Nikolaos Nakis

Spotlight @aistats_conf 2026🎉GraphHull: an explainable graph generative model with two-level convex hulls—global archetypes = “pure" communities, local prototypes = within-community variation. Identifiable by design (non-overlapping local hulls). Paper: arxiv.org/pdf/2602.21342

A central result is that these mechanisms play different roles across domains: interdependence contributes most strongly to social ties, whereas health and economic ties are shaped more by dependence and individual-level behavior.

We introduce the Multiplex Latent Trade-off (MLT) model, a framework for identifying how independence, dependence, and interdependence shape structure in multiplex networks.

Enrichment analysis revealed that archetypes from the positive space align with processes like cell cycle regulation and immune response, while negative space archetypes capture functions like antiviral defense and cytoskeletal organization.

We also performed Gene Ontology enrichment analysis on the archetypes derived from our dual latent spaces, revealing distinct biological tasks associated with archetypal groups formed by both positive and negative interactions.

We validated our approach on SIGNOR 3.0 signed PPI networks for Homo sapiens, Mus musculus, and Rattus norvegicus—ensuring robustness on real-world data. Our model achieves a 4.3% average F1 improvement over state-of-the-art methods while delivering clear biological insights.

Our S2-SPM leverages two independent archetypal latent spaces to capture both activating (positive) and inhibitory (negative) protein–protein interactions, truly reflecting biological complexity and offering a nuanced view of the interactome.

In PPIs, the sign differentiates activating from inhibitory interactions—each playing distinct regulatory roles. Overlooking the sign means missing how proteins modulate processes; negative interactions do not simply imply “no interaction” but represent vital inhibitory controls.

Protein–protein interactions are key to decoding cellular processes and disease mechanisms; however, experimental mapping remains both costly and challenging. A crucial detail often overlooked is the sign of PPIs.

You can read the preprint at: arxiv.org/pdf/2503.03904

Enrichment analysis revealed that archetypes from the positive space align with processes like cell cycle regulation and immune response, while negative space archetypes capture functions like antiviral defense and cytoskeletal organization.

We also performed Gene Ontology enrichment analysis on the archetypes derived from our dual latent spaces, revealing distinct biological tasks associated with archetypal groups formed by both positive and negative interactions.

We validated our approach on SIGNOR 3.0 signed PPI networks for Homo sapiens, Mus musculus, and Rattus norvegicus—ensuring robustness on real-world data. Our model achieves a 4.3% average F1 improvement over state-of-the-art methods while delivering clear biological insights.

Our S2-SPM leverages two independent archetypal latent spaces to capture both activating (positive) and inhibitory (negative) protein–protein interactions, truly reflecting biological complexity and offering a nuanced view of the interactome.

New paper accepted at #ICLR2025 ! Excited to share our findings on achieving ultra-low dimensional representations for exact graph reconstruction—extended to large-scale networks for the first time. Submission can be accessed here: (openreview.net/pdf?id=V71ITh…)

Exciting news! 🎉 Our paper on 2-level polarization in complex networks has been accepted to #AISTATS2025! Huge thanks to my amazing co-authors for their incredible contributions!