Transforming Behavioral Neuroscience Discovery with In-Context Learning and AI-Enhanced Tensor Methods

Transforming Behavioral Neuroscience Discovery with In-Context Learning and AI-Enhanced Tensor Methods

Overview

This paper presents an AI-enhanced pipeline for accelerating discovery in behavioral neuroscience research, with a focus on understanding fear discrimination and generalization in mice. The key components of the pipeline are:

• In-Context Data Preparation: Leveraging In-Context Learning (ICL) to automate tedious data preparation tasks, such as annotating mouse behaviors from videos.
• AI-Enhanced Tensor Analysis: Applying neural tensor decomposition methods to extract latent patterns from coupled neuroscience datasets (neural activity and behavioral observations).
• AI-Driven Pattern Interpretation: Using language models and retrieval-augmented generation to assist domain experts in interpreting the discovered latent patterns.

The authors demonstrate the effectiveness of this pipeline through extensive experiments and a case study on real neuroscience data, showing how it can transform a traditionally complex and manual research workflow into one that empowers domain experts to focus on scientific discovery.

Methodology

In-Context Data Preparation

• Identified ICL as a promising approach to automate data preparation tasks for domain experts without requiring extensive AI/ML expertise.
• Introduced "Autoregressive ICL" (AR-ICL) to capture temporal continuity in mouse behavior annotations, improving performance over standard ICL.
• Also explored using ICL for coarse extraction of neural activity from calcium imaging data, but found performance limited by the fine-grained nature of the signals.

AI-Enhanced Tensor Analysis

• Leveraged tensor decomposition methods, including CANDECOMP/PARAFAC decomposition (CPD) and the more advanced Neural Additive Tensor Decomposition (NeAT).
• NeAT achieves 46% lower test RMSE than CPD on the neuroscience datasets, demonstrating the benefit of its non-linear modeling capabilities.
• Coupled the neural activity and behavioral tensors to discover shared and unshared latent patterns.

AI-Driven Pattern Interpretation

• Designed a system to automatically generate plausible, literature-grounded hypotheses for the discovered latent patterns.
• Used a combination of In-Context Learning and Retrieval-Augmented Generation to leverage both visual information (factor plots) and relevant scientific literature.
• Demonstrated moderate agreement between model-generated interpretations and expert annotations, suggesting the system can serve as a useful interpretive aid.

Data & Experimental Setup

• Collected neural activity data using one-photon calcium imaging in the prelimbic region of the mouse medial prefrontal cortex during a fear conditioning and discrimination task.
• Simultaneously recorded mouse behaviors, which were manually annotated by domain experts as freezing, fleeing, or grooming/exploring.
• Datasets comprised 33 trials across 7 mice, with neural activity tensors of size (33 x 6000 x Ns) and behavior tensors of size (33 x 6000), where Ns ranged from 948 to 2339 neurons per subject.

Results

Behavioral Video Labeling

• AR-ICL achieved the best performance, with a macro F1 score of 0.545, balanced accuracy of 0.801, and MCC of 0.517.
• Outperformed a DINO-V2 baseline as well as standard ICL and a temporal context-only variant.

Tensor Decomposition

• NeAT outperformed the simpler CPD model, achieving 46% lower test RMSE on the coupled neural and behavioral tensors.
• NeAT was able to capture non-linear interactions between the data modalities more effectively.

AI-Driven Interpretation

• The language model-based interpretation system achieved moderate agreement with expert annotations, with a weighted Cohen's kappa of 0.59.
• While not as detailed as expert interpretations, the model-generated hypotheses provided a useful starting point for guiding scientific analysis.

Limitations & Uncertainties

• The ICL-based calcium imaging preprocessing approach had limited performance, suggesting the need for more sophisticated methods to handle the fine-grained and noisy neural signals.
• The interpretability of the tensor decomposition results, while enhanced by the NeAT model, is still dependent on the domain experts' ability to link the latent factors to neuroscientific concepts.
• The evaluation of the end-to-end pipeline was limited to a small dataset of two subjects, and further validation on larger-scale datasets is needed.

Future Work

The authors highlight several directions for future work:

• Investigating more advanced techniques for incorporating temporal dynamics into the ICL-based behavioral labeling.
• Exploring neural network architectures that can better capture the spatial and temporal structure of calcium imaging data.
• Expanding the AI-driven interpretation system to support more sophisticated reasoning and hypothesis generation capabilities.
• Validating the effectiveness of the full pipeline on larger-scale behavioral neuroscience datasets and across different experimental paradigms.