Source-linked AI summary
Condition aware learning enables robust prediction of oligonucleotide melting behavior across diverse chemistries and assay conditions
Danielle L. Ferreira, Lifeng Lin, Adam Aslam, Nicholas Chang, Rebekah G. Baig, Edgar Baculi, Zoey Cao, Melanie Senn
TL;DR
Generalization beyond the training distribution is a key challenge for data-driven approaches. This framework combines thermodynamic priors, sequence representations, explicit experimental context, and synthetic pretraining to achieve sub-degree accuracy across unmodified and LNA-modified oligonucleotides while maintaining strong external-benchmark performance.
Problem
Generalization beyond the training distribution is a key challenge for data-driven approaches.
Method
The framework integrates thermodynamic priors, sequence representations, explicit experimental context, and thermodynamic knowledge from synthetic pretraining.
Results
Sub-degree prediction accuracy was achieved across unmodified and LNA-modified oligonucleotides, with more accurate modification-effect capture and strong independent-benchmark performance.
Takeaways & Limitations
Condition-aware nucleotide language models provide an effective extension of conventional nearest-neighbor thermodynamic approaches for oligonucleotide melting-temperature prediction.
Takeaways & Limitations
Modified-oligonucleotide measurements came from a single laboratory, covered two LNA substitution types, and require broader replication.
Abstract
from arXiv · showhide
Oligonucleotide melting temperature is a fundamental determinant of nucleic acid hybridization and underpins the design of molecular diagnostics, polymerase chain reaction assays, and many other biotechnology applications. However, accurately predicting melting behavior remains difficult because it depends not only on sequence composition, but also on experimental conditions and chemical modifications commonly used in modern assay design. Existing thermodynamic models rely on fixed parameterizations that are often difficult to extend across diverse reaction environments and nucleotide chemistries. Here we show that a condition-aware nucleotide language model can accurately predict oligonucleotide melting behavior across diverse experimental conditions and both unmodified and chemically modified oligonucleotides. By combining contextual sequence representations with explicit information describing the reaction environment, the framework achieves sub-degree prediction accuracy and reduces prediction error for locked nucleic acid-modified oligonucleotides by up to 25% relative to nearest-neighbor thermodynamic approaches. The model also more accurately captures the thermal effects introduced by nucleotide modification and maintains strong performance on independent benchmark datasets spanning experimental conditions substantially different from those represented during training. Our results demonstrate that learned sequence representations can complement classical thermodynamic models by capturing context-dependent effects that are difficult to encode using fixed parameter tables alone. More broadly, this work provides a scalable framework for predicting oligonucleotide melting behavior across diverse chemistries and assay conditions, supporting more reliable molecular assay design.
AI framework for melting temperature prediction
The condition-aware AI framework predicts melting temperatures across unmodified and LNA-modified oligonucleotides, with strongest gains for chemically modified sequences. It maintains strong agreement across sequence classes, assay conditions, and independent benchmarks, including settings outside training conditions.
- Held-out prediction performance: The AI model achieved the strongest agreement and lowest prediction errors across unmodified and LNA-modified oligonucleotides.Its largest gains occurred for LNA-containing sequences, where contextual representations captured sequence-dependent modification effects.
- Modification effects: Up to 41% lower prediction error was observed for terminal LNA modifications relative to the thermodynamic baseline.The AI model maintained comparatively stable accuracy across LNA position classes, while Thermo showed larger errors and positive bias for terminal and near-terminal modifications.
Contributions
The study was conceived, developed, experimentally evaluated, and interpreted through contributions spanning computational modeling, laboratory work, and manuscript preparation.
- D.F., L.L., and M.S. conceived and designed the study.
- D.F. developed the computational framework, generated synthetic datasets, and performed model training and data analysis.
- A.A. and N.C. performed wet-laboratory experiments with guidance from L.L. and R.B.
- D.F., L.L., and M.S. interpreted the results and wrote the manuscript, which all authors reviewed and approved.
Additional Information
Additional information describes the study’s supplementary datasets and analyses, including condition-stratified splits, position-dependent LNA performance, and agreement comparisons across datasets and assay conditions.
- The supplementary materials include a condition-stratified split-composition table and a table of position-dependent Tₘ prediction performance for LNA-modified oligonucleotides.
- Mean absolute error and prediction bias were stratified by LNA chemistry and modification position, with lower error and bias across all position classes.The largest improvements were observed for terminal modifications.
- Agreement analyses compared experimental melting temperatures with Thermo and the proposed AI model across datasets, GC-content groups, and sequence-length groups.
- Across the reported strata, the AI model exhibited narrower limits of agreement and reduced residual variability than the thermodynamic baseline.
- Agreement analyses across oligonucleotide concentration, potassium concentration, and magnesium concentration showed stable AI-model agreement across experimental conditions.The AI model consistently had narrower limits of agreement than the thermodynamic baseline.