Research

We do AI for Science, that is, building machine intelligence that works alongside human reasoning to accelerate scientific discovery. Our core conviction is that AI systems must go beyond pattern recognition: they need to reason, explain, and collaborate with humans in the pursuit of scientific knowledge. To that end, we develop methods that combine the scalability of modern deep learning with the interpretability and structure of symbolic reasoning, and apply them to real-world scientific problems in computational biology and drug discovery. We welcome collaborations on complex, real-world problems; if you are interested in working together, please feel free to write to the PI.

Neurosymbolic AI

Statistical learning and symbolic reasoning have complementary strengths. Deep neural networks learn well from large, noisy data but are opaque and data-hungry. Symbolic methods such as Inductive Logic Programming (ILP) are interpretable and data-efficient but do not scale to raw data easily. We build neurosymbolic methods that integrate both: neural networks guided by symbolic domain knowledge, and symbolic engines that learn from neural representations.

Key contributions include Compositional Relational Machines (CRMs), knowledge-guided graph neural networks (BotGNN, VEGNN), symbolic neural generation (SNG), and methods for extracting interpretable rules from trained networks. Applications span drug discovery, genomics, and other scientific domains where both data and prior knowledge are available.

Deep Learning

We work on foundational aspects of deep learning relevant to scientific applications: sequence models and transformers for genomic data, model calibration and uncertainty quantification, knowledge distillation, and explainability. A recurring theme is reliability: predictions in science must come with well-calibrated confidence and interpretable justification.

We also study large language models (LLMs) for scientific tasks, including using LLMs with logical feedback for molecular design, and building genomic foundation models that capture regulatory and codon-usage patterns across eukaryotes.

AI for Science

Biology is a natural domain for AI for Science: the data are vast, the problems are hard, and mechanistic interpretability is essential. We work on:

• Gene regulation: modeling the relationship between DNA sequence and gene expression.
• Genomic foundation models: learning general-purpose representations of genomic sequences.
• Multi-omics analysis: interpretable and knowledge-guided ML for biomarker discovery and disease subtyping across omics data.
• Drug discovery: neurosymbolic and graph neural network models for molecular property prediction, lead generation, and explainable screening.
• Biomedicine: analyzing video, imaging, and clinical text to support diagnosis, assessment, and decision-making.

Research in these directions involves collaborating with scientists from different domains.

Causal Machine Learning

Correlation-based models are brittle: they fail under distribution shift and cannot answer interventional or counterfactual questions that science demands. We are interested in causal machine learning, specifically in methods that go beyond associational learning to discover and exploit causal structure in data.

This includes learning causal graphs from observational and interventional data, integrating causal priors with deep learning, and using causal reasoning to improve robustness and interpretability of models in biological and clinical settings. A longer-term goal is to build AI systems that can reason about why a biological phenomenon occurs, not just that it occurs, which is a prerequisite for trustworthy AI in science.

Most of our code is publicly available at: lab-github and PI-github.