Aptamer Predictive Triage (APT): Generative & Geometric Deep Learning for Drug Discovery

The problem: finding molecular binders is slow and expensive

Aptamers are short, single-stranded DNA or RNA oligonucleotides that fold into precise three-dimensional shapes and bind a chosen target — a protein, a small molecule, or a whole cell — with antibody-like specificity. Because they are chemically synthesised rather than grown in living systems, they are cheaper, more stable and easier to modify than antibodies, which makes them attractive building blocks for diagnostics, biosensors and targeted therapeutics.

The difficulty lies in discovery. The classical laboratory route, SELEX (Systematic Evolution of Ligands by EXponential enrichment), iterates through many rounds of binding, washing and amplification to enrich a random library toward high-affinity sequences. It works, but it is slow, costly and biased by the experimental protocol, and it explores only a vanishingly small corner of a sequence space that grows as 4^N with sequence length. For a research team trying to hit a novel target, each additional SELEX campaign is weeks of wet-lab effort with an uncertain outcome.

The idea: predict and triage candidates before the wet lab

Aptamer Predictive Triage reframes discovery as a machine-learning problem. Instead of searching the sequence space blindly in the lab, the system learns the relationship between an aptamer's structure and its ability to bind a given target, then uses that learned model to generate promising candidates de-novo and triage them — ranking a large pool of proposals so that only the most credible sequences are carried forward to costly experimental validation.

Two complementary capabilities sit at the core of the work:

• De-novo generation. A generative model proposes new candidate aptamers conditioned on a target, rather than mutating an existing pool. This lets the search reach regions of sequence space that iterative enrichment would never sample.
• Predictive triage. A scoring model estimates binding plausibility for each candidate, turning an unmanageable set of possibilities into a short, prioritised list — the "triage" that gives the project its name.

Why geometric deep learning

An aptamer's behaviour is dictated by its folded shape, not by its raw sequence of letters. Treating a molecule as flat text throws away exactly the information that determines binding. Geometric deep learning offers a better inductive bias: by representing molecular structure as graphs and reasoning over spatial relationships, the models respect the symmetries of the underlying chemistry — the same principles that power modern structure-aware architectures for proteins and nucleic acids.

In this project, geometric and generative deep learning are combined so that the generator produces structurally coherent candidates and the triage model evaluates them in a representation that reflects three-dimensional folding rather than surface-level token statistics. This structure-first perspective is what distinguishes the approach from sequence-only baselines.

How generation and triage work together

The two components are designed as a loop rather than a one-shot pipeline. The generator proposes a broad, diverse pool of candidate sequences for a given target; the triage model scores and ranks them; and the highest-ranked candidates are the ones a lab would synthesise and test. Crucially, this framing is compatible with the experimental reality of aptamer discovery — the model does not claim to replace validation, it prioritises it. Every real binding result that comes back can, in principle, sharpen the scoring model, so the system improves as it is used. The goal is not a single perfect prediction but a dramatically better hit rate per experiment, which is where the cost of discovery actually lives.

The technical stack

The pipeline is built in Python with PyTorch, using geometric deep-learning tooling to encode molecular graphs, generative modelling for de-novo proposal, and standard scientific-Python infrastructure (NumPy, pandas, scikit-learn) for data preparation, evaluation and analysis. The workflow is organised as a reproducible research pipeline — data curation, model training, candidate generation and triage scoring — so that results can be regenerated and extended rather than produced one-off.

Beyond therapeutics: diagnostics and biosensing

Because aptamers are cheap, stable and easy to functionalise, their usefulness extends well past drug candidates. The same de-novo design and triage machinery applies to aptamer-based biosensors ("aptasensors") for rapid diagnostics, to affinity reagents that can stand in for antibodies in laboratory assays, and to targeted delivery vehicles. Anywhere a highly specific, synthesisable molecular recognition element is needed, a computational method that proposes and ranks strong binders for a novel target shortens the path from idea to working assay — which is why predictive triage is best understood as a general capability for molecular design, not a single-use tool.

Why it matters for healthcare and medicine

Anything that shrinks the search at the front of the drug-development funnel compounds downstream. By filtering candidates computationally before committing wet-lab resources, predictive triage aims to reduce the number of physical SELEX rounds needed to reach a viable binder, lower cost, and open the door to targets that are difficult to address experimentally. The same machinery generalises beyond therapeutics to diagnostics and biosensing, wherever a highly specific, synthesisable molecular recognition element is required.

Aptamer Predictive Triage sits within a broader research focus on applied AI for healthcare and medicine — using generative and geometric deep learning to make biological discovery faster, cheaper and more targeted.