learning models like HinSAGE. Our model also has shown to converge quickly and achieve reasonable results after 8 to 16 epochs. To note, our model also achieves similar PCC even after lowering layer size to as low as 8, demonstrating siRNAttention’s scalability advantage.

5. Limitations & Future Work

Unlike many predecessors, our model did not undergo rigorous hyperparameter optimization using tools like Optuna. Further experimentation should examine the benefits of hyperparameter tuning for improving the model’s performance. Additionally, our model did not focus on generalizing to out-of-distribution data samples, which may explain the lack of improvement over existing methodologies when running inference on the testing set. The use of attention also improves interpretability and allows for statistical analysis on the contribution of siRNA/mRNA sequences to the final efficacy, which should be examined.

Future work should focus on combining inductive learning techniques [4, 5, 8] with attention mechanisms to be able to predict well on unseen data while retaining the benefits of modulating information communication between nodes. Moreover, in vitro trials should be conducted to validate efficacy prediction claims of siRNAttention. It may also be of interest to integrate siRNA sequence design into the model natively, as well as predicting and preventing off-target effects.

Although we were unable to employ the model in the design process of our siRNA sequence, in future work we can try to optimize our existing siRNA sequence to maximize efficacy using siRNAttention. Moreover, the development of a general purpose siRNA-mRNA efficacy prediction model reduces in vitro trial-and-error cost and allow us to quickly develop new siRNA sequences targeting different proteins. This allows us to expand our marketing endeavor in the future rapidly and possibly to other animals/cat proteins.

6. Reproducibility & Reuse

The official code for the siRNAttention model can be found in https://drive.google.com/drive/folders/1_HzlRG85fSJdpWSfEHdDAmFy2trxdf3l?usp=sharing.

The full HVUK dataset and data preprocessing utility functions can be found in siRNADiscovery’s official repository at https://github.com/BertramLoong/siRNADiscovery

7. References

[1] Safari, F., Barouji, S. R., & Tamaddon, A. M. (2017). Strategies for improving siRNA-induced gene silencing efficiency. Advanced Pharmaceutical Bulletin, 7(4), 583–592. https://doi.org/10.15171/apb.2017.072

[2] Filhol, O., Ciais, D., Lajaunie, C., et al. (2012). DSIR: assessing the design of highly potent siRNA by testing a set of cancer-relevant target genes. PLoS One, 7(10), e48057. https://doi.org/10.1371/journal.pone.0048057

[3] Ichihara, M., Murakumo, Y., Masuda, A., et al. (2007). Thermodynamic instability of siRNA duplex is a prerequisite for dependable prediction of siRNA activities. Nucleic Acids Research, 35(18), e123. https://doi.org/10.1093/nar/gkm699

[4] Long, R., Guo, Z., Han, D., Liu, B., Yuan, X., Chen, G., Heng, P.-A., & Zhang, L. (2024). siRNADiscovery: a graph neural network for siRNA efficacy prediction via deep RNA sequence analysis. Briefings in Bioinformatics, 25(6), bbae563. https://doi.org/10.1093/bib/bbae563

[5] La Rosa, M., Fiannaca, A., La Paglia, L., & Urso, A. (2022). A graph neural network approach for the analysis of siRNA-target biological networks. International Journal of Molecular Sciences, 23(22), 14211. https://doi.org/10.3390/ijms232214211

[6] Han, Y., He, F., Chen, Y., et al. (2018). siRNA silencing efficacy prediction based on a deep architecture. BMC Genomics, 19(Suppl 1), 59. https://doi.org/10.1186/s12864-018-5028-8

[7] Sanchez-Lengeling, B., Reif, E., Pearce, A., & Wiltschko, A. B. (2021). A gentle introduction to graph neural networks. Distill. https://doi.org/10.23915/distill.00033

[8] Hamilton, W. L., Ying, R., & Leskovec, J. (2017). Inductive representation learning on large graphs. ArXiv. https://arxiv.org/abs/1706.02216

siRNA Delivery Ligand Docking

Lipid Nanopariticle Packaging For SiRNA Delivery

To deliver our siRNA by means of lipid nanoparticle, we analyzed potent pathways, created novel feline M3 muscarinic acetylcholine receptor, and ran a ~30k molecule docking test using the ChemDiv’s GCPR targeted library.

1. Introduction & Why Modeling

As part of our team’s future design goal, we plan on designing a mechanism for transporting the Fel d1 silencing siRNA sequence into the cat’s bloodstream orally. By transporting orally through saliva gland cells, we minimize intrusiveness. Delivering orally allow us to package the siRNA sequence into consumer products such as cat food, increasing our product's marketing potential. One prominent way of doing so is to package the siRNA into a lipid nanoparticle (LNP). We can then decorate the LNP with a ligand that can bind to cell receptors, and deliver the LNP into the bloodstream through activating cell’s ingestive mechanism. In order to successfully deliver the ligand decorated LNP, a potent receptor should be selected. An ideal target should be abundant in the targeted cell type, have good track record for binding affinity, and is able to get into the bloodstream.

Figure 1. Summary of siRNA lipid nanoparticle packaging. From El Moukhtari, S. H., et al. Advanced Drug Delivery Reviews 201 (2023): 114940.

To select an ideal ligand target, we decided to use docking techniques in order to screen through a large databank of potential candidates, reducing the cost of in vitro experimentation and time consumption. Docking techniques allow us to simulate the binding pose and affinity of thousands of molecule in minutes to hours, drastically speeding up the process.

2. Background

In order to conduct docking experiments to find idea ligand targets, we first need to find a receptor capable of ingesting the lipid nanoparticle into the bloodstream. Many oral delivery pathways have shown great potential, namely the sublingual and buccal routes, which allow for the absorption into the bloodstream while bypassing the first-pass effect (i.e. metabolic breakdown before entering systemic circulation). The class of muscarinic acetylcholine receptors is particularly promising due to its high expression in salivary glands and importance in stimulating secretion, making it an ideal target for inducing receptor-mediated endocytosis.

The five types of muscarinic acetylcholine receptors are all promising targets, with M1R primarily being neural, M2R being cardiac, and M3R responsible for stimulating glandular secretion and smooth muscle contraction. Research across multiple species confirm that M3Rs are predominant within salivary glands [1].

Figure 2. Signaling pathways of muscarinic acetylcholine receptors. From Wess et al., Nat Rev Drug Discov, 2007

Despite lack of research and evidence, M3Rs are likely still the dominant receptor in felines. Evidence from reverese transcription polymerase chain reaction (RT-PCR) has shown M3R’s abundance within feline, with gene homology between feline M3R and human M3R estimated at 89% [8].

The nervous system can also influence secretion through A1-adrenergic receptors while activating b-adrenergic receptors triggers exocytosis through cyclic AMP. This is also prospective but secondary to M3R approaches.

Other receptors like vasoactive intestinal peptide, nerve growth factor, and transforming growth factor have specialized functions that are important for tissue repair. For general delivery purposes, these may not be ideal.

Overall, compared to M1R, adrenergic receptors, and other receptors for tissue repair, M3R has more empirical backing and should thus be experimented with. However, due to the lack of literature on feline specific genomic data and research, further experimental data may still be needed. Our goal is to have the binding ligand become engulfed into the cell through the process of receptor-mediated endocytosis.

3. Methods

3.1 Modeling Novel M3 Muscarinic Acetylcholine Receptor

The target feline M3 muscarinic acetylcholine receptor (M3R) is not found in public databases like UniProt [10]. M3R for felines can be found at NCBI Gene ID 101100919 and is verified by running a BLAST [2] to compare similarity to other M3Rs.

Figure 3. Running BLAST to check if NCBI Gene ID matches M3R

3 preliminary models were created on AlphaFold [4] using different seeds. The structures returned a null ipTM (i.e. prediction for accurate of protein-protein interaction, which is not present) and a pTM score of ~0.6, an indication of resemblance but not accurate enough.

Figure 4. Alphafold prediction of feline M3 muscarinic acetylcholine receptor

To ensure more fine-grained control, the ColabFold [6] library is used, which allowed hyperparameter control and template passing. Active-state templates were provided to AlphaFold for reference to prevent the generation of inactive-state proteins, which do not reflect real-life docking situations. Existing experimentally verified active GPCR structures were extracted from gpcrdb.org and BLAST was used to find structure similar to the feline M3R. The top eight results were 8E9W, 8E9Z, 8E9Y, 6U1N, 6OIJ, 7TRK, 7V68, and 4MQS. Multiple trials were ran and the results were compared.

Figure 5. Prediction of lDDT, the local distance difference that compares the structural similarity between the estimated ground truth and current structure. We see high confidence (~0.95) in transmembrane helix regions but significant dips (~0.25) in extracellular loops. The dips can be explained by intrinsic randomness of extracellular loops.

Figure 6. Top prediction from AlphaFold. Most of the important transmembrane helices score confidence intervals of >90 plDDT, indicating high accuracy. Qualitatively, the results are much better than from Google’s AlphaFold server, which had numerous transmembrane helices that scored <50 plDDT.

The final prediction shows high plDDT in transmembrane helices, which are regions that involve binding sites, while the less-relevant extracellular loops scored low plDDT due to inherent noise. The model is then relaxed and preprocessed prevent steric clashes and maximize docking capabilities. The AMBER algorithm is ran for 2000 iterations to minimize conformation energy.

3.2 Binding Site Prediction

To identify promising docking sites, the freely available docking site prediction tool PrankWeb is used. PrankWeb employs a P2Rank [5] backbone, which uses machine learning to make predictions based on scoring and clustering points on the protein’s solvent accessible surface.

Figure 7. Docking pockets identified by PrankWeb.

The top candidate scored 13.05 and a probability of 0.634. The cluster of docking sites aligned with literature position. Next, a volume test is conducted and reported 819.3 Angstrom cubed. A docking test with the receptor antagonist 4-DAMP is conducted and reported an energy of -7.186 kcal/mol, indicating potential but low efficacy.

3.3 Docking Experiment

After showing that the initial setup is valid, a docking test is conducted on around 30k molecule from ChemDiv’s GPCR dataset, a well-curated and focused library of synthetically accessible ligands. QVina2 [11], a faster and similarly accurate version of autodock vina, is used. The ligand library was converted to the PDBQT format and Gasteiger charges are added to provide electrostatic information for docking calculations.

Figure 8. Top ten scoring ligands.

Figure 9. The 3D conformation of the top three ligands, visualized in PyMol

The high binding affinity score from -16.5 to -15.7kcal/mol indicates high potential viability of the ligands for delivery. In the future, the top candidates should be tested in vitro to validate high affinity scores. Considering that autodock programs have the tendency to prioritize large and polar/charged molecules, the results are not definitive and should be further validated.

4. Conclusion

In summary, we modeled a novel M3 muscarinic acetylocholine receptor for feline and conducted a docking experiment using around 30k molecule taken from ChemDiv's GPCR library. In future work, we should experiment with the top ligand candidates in vitro and package our siRNA sequence into LNPs for production use. The novelty of the design ensures a high barrier of entry that gives us an advantage when entering the consumer market, increasing our design's profitability and consumer satisfaction.

5. References

[1] Nakamura T, Matsui M, Uchida K, Futatsugi A, Kusakawa S, Matsumoto N, Nakamura K, Manabe T, Taketo MM, Mikoshiba K. M(3) muscarinic acetylcholine receptor plays a critical role in parasympathetic control of salivation in mice. J Physiol. 2004 Jul 15;558(Pt 2):561-75. doi: 10.1113/jphysiol.2004.064626. Epub 2004 May 14. PMID: 15146045; PMCID: PMC1664962.

[2] Camacho, C., Coulouris, G., Avagyan, V., Ma, N., Papadopoulos, J., Bealer, K., & Madden, T. L. (2009). BLAST+: architecture and applications. BMC Bioinformatics, 10, 421.

[3] El Moukhtari, S. H., et al. (2023). Lipid nanoparticles for siRNA delivery in cancer treatment. Journal of Controlled Release, 361, 130-146.

[4] Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature. https://doi.org/10.1038/s41586-021-03819-2

[5] Krivak, R., & Hoksza, D. (2018). P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. Journal of Cheminformatics.

[6] Mirdita, M., Schütze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., & Steinegger, M. (2022). ColabFold: Making protein folding accessible to all. Nature Methods. https://doi.org/10.1038/s41592-022-01488-1

[7] Polak, L., Skoda, P., Riedlova, K., Krivak, R., Novotny, M., & Hoksza, D. (2025). PrankWeb 4: a modular web server for protein–ligand binding site prediction and downstream analysis. Nucleic Acids Research.

[8] Preiksaitis, H. G., & Laurier, L. G. Pharmacological and molecular characterization of muscarinic receptors in cat esophageal smooth muscle.

[9] Wess, J., Eglen, R. M., & Gautam, D. (2007). Muscarinic acetylcholine receptors: mutant mice provide new insights for drug development. Nature Reviews Drug Discovery.

[10] The UniProt Consortium, 2021. UniProt: the universal protein knowledgebase in 2021. Nucleic Acids Research, 49(D1), D480-D489.

[11] Alhossary, A., Handoko, S. D., Mu, Y., & Kwoh, C. K. (2015). Fast, accurate, and reliable molecular docking with QuickVina 2. Bioinformatics, 31(13), 2214-2216.