Biophysically Informed Sequence Encodings for Weak-Affinity DNA-DNA Binding Prediction

Disclaimer: the text on this page was generated with the help of AI, based on my own ideas.

Status: manuscript under review. This page describes the motivation and the general approach only; methods and results are withheld until publication.

Whether two DNA strands will bind is easy to predict when they are near-perfect complements. Sequence-distance heuristics and thermodynamic estimates of hybridization free energy separate binders from non-binders well in that regime. The difficulty lies in the much larger intermediate space of weak-affinity interactions, where strands deviate substantially from perfect complementarity. That space is no longer a curiosity: in DNA-based data storage, unintended weak interactions between a primer probe and off-target strands are a direct source of retrieval noise, while molecular similarity search deliberately exploits non-orthogonal hybridization to return approximate matches. In both settings, predicting partially complementary interactions is the factor that limits specificity and scale.

Deep learning now outperforms the conventional metrics at this task. What has gone largely unexamined is the step before the model: how a pair of sequences is turned into numbers in the first place. The standard choice, one-hot encoding, is deliberately uninformative. It presents the network with two independent strings and leaves it to rediscover base pairing, complementarity, and the energetics of hybridization purely from data, even though these rules are already well established.

This work asks whether that knowledge is better supplied than learned. I design sequence-pair encoders that embed Watson-Crick complementarity and thermodynamic priors directly into the model's input representation, so that the relationship between the two strands is expressed in the encoding rather than inferred from examples. The encoders are evaluated against standard one-hot representations along four axes that matter for practical use: predictive accuracy, generalization to unseen sequences, robustness, and computational cost.

This is joint work with Karishma Matange and my advisor, Prof. James Tuck, at NC State University.