A TOOLBOX FOR ROBUST ESTIMATES OF MULTIVARIATE INFORMATION FROM NEURAL DATA

Information theory has been widely adopted to study neural information processing and in particular to study how interactions between neurons or brain areas shape information processing. However, its application is often constrained by computational complexity, estimation problems due to limited-sampling biases, and scarcity of accessible up-to-date software.
We built MINT, an open-source toolbox for multivariate information analysis (Lorenz et al., PLoS Comput. Biol. 21, e1012934 (2025)), which provides state-of-the-art tools including partial information decomposition (PID), Transfer Entropy, Feature Specific Information Transfer, Intersection Information. PID disentangles synergistic, independent, and redundant information contributions. While limited-sampling biases are well known for Shannon information measures, their impact on PID remains under-explored. To address this, we use simulations and analytical calculations to characterise limited-sampling bias of synergy and redundancy in both discrete PID (spiking activity) and Gaussian PID (continuous signals such as MEG or fMRI). We demonstrate that two- and three-sources PID can be strongly biased. Synergy is consistently more affected than redundancy, with the difference increasing with the number of parameters specifying the probability distributions. We introduce effective bias-correction procedures and data-size guidelines, validate them on brain recordings from mice, monkeys and humans across two-photon calcium imaging, electrophysiology, MEG and fMRI, showing that they mitigate synergy overestimation.
Overall, MINT provides a comprehensive set of state-of-the-art information-theoretic methods for neural data spanning discrete (e.g., spike trains) and continuous signals (e.g., LFP, M/EEG, fMRI, calcium imaging), and incorporates validated PID bias-correction procedures and data-size guidelines to reduce synergy overestimation.