Charilaos I. Kanatsoulis

Problem. Existing pipelines produce sparse, noisy KGs from text.

Contribution. A Python package (pip install kg-gen) producing high-fidelity KGs via LLM extraction + entity clustering. Released MINE, the first benchmark for KG extraction.

Impact. 1.1k+ GitHub stars; widely used for biomedical and scientific reasoning.

Problem. KG embeddings for biomedical discovery require fusing structured and textual evidence.

Contribution. Coupled tensor-matrix factorization for biomedical knowledge graphs, applied to COVID-19 drug repurposing, an early AI-driven hypothesis-generation pipeline.

Problem. Single-cell transcriptomics foundation models treat genes as independent tokens, ignoring regulatory structure.

Contribution. A multimodal graph transformer that injects gene-regulatory-network structure as positional information.

Key insight. Biological priors during pretraining capture long-range gene-token dependencies for cancer, Alzheimer's, and other downstream tasks.

Problem. LoRA's matrix factorization leaves parameter-efficiency on the table for multi-axis adaptation.

Contribution. Generalizes LoRA to tensor decomposition, reducing trainable parameters by up to two orders of magnitude at matched performance.

Applications. DPO, instruction tuning, vision, and protein folding fine-tuning.

Problem. Eigenvector-based positional encodings (e.g., Laplacian PEs) give strong inductive bias to graph transformers but require expensive eigendecomposition and suffer from instability and limited expressivity.

Contribution. PEARL is a learnable PE framework that approximates equivariant functions of eigenvectors with linear complexity, matching or surpassing full eigenvector PEs at one-to-two orders of magnitude lower cost.

Key insight. Message-passing GNNs initialized with random / basis-vector node features compute nonlinear maps of eigenvectors, unlocking expressive positional encodings without explicit eigendecomposition. A foundational primitive for our GFM work.

Problem. Common belief: Weisfeiler-Lehman bounds GNN expressivity.

Contribution. GNNs can discriminate the majority of real graphs; WL is not the real limit. We design convolutional architectures that are provably more expressive than WL.

Problem. Counting substructures (triangles, cycles, paths, cliques) is fundamental to chemistry, biology, and graph reasoning, yet standard message-passing GNNs are bounded by 1-WL and provably cannot count most of them.

Contribution. A theoretical and architectural framework characterizing exactly when GNNs can count substructures, together with provably expressive architectures that go beyond 1-WL.

Key insight. Substructure counting is a much finer-grained yardstick for GNN expressivity than the 1-WL test, and reveals the latent power of message-passing models.