AI-Driven Machine Learning Framework for Therapeutic Co-Design: TP53-Targeted ASOs and Polymeric Nanoparticles as a Benchmark in Ovarian Cancer

Background

Therapeutic design across nucleic acids and nanocarriers is an AI/ML (artificial intelligence / machine learning) problem: vast, coupled design spaces; hard, multi-objective trade-offs (selectivity, stability, delivery, toxicity); and expensive experiments. Our central premise is to place AI/ML at the core—using generative models with physics-informed constraints and active, uncertainty-aware optimisation—while using biology strictly as the validation loop. The outcome we seek is a general, transferable methodology for co-designing sequence-level therapeutics and delivery systems that can extend far beyond any single disease.

We focus on two interlocking spaces. First, antisense oligonucleotides (ASOs) engineered as hairpins can recognise single-nucleotide variants and intercalate small-molecule chemotherapeutics, enabling dual-layer precision: (i) drug release triggered only upon target hybridisation and (ii) sequence-level selectivity. Second, poly-(β-amino ester) (pBAE) nanoparticles offer highly tunable delivery (monomers, end-groups, ratios) but are intractable to optimise by trial-and-error. Jointly searching these spaces demands generative modelling + Bayesian optimisation under structural/energetic constraints and closed-loop experimental feedback.

Clinical nanotechnology’s persistent bottleneck—limited tumour selectivity and off-target toxicity despite improved pharmacokinetics—reinforces the need for this AI/ML-first approach. Decades after Doxil, relatively few nanotherapies have translated; targeted constructs often stall in late-phase trials due to heterogeneity and residual toxicity. Rather than incremental formulation tweaks, we treat the problem as data-efficient global search over coupled design variables, guided by priors from RNA structure and hybridisation thermodynamics.

Within this AI/ML framework, ovarian cancer is a first benchmark—chosen not to centre the biology but to stress-test the platform under realistic constraints. High-grade serous ovarian cancer (HGSOC) exhibits frequent TP53 point mutations (e.g., R273C/H/L), creating a crisp discrimination task for mutation-selective ASOs and a demanding delivery scenario for pBAEs. Standard care (surgery + IV chemotherapy), prior intraperitoneal strategies, and HIPEC have improved some outcomes yet still face high recurrence and mortality; this underscores the difficulty of the benchmark, not the scope of our ambition. Success here validates the method, which is then portable to other mutations, cancers, and non-oncology indications.

Hypothesis

A multi-objective, uncertainty-aware AI/ML pipeline that co-designs mutation-selective hairpin ASOs and optimises pBAE formulations—constrained by RNA structure/energetics and trained in an active-learning loop—will yield synergistic selectivity and lower effective doses in the ovarian cancer testbed, while establishing a generalizable AI/ML methodology for precision nanotherapies.

Objectives

1) AI/ML-driven ASO design (M1–6).
Develop generative, physics-informed models to propose hairpin ASOs with single-mutation selectivity for TP53 R273C/H/L, structural stability compatible with drug intercalation, and minimal off-targets.

2) AI/ML-guided nanoparticle optimisation (M4–9).
Use Bayesian optimisation and active learning over pBAE chemistry and formulation to maximise uptake and endosomal escape in the test system while minimising toxicity in controls.

3) High-throughput, closed-loop validation (M6–12).
Engineer a biosensor-style assay that quantifies mutation selectivity, cellular uptake, and triggered drug release; stream results back to continuously retrain AI/ML models.

4) Ex vivo benchmark on ovarian cancer (M6–12).
Fabricate top AI/ML-predicted hASO–pBAE constructs and validate selective delivery to TP53-mutant cells vs healthy comparators, quantifying synergy between ASO silencing and triggered chemotherapy release. (This constitutes the first testbed of a broader, disease-agnostic platform.)

Evidence of Recent Research Activity

Papers
Tsikonofilos, K., Bruyns-Haylett, M., May, H. G., Donat, C. K., & Kozlov, A. S. (2025). Alterations in topology, cost, and dynamics of gamma-band EEG functional networks in a preclinical model of traumatic brain injury. Network Neuroscience, 1–26.
May, H. G., Tsikonofilos, K., Donat, C. K., Sastre, M., Kozlov, A. S., Sharp, D. J., ... (2024). EEG hyperexcitability and hyperconnectivity linked to GABAergic inhibitory interneuron loss following traumatic brain injury. Brain Communications, 6(6), fcae385. https://doi.org/10.1093/braincomms/fcae385
May, H. G., Tsikonofilos, K., Donat, C. K., Sastre, M., Kozlov, A. S., Sharp, D. J., ... (2023). Mild blast TBI raises gamma connectivity, EEG power, and reduces GABA interneuron density. bioRxiv. https://doi.org/10.1101/2023.12.01.569541
Azeem, A., Julleekeea, A., Knight, B., Sohail, I., Bruyns-Haylett, M., & Sastre, M. (2023). Hearing loss and its link to cognitive impairment and dementia. Frontiers in Dementia, 2, 1199319. https://doi.org/10.3389/frdem.2023.1199319
Loecher, A., Bruyns-Haylett, M., Ballester, P. J., Borrós, S., & Oliva, N. (2023). A machine learning approach to predict cellular uptake of pBAE polyplexes. Biomaterials Science, 11(17), 5797–5808. https://doi.org/10.1039/D3BM00626C
Cotur, Y., Olenik, S., Asfour, T., Bruyns-Haylett, M., Kasimatis, M., Tanriverdi, U., Gonzalez-Macia, L., Lee, H. S., Kozlov, A. S., & Güder, F. (2022). Bioinspired stretchable transducer for wearable continuous monitoring of respiratory patterns in humans and animals. Advanced Materials, 34(33), 2203310. https://doi.org/10.1002/adma.202203310

Grants applied to:
Horizon Europe – Health Cluster (HORIZON-RIA)
•Title: Synergistic Modulation of Neural Inhibition and Parasympathetic Tone to Mitigate Pollution-Accelerated Brain Ageing
•Acronym: NEUROCLEAR
•Call/Topic: HORIZON-HLTH-2025-03-ENVHLTH-01-two-stage
•Proposal number: SEP-211227004
•Duration: 48 months
•Submission date: 16/09/2025
•Budget asked: 7,000,000 euros
________________________________________
2. Indústria del Coneixement 2025 (Llavor – Modalitat A)
•Title: Multi-Modal Non-Invasive Vagus Nerve Stimulation Device with Adaptive Biofeedback
•Applicant: Universitat Ramon Llull
•PI: Michael Bruyns-Haylett
•Expedient: 2025 LLAV 00141
•Submission: 2025 call
•Budget asked: 20,000 euros
________________________________________
3. Indústria del Coneixement 2024
•Title: Bio- and Neuro-feedback based educational platform using real time adaptive generative AI (EduSmart)
•Applicant: Universitat Ramon Llull
•PI: Michael Bruyns-Haylett
•Submission: 2024 call (detailed memory prepared)
•Budget asked: 20,000 euros