APTAi: De Novo Aptamer Design for Proteomic Biomarker Detection Using a Physics-Informed AI Model

Transcript:

Sepsis affects around 50 million lives annually, but treatment is delayed by ineffective diagnostics, having a misdiagnosis rate of around 8-20%. Aptamers are promising for detecting proteomic biomarkers like Procalcitonin, yet their discovery is bottlenecked by the expensive, labor-intensive SELEX process.

To solve this, I developed APTAi, a novel computational pipeline that utilizes a Conditional Variational Autoencoder and Monte-Carlo Tree Search to generate high-affinity sequences in-silico. By mapping biomarker features onto a 423-dimensional Riemannian manifold, APTAi navigates the target’s unique surface topography.

Integrating physics-based validation, including ΔG thermodynamics, molecular docking, and toxicity assessment ensures stable binding and manufacturability. This de novo approach replaces months of laboratory research with a rapid, generalizable solution for the future of disease diagnostics.

Proteomic biomarkers serve as key metrics that can signal the onset and presence of many diseases. One of the world's most life threatening conditions, Sepsis-induced Systemic Inflammatory Response Syndrome(SIRS) claims over 11 million lives annually, and has symptoms that are not very specific to the disease, leading to late or inaccurate diagnoses, making up around 8.2-20% of cases (Figure 1). The progression of Sepsis is shown in Figure 2, with an infection eventually leading to non-localized inflammation and organ failure. The current "gold standard" is a CMIA blood test (Figure 3) that has require a 24-to-48 hour latency period. In Sepsis treatment, every hour of delayed treatment increases mortality by 8%, making this diagnostic gap fatal.

Aptamers—small, synthetic ssDNA/ssRNA strands—have many characteristics ideal for the next generation of rapid diagnostics. Unlike bulky antibodies, aptamers are "chemical antibodies" that fold into specific 3D shapes: (a) Hairpins, (b) Stem-loops, (c) Pseudoknots, and (d) G-Quartets (Figure 4). These diverse morphologies allow them to bind to targets like Procalcitonin with high affinity. They offer superior thermal stability, lower immunogenicity, and faster, cheaper synthesis in vitro, making them ideal for point-of-care biosensors.

Currently, the development of these aptamer sequences is bottlenecked by the traditional SELEX (Systematic Evolution of Ligands by Exponential Enrichment) process(Figure 5). SELEX is a stochastic search that begins with a massive library of trillions of sequences, which are fed through an affinity column to find binders. This trial-and-error methodology has four major limitations as highlighted in Figure 6.

1. Feature Engineering & Manifold Mapping

The system begins by extracting the features of the proteomic target into a 423-dimensional vector. This vector captures 5 key biochemical signatures: Amino Acid Composition (20 dim), Dipeptide Distribution (20 x 20 = 400 dim), Aromaticity (1 dim), Hydropathy (1 dim), and Isoelectric Point (1 dim). Unlike standard models that treat search as a "flat" Euclidean problem, APTAi transforms this space into a Riemannian Manifold. By calculating the Metric Tensor (gij), the system accounts for the complex topography of the protein surface. This allows the AI to calculate the Geodesic Path (dg), which is the shortest, most stable sequence orientation that stays flush with the target's unique molecular curvature. This allows the aptamer to be designed to go over the protein's surface rather than through it like a traditional Euclidian path.

2. Generative Search (CVAE + MCTS)

To navigate a search space of 432 (18 quintillion) possible combinations, APTAi utilizes a Conditional Variational Autoencoder (CVAE), and generates the sequences using a Monte-Carlo Tree Search (MCTS).

The ELBO Constraint: The Evidence Lower Bound (ELBO) serves as the objective function, balancing sequence "memory" (reconstructing chemical grammar) with "manifold exploration" (regularizing the latent space).

MCTS Navigation: The MCTS agent explores this latent space. Using Upper Confidence Bound (UCT) , the agent balances the exploitation of known high-affinity motifs with the exploration of novel regions on the manifold. This prunes away unstable designs, converging on the theoretical binding optimum.

3. Physics-Based Validation & Deep Filtering

To bridge the gap between potential AI hallucination and biological reality, candidates pass through a Deep Neural Network (DNN) filter, which reduces the search space by 1.84 x 1015. Lead candidates undergo molecular dynamics simulations to ensure a Dissociation Constant (Kd) in the low nanomolar range and an RMSD < 1.5 Å, guaranteeing both stability and clinical potency.

1. Core Validation Metrics

The lead candidate, was evaluated across a high-dimensional feature set to ensure clinical and biophysical viability:

Predictive Accuracy: The high-throughput predictor reached a Mean Squared Error (MSE) of 0.07, allowing for the digital screening of millions of sequences with high confidence.

Binding Strength: The lead candidate achieved a predicted G of -12.5 kcal/mol and a Dissociation Constant ( ) in the low nanomolar range, rivaling commercial monoclonal antibodies.

Structural Integrity: Molecular docking simulations (HADDOCK 2.4) yielded a 1.42 Å RMSD, proving a near-perfect geometric fit with the Procalcitonin surface.

2. Physics-Informed Architecture & Selection

The pipeline utilizes a Composite Scoring Formula to ensure Explainable AI (XAI) and optimization of all metrics:

(Weights: Predictor 40%, Docking 20%, Folding 20%, Specificity 15%, Motif 5%)

As shown in the data, APTAi's thermodynamic optimization process successfully shifted the entire candidate pool from a random distribution (avg. -4 kcal/mol) to a high-stability cluster (avg. -12 kcal/mol).

Pareto Optimization: The final candidate was selected via Pareto analysis, identifying the "Goldilocks zone" that maximizes binding affinity ( G) while minimizing nucleotide length to reduce manufacturing costs and steric hindrance.

3. Performance Analysis (Figure Interpretations)

CVAE Training Loss: The convergence of the KL-Divergence and Reconstruction loss (reaching a stable 10.0) confirms the latent space is well-organized and capable of generating foldable, biologically "grammatical" sequences.

Neural Network: The Neural Network Binding Predictor also achieved a <0.1 MSE after 100 epoch training on the dataset.

MCTS Convergence: The selection path reached stability at ~800 iterations, proving the algorithm effectively navigated the 432 (18 quintillion) possibilities to find the mathematical global optimum.

Biomolecular Manifold Mapping: By treating the protein surface as a Riemannian Manifold, APTAi accounted for Van der Waals forces and electrostatic potentials, allowing the agent to follow the "path of least resistance" to the optimal binding pocket.

4. Statistical Significance: The Ablation Study

An ablation study was conducted by removing different components of the pipeline, comparing the probability densities of each ablated pipeline resulting in a certain binding affinity.

The Ablations were:

No ablation (full pipeline)

No CVAE Latent Space

No Binding Neural Network

Baseline, with no MCTS as it is the model that creates the sequences.

The results showed that all of the pipeline components were necessary for the improvement of the aptamer designs, the most important being the Binding Neural Network, apart from the MCTS, which is necessary as it creates the sequences themselves.

APTAi allows for rapid, point-of-care diagnostics to be developed, by first manufacturing the oligonucleotide sequence, as well as any possible modifications for integrations

There are 2 ways a commercial lab would be able to manufacture the aptamer. Firstly, these custom sequences can be ordered from oligo synthesis companies who use various methods such as solid phase synthesis to create the exact sequence for further testing on things like binding affinity and toxicity. Secondly, the sequence may already be used for another purpose. In this scenario, labs should check the off-target score of the aptamer relative to the other target, making sure that the aptamer doesn’t accidentally bind to the wrong target.

In figure 23, the aptamer is integrated into a biosensor, where it poses as the main part for the detection of the biomarker. When the aptamer binds with the biomarker, it creates resistance in a current given by stage 4 using a process called Electrochemical Impedance Spectroscopy (EIS). As the concentration of the biomarker, in this case Procalcitonin, increases, the resistance increases, causing the microchip to accurately find the amount of biomarker present. With Sepsis, the more Procalcitonin found indicates higher risk and likelihood of death, allowing doctors to quickly know which patients need more urgent care, ensuring that proper treatment is given, when seconds can be the difference between life and death.

Future development of APTAi will focus on both the biological realism of the aptamers as well as the computational power of the platform.

1. The model should be able to recognize new binding sites for proteins with post-translational modifications like glycosylation (Fig 24).

2. Incorporating 3D structure prediction modules to penalize unstable conformations and reward sequences with consistent, high-affinity binding pockets.

3. Enhanced Filtering for advancing safety and off-target modules using motif analysis to automatically reject high-risk or cross-reactive motifs.

4. Wet-Lab Validation from binding assays and SELEX data to refine and further confirm the model’s predictive accuracy.

I would like to thank my parents for supporting me and giving me help and advice throughout this project. I would also like to thank both the TSF and CWSF organizers and judges for giving your time to support youth STEM. Thank you!

References

[1] S. A. Abid, A. Ahmed Muneer, I. M. S. Al-Kadmy, et al., "Biosensors as a future diagnostic approach for COVID-19," Life Sciences, vol. 273, p. 119117, 2021.

[2] Aptamer Group, "What is an aptamer?" YouTube, Jan. 8, 2021. [Online]. Available: https://www.youtube.com/watch?v=JKgymhUFPck

[3] SomaLogic, "Aptamers vs. Antibodies," Apr. 28, 2025. [Online]. Available: http://somalogic.com/aptamers-vs-antibodies/

[4] A. V. Lakhin, V. Z. Tarantul, and L. V. Gening, "Aptamers: Problems, Solutions and Prospects," Acta Naturae, vol. 5, no. 4, p. 34, 2024.

[5] A. Bashir, Q. Yang, J. Wang, et al., "Machine learning guided aptamer refinement and discovery," Nature Communications, vol. 12, no. 1, 2021.

[6] A. Bateman, M.-J. Martin, A. Adesina, et al., "UniProt: the Universal Protein Knowledgebase in 2025," Nucleic Acids Research, vol. 53, no. D1, 2024.

[7] M. Biesiada, K. J. Purzycka, M. Szachniuk, et al., "Automated RNA 3D Structure Prediction with RNAComposer," RNA Structure Determination, pp. 199–215, 2016.

[8] Bio-Helix, "Bio-Helix.com," 2016. [Online]. Available: https://www.bio-helix.com/products/237

[9] University of Bath, "Biosensors: sensing elements and techniques," n.d. [Online]. Available:

https://www.bath.ac.uk/corporate-information/biosensors-sensing-elements-and-techniques/

[10] CDC, "About Sepsis," Aug. 19, 2025. [Online]. Available: https://www.cdc.gov/sepsis/about

[11] E. Chovancova, A. Pavelka, P. Benes, et al., "CAVER 3.0: A Tool for the Analysis of Transport Pathways in Dynamic Protein Structures," PLoS Computational Biology, vol. 8, no. 10, p. e1002708, 2012.

[12] J. Cruz-Toledo, M. McKeague, X. Zhang, et al., "Aptamer Base: A collaborative knowledge base to describe aptamers and SELEX experiments," Database, vol. 2012, 2012.

[13] D. G. Lemay, "Glyco-what? Glycoproteins," International Milk Genomics Consortium, Nov. 22, 2023.

[14] C. Doersch, "Tutorial on Variational Autoencoders," arXiv, 2016. [Online]. Available:

https://arxiv.org/pdf/1606.05908

[15] Edinburgh Instruments, "The Beer-Lambert law," Jul. 8, 2021. [Online]. Available: https://www.edinst.com/resource/the-beer-lambert-law/

[16] N. Emami and R. Ferdousi, "AptaNet as a deep learning approach for aptamer–protein interaction prediction," Scientific Reports, vol. 11, no. 1, 2021.

[17] L. Evans, A. Rhodes, W. Alhazzani, et al., "Surviving Sepsis campaign: International Guidelines for Management of Sepsis and Septic Shock 2021," Intensive Care Medicine, vol. 47, no. 11, pp. 1181–1247, 2021.

[18] C. Fleischmann-Struzek, D. M. Goldfarb, P. Schlattmann, et al., "The global burden of paediatric and neonatal sepsis: a systematic review," The Lancet Respiratory Medicine, vol. 6, no. 3, pp. 223–230, 2018.

[19] F. Gao, J. Yin, Y. Chen, et al., "Recent advances in aptamer-based targeted drug delivery systems for cancer therapy," Frontiers in Bioengineering and Biotechnology, vol. 10, 2022.

[20] R. Gómez-Bombarelli, J. N. Wei, D. Duvenaud, et al., "Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules," ACS Central Science, vol. 4, no. 2, pp. 268–276, 2018.

[21] J. G. Greener, S. M. Kandathil, L. Moffat, and D. T. Jones, "A guide to machine learning for biologists," Nature Reviews Molecular Cell Biology, vol. 23, no. 1, 2021.

[22] R. Hamnett, "Aptamers | Antibodies.com," Antibodies.com, 2015.

[23] S. Hassibian, M. Amin, S. M. Taghdisi, et al., "Aptamers: Design, Theory, and Applications to

Diagnosis and Therapy for Diseases," MedComm, vol. 6, no. 5, 2025.

[24] T. Hermann, "Adaptive Recognition by Nucleic Acid Aptamers," Science, vol. 287, no. 5454, pp. 820–825, 2000.

[25] I. L. Hofacker, "Vienna RNA secondary structure server," Nucleic Acids Research, vol. 31, no. 13, pp. 3429–3431, 2003.

[26] R. V. Honorato, M. E. Trellet, B. Jiménez-García, et al., "The HADDOCK2.4 web server for integrative modeling of biomolecular complexes," Nature Protocols, 2024.

[27] IBM, "Latent Space," Jan. 28, 2025. [Online]. Available: https://www.ibm.com/think/topics/latent-space

[28] T. J. Iwashyna, E. W. Ely, D. M. Smith, and K. M. Langa, "Long-term Cognitive Impairment and Functional Disability Among Survivors of Severe Sepsis," JAMA, vol. 304, no. 16, p. 1787, 2010.

[29] S. James, G. Konidaris, and B. Rosman, "An Analysis of Monte Carlo Tree Search," Proceedings of the 31st AAAI Conference, 2017.

[30] Y. Jiang, Y. Song, Q. Zeng, and B. Jiang, "Mesenchymal Stem Cells and Their Extracellular Vesicles Are a Promising Alternative to Antibiotics for Treating Sepsis," Bioengineering, vol. 11, no. 11, p. 1160, 2024.

[31] J. Jumper, R. Evans, A. Pritzel, et al., "Highly Accurate Protein Structure Prediction with Alphafold," Nature, vol. 596, no. 7873, pp. 583–589, 2021.

[32] N. Khalid, "Aptamers Provide Instant Photonics-Based Diagnostics," Photonics.com, Mar. 17, 2026.

[33] J. Kim, S. Park, D. Min, and W. Kim, "Comprehensive Survey of Recent Drug Discovery Using Deep Learning," International Journal of Molecular Sciences, vol. 22, no. 18, p. 9983, 2021.

[34] M. Kohlberger and G. Gadermaier, "SELEX: Critical factors and optimization strategies for successful aptamer selection," Biotechnology and Applied Biochemistry, 2021.

[35] H. Y. Kong and J. Byun, "Nucleic Acid Aptamers: New Methods for Selection, Stabilization, and Application in Biomedical Science," Biomolecules and Therapeutics, vol. 21, no. 6, pp. 423–434, 2013.

[36] A. Kumar, D. Roberts, K. E. Wood, et al., "Duration of Hypotension before Initiation of Effective Antimicrobial Therapy Is the Critical Determinant of Survival in Human Septic Shock," Critical Care Medicine, vol. 34, no. 6, pp. 1589–1596, 2006.

[37] J. Li, Q. Zhao, H. Gao, et al., "Early biomarkers for predicting sepsis-induced shock: insights from inflammatory pathways and immune response," Frontiers in Pharmacology, vol. 17, 2026.

[38] J. Liu and Y. Lu, "Fast Colorimetric Sensing of Adenosine and Cocaine Based on a General Sensor Design Involving Aptamers and Nanoparticles," Angewandte Chemie, vol. 45, no. 1, pp. 90–94, 2006.

[39] Q. Liu, W. Zhang, S. Chen, et al., "SELEX tool: a novel and convenient gel-based diffusion method

for monitoring of aptamer-target binding," Journal of Biological Engineering, vol. 14, no. 1, 2020.

[40] M. Lönne, G. Zhu, F. Stahl, and J.-G. Walter, "Aptamer-Modified Nanoparticles as Biosensors," Advances in Biochemical Engineering/Biotechnology, vol. 140, pp. 121–154, 2013.

[41] R. Lorenz, S. H. Bernhart, C. Höner zu Siederdissen, et al., "ViennaRNA Package 2.0," Algorithms for Molecular Biology, vol. 6, no. 1, 2011.

[42] Z. Lu, "Thrombin aptamer," Ribocentre.org, Apr. 3, 2024.

[43] Lucas Learns, "Aptamers and SELEX Simply Explained," YouTube, Apr. 28, 2022.

[44] D. H. Mathews, M. D. Disney, J. L. Childs, et al., "Incorporating chemical modification constraints into a dynamic programming algorithm for prediction of RNA secondary structure," PNAS, vol. 101, no. 19, pp. 7287–7292, 2004.

[45] Mayo Clinic, "Sepsis - Diagnosis and Treatment," Feb. 10, 2023.

[46] M. Meisner, "Update on Procalcitonin Measurements," Annals of Laboratory Medicine, vol. 34, no. 4, pp. 263–273, 2014.

[47] G. M. Morris, R. Huey, W. Lindstrom, et al., "AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility," Journal of Computational Chemistry, vol. 30, no. 16, pp. 2785–2791, 2009.

[48] MultiMolecule, "RNABERT - MultiMolecule," 2022. [Online]. Available: https://multimolecule.danling.org/models/rnabert/

[49] N. M. O’Boyle, M. Banck, C. A. James, et al., "Open Babel: An open chemical toolbox," Journal of Cheminformatics, vol. 3, no. 1, 2011.

[50] C. J. Paoli, M. A. Reynolds, M. Sinha, et al., "Epidemiology and costs of sepsis in the United States," Critical Care Medicine, vol. 46, no. 12, pp. 1889–1897, 2018.

[51] Radiometer, "Procalcitonin (PCT) – the biomarker of choice to aid in the diagnosis of sepsis," 2018.

[52] P. Prommapan, N. Brljak, T. W. Lowry, et al., "Aptamer Functionalized Lipid Multilayer Gratings for Label-Free Analyte Detection," Nanomaterials, vol. 10, no. 12, p. 2433, 2020.

[53] PubChem, "Procalcitonin," NIH, 2025.

[54] K. Reinhart, R. Daniels, N. Kissoon, et al., "Recognizing Sepsis as a Global Health Priority — A WHO Resolution," New England Journal of Medicine, vol. 377, no. 5, pp. 414–417, 2017.

[55] Roger, "Variational Autoencoder(VAE)," Medium, Jul. 8, 2021.

[56] K. E. Rudd, S. C. Johnson, K. M. Agesa, et al., "Global, regional, and National Sepsis Incidence and mortality, 1990–2017," The Lancet, vol. 395, no. 10219, pp. 200–211, 2020.

[57] M. Schütte, S. Hocker, et al., "A Machine Learning Framework for Scattering Kernel Derivation Using Molecular Dynamics Data in Very Low Earth Orbit," Research Gate, Jul. 7, 2025.

[58] Sepsis Alliance, "Symptoms," 2020. [Online]. Available: https://www.sepsis.org/sepsis-basics/symptoms/

[59] C. W. Seymour, F. Gesten, H. C. Prescott, et al., "Time to Treatment and Mortality during Mandated Emergency Care for Sepsis," NEJM, vol. 376, no. 23, pp. 2235–2244, 2017.

[60] M. Singer, C. S. Deutschman, and C. W. Seymour, "The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3)," JAMA, vol. 315, no. 8, pp. 801–810, 2016.

[61] Y. Tan, M. Li, Z. Zhou, et al., "PETA: evaluating the impact of protein transfer learning with sub-word tokenization on downstream applications," Journal of Cheminformatics, vol. 16, no. 1, p. 92, 2024.

[62] University of Vienna, "TBI - ViennaRNA Package 2," n.d. [Online]. Available:

https://www.tbi.univie.ac.at/RNA/

[63] Biosearch Technologies, "Where are all the therapeutic aptamers?" Jun. 9, 2022.

[64] R. Thevendran and M. Citartan, "Assays to Estimate the Binding Affinity of Aptamers," Talanta, vol. 238, p. 122971, 2022.

[65] O. Trott and A. J. Olson, "AutoDock Vina: Improving the Speed and Accuracy of Docking," Journal of Computational Chemistry, vol. 31, no. 2, 2009.

[66] A. L. Vijayan, Vanimaya, S. Ravindran, et al., "Procalcitonin: a promising diagnostic marker for sepsis and antibiotic therapy," Journal of Intensive Care, vol. 5, no. 1, 2017

.

[67] K. Vo, "Spectrophotometry," Chemistry LibreTexts, 2022.

[68] Base Pair Bio, "What is an Aptamer? – Aptamers and SELEX," Apr. 6, 2015.

[69] WHO, "Global report on the epidemiology and burden of sepsis," Sep. 9, 2020.

[70] Wikipedia Contributors, "Procalcitonin," Apr. 11, 2026. [Online]. Available: https://en.wikipedia.org/wiki/Procalcitonin

[71] WuXi AppTec, "How Does It Work? | Aptamers," YouTube, Sep. 13, 2022.

[72] Z. Zhao, H. Lin, H. Y. Lau, et al., "Structure-enhanced deep learning accelerates aptamer selection for small molecule families like steroids," Briefings in Bioinformatics, vol. 26, no. 6, 2025.

[73] J. Zhou and J. Rossi, "Aptamers as targeted therapeutics: current potential and challenges," Nature Reviews Drug Discovery, vol. 16, no. 3, pp. 181–202, 2016.