Inspiration

Researchers can generate thousands of antisense oligonucleotide (ASO) candidates, but testing them takes time and money. The challenge is deciding which experiments deserve a limited budget.

We built OligoAL around one question: How can every experiment help researchers find stronger candidates and choose the next experiment more effectively?

What it does

OligoAL helps researchers generate ASO candidates, predict supported biological properties, and prioritize which candidates to test.

Our experimental-selection harness works within a fixed assay budget. As measured results arrive, an adaptive efficacy model updates its estimates to inform subsequent selections.

We also designed a research portal with named batch tabs, candidate details, measurement logging, and persistent experimental context. Connecting that portal to the research harness is our next integration milestone.

How we built it

We prepared public oligonucleotide datasets while preserving chemistry, experimental conditions, and measurement provenance. The Atlas collection contains 283,521 assay rows across inhibition, dose-response, and toxicity tables.

We developed endpoint-specific predictors, explored RNA foundation-model adaptation, and built a sequential experimental-selection environment. A neural policy trained with proximal policy optimization (PPO) chooses among candidate-ranking strategies. A separate Bayesian correction updates efficacy estimates using acquired measurements.

We evaluated the system through historical replay, comparing methods under the same budget: 10 initial measurements and 40 additional selections.

Challenges we ran into

Biological datasets contain measurements collected under different conditions. We had to preserve distinctions between chemistry, dose, duration, and endpoints while preventing evaluation labels from influencing training or selection.

Another challenge was making reinforcement learning useful against strong conventional methods. Larger or more complex models did not automatically perform better. We kept random screening and simple ranking methods as controls and reported negative results alongside improvements.

Accomplishments that we're proud of

  • Useful ASO ranking: On two recorded panels, our model’s top-20 candidates averaged approximately 10 percentage points more measured inhibition than random selection.
  • More hits within a fixed budget: In one five-campaign retrospective replay, PPO found 8.36 new strong hits versus 3.76 for random screening, or 2.22 times as many at the same 50-assay budget. Adaptive mean ranking performed better still, at 9.20 hits.
  • Fresh-seed validation: A separate siRNA study showed a modest, statistically supported improvement in predictive accuracy from uncertainty-based acquisition across 40 fresh paired split seeds.
  • A reusable research foundation: We built provenance-aware data preparation, prediction tools, an experimental-selection harness, and a portal concept that retains context across batches.

These are retrospective findings, rather than prospective assay savings or validated therapeutic outcomes.

What we learned

Predicting activity and choosing experiments are different problems. A useful predictor does not automatically produce the best experimental strategy.

We learned that simple baselines can be difficult to beat, and that honest comparisons make the system stronger. We also learned to keep predictions separate from observations: model outputs guide hypotheses, while measured results provide evidence.

What's next for OligoAL

Our next step is to run a prospective research pilot.

Researchers would record each batch’s measurements and experimental conditions, then receive recommendations informed by their accumulated results. We also want to evaluate more independent target genes and improve selection across multiple compatible endpoints.

Our goal is to help researchers make every assay contribute to a better next decision.

Built With

Share this project:

Updates

Submission history