Explainable Artificial Intelligence (XAI)

Improving Data Quality, Model Transparency and Performance in Lung Histopathology with Explainable AI

This paper applies explainable AI techniques to improve transparency, data quality, and performance in lung histopathology image classification.

Biomarker Fidelity Score - A Quantitative Framework for Individual-Level Validation of Explainability Methods in 3D Alzheimer's Disease MRI Classification

The paper introduces a quantitative framework to validate explainability methods in Alzheimer's disease MRI classification at the individual level using biomarker-aligned spatial overlap scores.

Efficient Game-Theoretic Explanations for Tree-Based Ensembles via Owen Values

This paper introduces TreeOwen, algorithms for computing Owen values in tree-based ensembles to provide structured, multi-resolution explanations for grouped features in scientific applications.

Trustworthy in silico labeling via semantic visual interpretability of image-to-image translation

The paper introduces Mask Interpreter, a method to improve interpretability of image-to-image translation models, validating biological relevance and identifying artifacts in in silico labeling.

GDTR: Layer-wise Settling Depth Reveals Biological Grammar in Genomic Foundation Models

GDTR measures nucleotide settling depth in genomic models to reveal biological grammar stability across layers, showing splice sites settle earlier than intronic regions and identifying variant impact through differential depths.

Does ensembling improve feature attributions from sequence-to-activity models?

This study evaluates if ensembling improves feature attributions from sequence-to-activity models, finding that it enhances downstream applications like motif identification and variant prediction.

Benchmarking attention-based methods for vision transformers' interpretability in retinal fundus imaging

This study benchmarks attention-based interpretability methods for Vision Transformers in retinal fundus imaging, finding gradient-weighted attention rollout provides the most faithful and biologically relevant attributions.

VFUSE: Virulent Feature Understanding with Sparse autoEncoders

VFUSE uses sparse autoencoders to audit protein design models for hazardous features, improving interpretability without sacrificing performance.

Mechanistic Interpretability for Protein Language Models: A Validation Framework

This paper evaluates two mechanistic interpretability methods on a protein language model to identify features distinguishing Class A beta-lactamase from other classes.

Interpretable Predictive Modeling for Medical Data Using Boolean Rule-aware Regression

This paper introduces RBBR, a framework that generates interpretable Boolean rules for medical prediction, achieving strong performance and clinical relevance across six datasets.

The illusion of interpretability in biologically informed neural networks

This paper shows that biologically informed neural networks (BINNs) cannot reliably recover internal biological structures from outputs alone due to nonidentifiability.

TPCAV: Interpreting deep learning genomics models via concept attribution

This paper introduces TPCAV, a method to interpret genomics deep learning models by adapting concept attribution, showing it can analyze both motif features and general genomic concepts.

DVPNet: A New XAI-Based Interpretable Genetic Profiling Framework Using Nucleotide Transformer and Probabilistic Circuits

The paper introduces DVPNet, an XAI framework combining Nucleotide Transformer and probabilistic circuits to identify important genes in cancer classification using single-cell data.

What Do Biological Foundation Models Compute? Sparse Autoencoders from Feature Recovery to Mechanistic Interpretability

This paper reviews sparse autoencoders applied to biological foundation models, finding they recover interpretable biological features but lack experimental validation of mechanistic understanding.

How to gain valuable insight from scarce data with Machine Learning: a post-hoc explanation tool to identify biases in biological images classification

This study uses ML to classify regenerating vs. scarring mouse tissue images, finding models overfit to individual mice rather than biological features, but successfully extract relevant biological information when tasks align with data capabilities.

XAI-ecDNA: A Proof-of-Concept Framework for Explainable Multi-Modal Extrachromosomal DNA Detection Using Synthetic FISH and Genomics

This paper introduces XAI-ecDNA, a framework for detecting extrachromosomal DNA using synthetic FISH and genomic data, achieving high performance and interpretability.

The eXplainable Artificial Intelligence (XAI) Triad: Models, Importances, and Significance at Scale

This paper evaluates XAI methods for variable selection in biological data, finding that RF/Accuracy/PIMP performs best but is computationally intensive, while DT/Gini/simThresh offers a more scalable alternative.

Keeping SCORE enables interpretable uncertainty-aware classification from diffusion models for genomics

Keeping SCORE transforms diffusion models into probabilistic classifiers for genomics, providing accurate predictions with uncertainty and interpretability across multiple tasks.

Systematic evaluation of peptide property predictors with explainable AI technique SHAP

The paper evaluates peptide property predictors using SHAP to explain model predictions and finds strong correlations with known amino acid indices.

Decode-gLM: A toolkit for interpreting, auditing, and steering genomic language models

This paper introduces Decode-gLM, a toolkit using sparse autoencoders to interpret, audit, and steer genomic language models, revealing biological features and identifying data contamination.

A layer-resolved diagnostic identifies bias-driven decisions in deep neural networks

This paper introduces the Bias Dominance Index (BDI) to diagnose whether neural network decisions are feature-supported or bias-driven, showing that high confidence can coexist with bias-driven decisions.

DeepDiff-SHAP: Interpretable deep learning for subgroup-specific causal inference using conditional SHAP

DeepDiff-SHAP identifies subgroup-specific causal relationships in biomedical data using deep learning and SHAP, revealing clinically meaningful changes in feature interactions across patient subgroups.

Trustworthy detection of exencephaly in high-throughput micro-CT embryo screens with focal-loss transformers

The paper develops a focal-loss transformer model to detect exencephaly in imbalanced micro-CT embryo data, achieving high accuracy and interpretable saliency maps.

Identifying intervention strategies from machine learning models with COALA: a counterfactual optimization framework

COALA identifies optimal counterfactual interventions for individuals using gradient boosted trees on NHANES data, achieving 85.4% accuracy in predicting feature changes.

XAI-based Data Visualization in Multimodal Medical Data

This paper examines advanced XAI techniques for multimodal medical data, focusing on perturbation-based, concept-based, and example-based explanations to improve model interpretability in healthcare.

Adaptive Example Selection: Prototype-Based Explainability for Interpretable Mitosis Detection

This paper introduces Adaptive Example Selection (AES), an XAI framework for mitosis detection that enhances model interpretability by comparing test cases with annotated prototypes.

Biophysical validation of explainable AI for functional brain imaging: bridging cellular mechanisms and network dynamics

This paper validates an explainable AI framework to identify brain regions affected by excitation/inhibition imbalance using simulations and empirical data from autism.

Sensitivity based model agnostic scalable explanations of deep learning

This paper introduces SensX, an XAI framework that improves explanation accuracy and speed for deep learning models, validated on vision transformers and single-cell RNA-seq data.

Explainable machine learning for health disparities: type 2 diabetes in the All of Us research program

This paper uses explainable machine learning to identify and quantify T2D risk factors contributing to health disparities among different racial and ethnic groups in the All of Us research program.

Ensemble Post-hoc Explainable AI for Multilead ECG: Identifying Disease-Relevant Features in Single-Lead Interpretations

The paper evaluates XAI methods for a 1D-ResNet ECG classifier, finding some methods consistently highlight clinically relevant features while others do not.