Shared data for Disinhibitory signaling enables flexible coding of top-down information

This dataset contains the source data underlying main and supplementary figures of Aquino, Kim, and Rungratsameetaweemana (2025), which investigates how biologically constrained recurrent neural networks (RNNs) integrate bottom-up sensory inputs with top-down task and attention signals during flexible decision-making. Overview The data were generated from RNN models trained on two dynamic decision-making tasks, a one-modality delayed match-to-sample (DMS) task and a two-modality DMS task, designed to dissociate bottom-up sensory inputs from top-down task and attention cues. Twenty RNNs were independently trained for each task variant, including single-module networks and hierarchical two-module networks composed of a sensory and a non-sensory module. The dataset captures model behavior, network architecture properties, neural dynamics, and the effects of targeted synaptic perturbations on task performance. Contents The repository is organized by figure, with each folder containing the processed source data (CSV or XLSX) used to generate the corresponding panel. The data span six categories: Model performance and behavior. Test accuracy of trained RNNs across early- and late-instruction conditions for both the one- and two-modality DMS tasks (Fig. 1C), along with comparisons of training trials and final accuracy between one-module and two-module networks (Fig. S15). Synaptic time constants. Optimized synaptic decay time constants for excitatory and inhibitory units in one-modality and two-modality RNNs (Fig. 3A, 3B), as well as in the sensory and non-sensory modules of two-module RNNs (Fig. 5D, 5E). Synaptic connection strengths. Average excitatory and inhibitory connection strengths within and between modules in the two-module RNNs (Fig. 5F, 5G; Fig. S16), and within- versus across-selectivity connection weights across rule-selective populations (Fig. S6), with counts of outgoing synaptic connections by cell-type and selectivity (Fig. S7). Lesion-based perturbation results. Task performance of all 20 trained RNNs after selectively reducing synaptic weights (50% lesions) targeted by presynaptic/postsynaptic cell type (E→E, E→I, I→E, I→I) and task- or attention-cue selectivity. These data span the one-modality task (Fig. 4C, S9), the two-modality task under different attended modalities (Fig. 4D, 4E, S8, S10–S13), and the two-module hierarchical networks across within-module, feedforward, and feedback connections (Fig. 6C–F, S18–S21). Latent trajectory analyses. Euclidean distances between CEBRA latent trajectories corresponding to opposite first-stimulus identities, computed separately for attended and unattended modalities (Fig. S2). File formats Data are provided as comma-separated values (.csv) and Excel workbooks (.xlsx). Each file maps to one or more figure panels as listed in the accompanying README, with several panels sharing source files when generated from a common analysis (e.g., the two-module lesion results in 6cdef).