Nabila Siregar

Hi, I’m an AI engineer and researcher based in Amsterdam. Currently, I work at IBM, building autonomous agents for open-ended discovery. My research interests are in deep learning and reinforcement learning, especially AI for science.

Previously: I did my MSc in Computational Science at the University of Amsterdam, supervised by Dr. Phong Le and Dr. Raquel G. Alhama, researching emergent communication with deep reinforcement learning. Prior to my MSc, I spent two years as a software engineer, I worked on platform engineering and internal tooling for production systems.

Why AI for Science?
I had a bit of background in chemistry. I worked on synthesizing silver nanoparticles for drug delivery, I realized the broader challenge: discovering new materials or drugs requires searching an impossibly large space of possibilities, where the overall process is slow and expensive to test. That holds true across most scientific domains and real-life systems.

It’s also why I think science is a good thing to point capable AI at. Discovery, whether exploring promising areas or producing insights, is one of the few applications where the value of AI is obvious and the failure mode is mostly wasted effort rather than harm.

Outside of work, you can find me playing piano, sudoku, chess, hiking, or scuba diving.

news

Apr 15, 2026	Started AI Engineer role at IBM
Apr 09, 2026	🥳 Paper podium presentation at Evolang 2026
Feb 01, 2026	Started an independent research project with my supervisors
Sep 01, 2025	Extending thesis work on emergent communication — created and manipulated datasets for training, trained and fine-tuned models on GPU.
Aug 01, 2025	Defended my master’s thesis, “Emergent Communication via Cognitive Graphs” — exploring how AI agents develop a communication protocol from scratch to solve a navigation task.

publications

Immuno-Metabolic Pathways: Investigating Mediators of Major Depressive Disorder and Atherosclerotic Cardiovascular Disease Comorbidity

Angela Koloi, Nabila P.R. Siregar, Rick Quax, and 7 more authors

Biological Psychiatry Global Open Science, 2025

Abs DOI

Background: Major depressive disorder (MDD) and cardiovascular diseases (CVD) often co-occur whereby comorbidity results in poorer clinical outcomes, presumably due to shared immune-metabolic pathways. Identifying shared biomarkers for MDD-CVD comorbidity may provide targets for prevention or treatment. Methods: Using data from the Netherlands Study of Depression and Anxiety (NESDA: N=2,256, 66.3% female, mean age 41.75 ± 13.11 years), validated with the UK Biobank data (UKB, N=35,668, 56.14% female, mean age 63.95 ± 7.74 years), this study aimed to identify (i) biomarkers, closely associated with current MDD, and (ii) longitudinal pathways linking MDD and atherosclerotic CVD. Plasma metabolites (NMR, Nightingale) and inflammatory markers were used as exposures within a Machine Learning framework. Influential biomarkers were integrated into a temporal network analysis linking MDD to subsequent CVD, exploring longitudinal pathways through Causal Discovery, validated by sensitivity analysis and centrality assessment. External validation included mediation and regression analysis adjusting for covariates. Results: Network analysis identified stable direct paths from MDD to CVD via TNF-α, tyrosine, and fatty acids, and indirect paths via acetate, HDL diameter, IL-6, alpha-1-acid glycoprotein (AGP), hs-CRP, and LDL triglycerides. Among these, acetate, tyrosine, AGP, and HDL diameter potentially mediate the MDD-CVD connection, as these were identified as key nodes within the network. UKB validation confirmed HDL diameter (β = 0.004) and AGP (β = 0.003) as significant depression-CVD mediators (both p < .001), after adjusting for age, sex, deprivation index, alcohol consumption, smoking status, physical activity, and BMI. Conclusions: These analyses identified biomarkers shared in MDD and CVD and may drive comorbid pathology risk.
An emergent communication framework for honeybee waggle dance

Nabila Siregar, Phong Le, and Raquel G Alhama

In International Conference on the Evolution of Language, 2026

Abs

We present a computational framework to investigate whether artificial agents in a foraging environment develop a communication protocol with features of the honeybee waggle dance. Analyzing communicative success, compositionality, and emergent semantics, we find that (i) agents tend to use directional signals for communication, especially when they have the freedom to explore the environment, (ii) in the absence of direct distance representations, agents develop a system based on shortest-path routes when their mental map representations overlap, and (iii) constraining their communication channel to simulate kinematic properties of honeybees provides a slight positive benefit. Overall, we find that, under certain constraints, agents develop a (relatively) compositional communication system reminiscent of the waggle dance.