From Black Box to Biological Insight: Interpretable DNA Sequence Analysis with DEFT

A Q&A with Nicolas Huynh and Prof David Bentley, FRS on their AISTATS 2026 paper introducing a new framework for explainable genomics

We are delighted to share our first major paper, “Interpretable DNA Sequence Classification via Dynamic Feature Generation in Decision Trees”, at the intersection of machine learning and biology. DEFT (Dynamic Engineering of Features in Trees) represents a new way of thinking about AI for genomics: rather than trading accuracy for interpretability, it achieves both. We sat down with first author Nicolas Huynh and collaborator Prof David Bentley, FRS to discuss the work, its motivation, and what comes next.

A bit about our authors:

Nicolas Huynh

(referred to in the Q&A as NH), is a fourth-year PhD student in Machine Learning at the University of Cambridge (department of Applied Mathematics and Theoretical Physics), working in the van der Schaar lab (Machine learning and AI for medicine). He previously completed a Diplome d’ingénieur from Ecole Polytechnique (majoring in applied mathematics) and a MSc in Machine Learning from Ecole Normale Supérieure Paris-Saclay.

Prof David Bentley FRS

(referred to in the Q&A as DB), is a CCAIM Affiliated Faculty member, Professor of Biochemistry and Molecular Genetics and Co-Director of the RNA Bioscience Initiative at the U. Colorado School of Medicine. He previously held group leader positions at the Amgen Institute U. Toronto (1995-’98), and Cancer Research UK (1987-1995). His research focuses on the mechanisms controlling gene expression, particularly transcription by RNA polymerase II and the coordination of mRNA biogenesis. His “mRNA factory” model has been highly influential in shaping research on transcription-coupled RNA processing.

1. In simple terms, what is this paper about?

NH: This paper is about discovering human-interpretable patterns in DNA sequences that are associated with a biological phenomenon. In the paper’s main example, that phenomenon is RNA Polymerase II pausing. Instead of using a black-box model, we introduce our method that we call DEFT, which takes DNA sequences as input and builds a decision tree whose splits are based on automatically discovered, human-readable sequence features. The goal is not just prediction, but discovery: to identify features a biologist can inspect, understand, and test in the wet lab. 

DB: This paper reports a powerful new way of discovering DNA sequences that are important for a biological phenomenon chosen by the investigator.  As a proof of principle, we used this method called DEFT to reveal previously hidden features in the DNA that determine where RNA polymerase II pauses as it transcribes genes.

2. What limitations in current approaches to DNA sequence analysis motivated this work?

NH: The main limitation is that the strongest-performing methods in genomics are often black boxes. They can predict well, but they do not make it easy to understand why they make a prediction, which limits their usefulness for scientific discovery. Post-hoc explanation methods can help, but they are not always reliable or faithful to the model’s true prediction mechanism. On the other hand, standard interpretable models like decision trees are too limited for DNA: they usually split on one sequence position at a time, so they struggle to capture higher-order sequence patterns unless they grow very deep, which hurts both interpretability and generalization. Manual feature engineering is also limiting, because it depends on existing biological knowledge and is not adaptive to the local data structure encountered during tree construction. 

DB: If the DNA elements important for a particular mechanism controlling some aspect of gene expression are complex, rather than simple consensus sequences, then current methods are inadequate to identify them. DEFT brings the power of LLM’s to bear on this problem to reveal new insights.

3. Why is interpretability particularly important in this setting?

NH: Interpretability matters here because, in biology, prediction is usually not the end goal, but it is rather understanding. Biologists want models that can help them discover plausible mechanisms, validate findings against known biology, and generate new hypotheses. In our setting, interpretability is what turns prediction into a tool for discovery. DEFT is transparent by design: each feature comes with a name, a natural-language description, and executable code, so a practitioner can understand what the model is using and trace a prediction step by step. This is especially valuable in genomics, where researchers need to inspect proposed sequence determinants rather than just accept a score from a black box.

DB: Interpretability is critical if the results are to yield “actionable intelligence” for the biologist to use when she designs the next experiment.

4. How does your approach overcome the limitations of existing methods?

NH: Compared with black-box models, it is transparent by design, because the final predictor is a decision tree built from interpretable features. Compared with standard decision trees, it is much more expressive, because it can automatically discover new high-level sequence features during tree construction rather than relying only on raw positions or hand-crafted inputs. Technically, DEFT uses a large language model to propose candidate features, represent them in human-readable form, translate them into executable code, and iteratively improve them through a reflection mechanism. Crucially, this happens locally at each node of the tree, so feature generation is adaptive to the subset of sequences being split. That lets the model discover biologically meaningful features while remaining interpretable.

5. What kind of impact could this have for biological research or applications going forward?

NH: The broader impact is that it could help make AI for genomics more useful for scientific discovery, not just prediction. By producing human-understandable sequence features, DEFT can help researchers identify candidate mechanisms, assess whether the model’s behaviour is biologically plausible, and design follow-up experiments. More broadly, this kind of approach could increase trust in AI systems used in the life sciences and support a more collaborative human-AI workflow, where models suggest interpretable hypotheses that experts can review and validate.

DB: The impact of our approach is that it makes it feasible to ask questions of how complex features DNA sequence affect aspects of gene regulation that were not previously approachable.

6. What are you most curious to explore next following this work?

NH: We are excited about methods that let us discover richer cross-modal or modality-specific features, as well as better reasoning algorithms for these biological domains, using natural language as an interface for interpretability. It is also exciting to envision a human-in-the-loop version of DEFT, where biologists can guide the hypothesis-generation process with their own expertise, making the system more like a copilot for scientific discovery.

DB: I think it would be exciting to apply the DEFT approach to the question of how DNA sequence elements influence epigenetic regulation of gene activity.

This work highlights the potential of interpretable machine learning to drive biological discovery, opening new avenues for understanding complex genomic mechanisms. As approaches like DEFT continue to evolve, they offer a promising step towards more transparent and collaborative AI in the life sciences.

To explore more of our contributions at AISTATS 2026, please see here: https://www.vanderschaar-lab.com/aistats-2026/