Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Predicting Functional Brain Connectivity with Context-Aware Deep Neural Networks
Authors: Alexander Ratzan, Sidharth Goel, Junhao Wen, Christos Davatzikos, Erdem Varol
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We rigorously benchmark context-aware neural networks, including SMT and a single-gene resolution Multilayer-Perceptron (MLP), to established rules-based and bilinear methods. Context-aware neural networks outperform linear methods, significantly exceed our stringent null map estimates, and generalize across diverse connectomic datasets and parcellation resolutions. |
| Researcher Affiliation | Academia | 1Department of Computer Science, New York University 2Neuroscience Institute, Grossman School of Medicine, New York University 3Department of Radiology, Columbia University 4Department of Radiology, University of Pennsylvania Correspondence: EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 CGE-Matched Spatial Null Brain Map Generation |
| Open Source Code | Yes | Code to reproduce our results is available at: github.com/neuroinfolab/Gene Ex2Conn. |
| Open Datasets | Yes | Raw data is available at https://portal.brain-map.org/. UK Biobank (UKBB). The primary connectivity dataset in our study is a subset of n = 1814 healthy participants from the UK Biobank (mean age=63.3 years; age range=45-82; 55% female) with available resting-state functional MRI (rs-f MRI) [22]. Human-Connectome-Project Young-Adult (HCP-YA) [23] and Max Planck Institute Leipzig Mind-Brain-Body Dataset (MPI-LEMON) [24] are used as validation connectomic datasets. |
| Dataset Splits | Yes | To rigorously evaluate model performance, we employ two train-test split protocols: random and spatial. Each protocol is repeated across 10 rounds of 4-fold cross-validation, resulting in 40 total splits per model. For each random split, 75% of brain regions are used for training and 25% for testing, with models trained on the rtrain 2 training-region edges and evaluated on the disjoint rtest 2 test edges (Figure 2B). |
| Hardware Specification | Yes | All experiments are run on A100 or H100 NVIDIA GPUs on a high-performance compute cluster, taking no more than 1 hour per fold. |
| Software Dependencies | No | The paper mentions several software tools and packages like 'abagen package', 'XCP-D [54], implemented in Nipype [55]', and 'Flash Attention [36]' but does not provide specific version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | To select the optimal hyperparameter configurations for each model in Table 1 and Table 2, we perform nested inner cross-validation using random sampling from a hyperparameter grid defined from wider grid searches and manual fine-tuning. For each model, 3–6 hyperparameter combinations are sampled... All gradient-based models, including both linear baselines and MLP architectures, are trained using the Adam W optimizer. The mean squared error (MSE) between predicted and ground-truth connectivity values are used as the training loss. ... Table 4: Hyperparameter search space for the Spatiomolecular Transformer. Default values in bold. |