The Effect of Preprocessing Choices on Single-Cell Foundation Models

Four scRNA-seq foundation models (scGPT, Geneformer, scFoundation, and scVI) were fine-tuned or trained under alternative denoising, normalization, gene-selection, value-encoding, and neighbor-graph choices. Changing any single step replaced 86% of a cell's nearest neighbors on average and relabeled up to 9% of held-out cells in cell-type prediction, with rare populations disproportionately affected; mismatched preprocessing between fine-tuning and inference could drop accuracy to chance.

Venue: NeurIPS 2026 Workshop on I Can't Believe It's Not Better: Failure Modes of AI in Biology (ICBINB)

Rich J, Pachter L. (2026). "The Effect of Preprocessing Choices on Single-Cell Foundation Models." NeurIPS 2026 Workshop on I Can't Believe It's Not Better: Failure Modes of AI in Biology (ICBINB).

RADAR: An Agentic Assistant for Renal Mass Assessment and Management Planning on CT

RADAR, a multimodal agentic assistant that pairs a large language model with CT image-analysis tools and retrieval-augmented generation to answer renal mass assessment and management questions (surveillance, biopsy, ablation, partial or radical nephrectomy) with its measurements and reasoning. On a 418-item benchmark built from 20 KiTS23 cases, RADAR with a Qwen2.5-7B model selected the reference answer in 80.6% of items, versus 63.4% for GPT-4o given lesion-centered slices.

Venue: NeurIPS 2026 Workshop on Agentic Intelligence for Medical Imaging and Multimodal Clinical Data (AIM)

Rich JM, Venkatesh S, Aminy M, Gill I, Weight C, Oberai A, et al. (2026). "RADAR: An Agentic Assistant for Renal Mass Assessment and Management Planning on CT." NeurIPS 2026 Workshop on Agentic Intelligence for Medical Imaging and Multimodal Clinical Data (AIM).

Comparative Effectiveness and Cost-Effectiveness of an Artificial Intelligence Workflow for Small Renal Mass Diagnosis on Computed Tomography

A cost-effectiveness modeling study of AI-assisted imaging for small renal masses, using a Markov microsimulation of 10,000 patients over a 10-year horizon to compare standard of care, an MRI+AI workflow, and two CT+AI workflows. The CT-based AI strategies yielded the highest quality-adjusted life years at the lowest cost — dominating standard of care and MRI+AI — suggesting AI risk stratification could reduce overtreatment cost-effectively; as a decision-analytic simulation, the findings are economic projections that motivate, rather than substitute for, prospective clinical validation.

Venue: Urology Practice

Rich JM, Bajakian T, Gala T, Aminy M, Forsyth E, et al. (2026). "Comparative Effectiveness and Cost-Effectiveness of an Artificial Intelligence Workflow for Small Renal Mass Diagnosis on Computed Tomography." Urology Practice.

The impact of package selection and versioning on single-cell RNA-seq analysis

A benchmarking study showing that the two dominant scRNA-seq pipelines, Seurat and Scanpy, produce substantially different results across filtering, dimensionality reduction, clustering, and differential expression - differences as large as those from sequencing under 5% of reads - and that even package version changes can alter outcomes, raising reproducibility concerns.

Venue: Cell Systems

Rich JM, Moses L, Einarsson PH, Luebbert L, Booeshaghi AS, Antonsson S, et al. (2026). "The impact of package selection and versioning on single-cell RNA-seq analysis." Cell Systems.

Analyzing foundation models for segmentation of osseous metastatic lesions in prostate cancer on CT scans

A comparative study pitting segmentation foundation models (SAM, SAM2, SAM-Med2D) - prompted with bounding boxes from human experts or from nnU-Netv2 predictions - against a supervised nnU-Netv2 baseline for outlining prostate-cancer bone metastases on CT, finding nnU-Netv2 still segments best but that prompt-guided foundation models are a viable low-data, low-compute alternative.

Venue: European Journal of Radiology Artificial Intelligence

Pawan SJ, Malewar S, Buren IV, Smith I, Rich J, Prajwal R, et al. (2025). "Analyzing foundation models for segmentation of osseous metastatic lesions in prostate cancer on CT scans." European Journal of Radiology Artificial Intelligence.

Image imputation with conditional generative adversarial networks captures clinically relevant imaging features on computed tomography

A study using a conditional GAN (trained on 333 patients) to synthesize any missing phase of a four-phase kidney CT from the other three, then showing the imputed images preserve clinically relevant features - over 85% agreement on all categorical imaging features plus key radiomic measures - as well as the real images do.

Venue: PLOS Digital Health

Rich J, Le J, Raad R, Tejura T, Rastegarpour A, Gill I, et al. (2025). "Image imputation with conditional generative adversarial networks captures clinically relevant imaging features on computed tomography." PLOS Digital Health.

Evaluation of nnU-Net for kidney tumor segmentation on a large external patient cohort

A generalizability study that trains nnU-Netv2 to segment kidneys and kidney tumors on CT and cross-tests it between the public KiTS19 dataset and a larger single-institution cohort, showing Dice accuracy drops substantially on external data (e.g., 0.85 to 0.66) and underscoring the need for diverse training data.

Venue: European Journal of Radiology Artificial Intelligence

Raman AG, Fisher D, Rich JM, Weight C, Heller N, Desai M, et al. (2025). "Evaluation of nnU-Net for kidney tumor segmentation on a large external patient cohort." European Journal of Radiology Artificial Intelligence.

An analysis of child abuse detected by skeletal surveys before and during the COVID-19 pandemic

A retrospective study of 479 children who underwent skeletal surveys at one academic hospital, comparing suspected physical-abuse cases before versus during the COVID-19 pandemic; overall abuse incidence stayed roughly the same, but intracranial and retinal hemorrhages dropped significantly during the pandemic, suggesting a shift in injury patterns or reporting.

Venue: Pediatric Discovery

Li VR, Pickering TA, Imagawa KK, Rich JM, Sura AS. (2025). "An analysis of child abuse detected by skeletal surveys before and during the COVID-19 pandemic." Pediatric Discovery.

Deep learning-based detection and segmentation of osseous metastatic prostate cancer lesions on computed tomography

A benchmarking study comparing deep-learning classifiers (EfficientNet, ResNet34, DenseNet) and segmentation networks (nnU-Netv2, U-Net, ResUNet, ResAttUNet) for automatically detecting and outlining prostate-cancer bone metastases on CT, finding EfficientNet best for detection and nnU-Netv2 best for segmentation (Dice 0.74) with radiomic features matching manual segmentations.

Venue: European Journal of Radiology Artificial Intelligence

Pawan SJ*, Rich J*, Malewar S, Patel D, Muellner M, et al. (2025). "Deep learning-based detection and segmentation of osseous metastatic prostate cancer lesions on computed tomography." European Journal of Radiology Artificial Intelligence.

Assessing the agreement of chronic lung disease of prematurity diagnosis between radiologists and clinical criteria

A study of 95 preterm infants comparing radiologists' chest X-ray diagnoses of chronic lung disease of prematurity against several standard clinical definitions, finding over 90% agreement and that radiologists often diagnose the condition, in the more severely affected infants, before the 36-week postmenstrual-age milestone those definitions use.

Venue: Maternal Health, Neonatology and Perinatology

Rich JM, Lin LJ, Le JL, Abe JRC, Sura A. (2024). "Assessing the agreement of chronic lung disease of prematurity diagnosis between radiologists and clinical criteria." Maternal Health, Neonatology and Perinatology.

A widely-used negative control formin mutant retains some actin polymerization activity

A biochemistry study finding that a formin mutant commonly relied on as an inactive negative control in actin-assembly experiments in fact retains some residual actin polymerization activity, complicating its use as a true zero-activity baseline.

Venue: UCLA MCDB Honors Thesis (Highest Honors)

Rich JM, Bradley AO, Quinlan ME. (2020). "A widely-used negative control formin mutant retains some actin polymerization activity." UCLA MCDB Honors Thesis (Highest Honors).

TCIA Radiology Image Processing for AI and Radiomics

A standardized, reproducible pipeline for turning heterogeneous multi-institutional CT data (e.g., from The Cancer Imaging Archive) into analysis-ready volumes for radiomics and deep learning. The notebook-based workflow handles series filtering, DICOM-to-NIfTI conversion, orientation and voxel-spacing harmonization, intensity normalization, and metadata validation, reducing the non-biological variability across scanners and protocols that otherwise destabilizes radiomic features and model performance; it is demonstrated on TCGA-KIRC renal-cancer scans.

Venue: medRxiv

Rich JM, Kang R, Jin D, Subramanian S, Duddalwar V, Pachter L. (2026). "TCIA Radiology Image Processing for AI and Radiomics." medRxiv.

fastQpick: scalable bootstrap and subsampling of FASTQ reads

A command-line tool and Python library for sampling sequencing reads with replacement, so a single FASTQ file can be resampled into many bootstrap replicates for uncertainty quantification at the level of raw reads — answering questions like whether a low-abundance call would survive resequencing. It streams large libraries efficiently (a full bootstrap of a 500-million-read library in under 30 minutes at 9.4 GB peak memory, or 1.4 GB in low-memory mode), and its replicates recover the multinomial sampling uncertainty of transcript abundance estimates to within a few percent on real yeast RNA-seq data.

Venue: bioRxiv

Rich JM, Pachter L. (2026). "fastQpick: scalable bootstrap and subsampling of FASTQ reads." bioRxiv.

Single-Cell Genomics Decontamination with CellSweep

CellSweep, a single-cell genomics decontamination tool that removes free-floating ambient molecules and bulk contamination introduced during library preparation; it works across a wide variety of platforms and outperforms existing decontamination methods across benchmarks.

Venue: bioRxiv

Caskey M*, Rich J*, Weber R, Mortazavi A, Pachter L, Hallgrimsdottir I. (2026). "Single-Cell Genomics Decontamination with CellSweep." bioRxiv.

Optimizing alluvial plots

An R package (wompwomp) for alluvial plots that formalizes the ordering and coloring of alluvia as optimization problems and adapts the NeighborNet algorithm from phylogenetics to produce clearer, less-cluttered layouts.

Venue: arXiv

Rich J, Oakes C, Pachter L. (2025). "Optimizing alluvial plots." arXiv.

Reference-based variant detection with varseek

varseek, a reference-based variant-detection tool that uses k-mer pseudoalignment to screen reads against an index of known variants, improving detection of indels and low-coverage variants and enabling variant calling at single-cell resolution (demonstrated on tumor-specific COSMIC variants in glioblastoma single-cell data).

Venue: bioRxiv

Rich JM, Luebbert L, Sullivan DK, Rosa R, Pachter L. (2025). "Reference-based variant detection with varseek." bioRxiv.

Geospatially informed representation of spatial genomics data with SpatialFeatureExperiment

SpatialFeatureExperiment, a Bioconductor R/S4 data class that extends SpatialExperiment with the Simple Features geospatial standard to store and run spatial operations on spatial-omics data from platforms like Visium, Xenium, MERFISH, SeqFISH, and Slide-seq within the SingleCellExperiment ecosystem.

Venue: bioRxiv

Moses L, Huseynov A, Rich JM, Pachter L. (2025). "Geospatially informed representation of spatial genomics data with SpatialFeatureExperiment." bioRxiv.

An upper bound on genomic recoverability from computed tomography in kidney cancer

Developed a mathematical model and software package to describe an upper bound on the recoverable genomic information from computed tomography (CT) scans in kidney cancer.

Venue: Radiological Society of North America (RSNA) 2026 Annual Meeting

Rich JM, Kang R, Jin DL, Pachter L. (2026). "An upper bound on genomic recoverability from computed tomography in kidney cancer." Radiological Society of North America (RSNA) 2026 Annual Meeting. Oral presentation, Cutting-Edge Research.

Assessment of Radiology Recommendation Adherence on Pediatric Ultrasound Reporting

Assessed how often follow-up recommendations made in pediatric ultrasound reports were adhered to in subsequent clinical care.

Venue: International Pediatric Radiology (IPR) 2026

Vu NM, Rich JM, Cen S, Sura AS. (2026). "Assessment of Radiology Recommendation Adherence on Pediatric Ultrasound Reporting." International Pediatric Radiology (IPR) 2026. Oral presentation.

Assessing the agreement of chronic lung disease of prematurity diagnosis between radiologists and clinical criteria

Described how radiologists' chest X-ray diagnoses of chronic lung disease of prematurity compare to several standard clinical definitions, finding over 90% agreement and that radiologists often diagnose the condition, in the more severely affected infants, before the 36-week postmenstrual-age milestone those definitions use.

Venue: Society for Pediatric Radiology (SPR) 2024 Annual Meeting

Rich JM, Lin LJ, Le JL, Abe JRC, Sura A. (2024). "Assessing the agreement of chronic lung disease of prematurity diagnosis between radiologists and clinical criteria." Society for Pediatric Radiology (SPR) 2024 Annual Meeting. Oral presentation.

Single-Cell Genomics Decontamination with CellSweep

Venue: Biology of Genomes, Cold Spring Harbor Laboratory

Caskey M*, Rich J*, Weber R, Mortazavi A, Pachter L, Hallgrimsdottir I. (2026). "Single-Cell Genomics Decontamination with CellSweep." Biology of Genomes, Cold Spring Harbor Laboratory. Poster presentation.

Reference-based variant detection with varseek

Venue: Cold Spring Harbor Laboratory Single Cell Analyses Conference

Joseph Rich, Laura Luebbert, Delaney Sullivan, Reginaldo Rosa, Lior Pachter. (2025). "Reference-based variant detection with varseek." Cold Spring Harbor Laboratory Single Cell Analyses Conference. Poster presentation.

Alluvial plot optimization with wompwomp

Venue: Cold Spring Harbor Laboratory Single Cell Analyses Conference

Joseph Rich, Conrad Oakes, Lior Pachter (2025). "Alluvial plot optimization with wompwomp." Cold Spring Harbor Laboratory Single Cell Analyses Conference. Poster presentation.

Contextualized surgical gestures in nerve-spare reveal heterogeneity in techniques and predicts erectile function recovery

Annotated nerve-sparing video from robot-assisted radical prostatectomies at four international centers with surgical gestures tied to their anatomical location and surgical task, revealing variation in technique across surgeons and using these contextualized gestures in a machine learning model to predict 12-month erectile function recovery.

Venue: American Urological Association (AUA) 2024 Annual Meeting

Ma R, Wang J, Rich J, Huang SL, Otiato MX, Dalieh I, You J, Goh AC, Sonn GA, Wagner C, Anandkumar A, Vickers AJ, Hu JC, Hung AJ. (2024). "Contextualized surgical gestures in nerve-spare reveal heterogeneity in techniques and predicts erectile function recovery." American Urological Association (AUA) 2024 Annual Meeting; J Urol 211(5S):e109. Moderated poster presentation (MP07-13).

Artificial intelligence and radiomics applied to prostate cancer bone metastasis imaging: a review

A scoping review of quantitative imaging methods - radiomics, machine learning, and deep learning - for detecting, diagnosing, and monitoring prostate cancer bone metastases, summarizing current approaches and calling for more clinically actionable tools.

Venue: IRADIOLOGY

Pawan SJ, Rich J, Le J, Yi E, Triche T, Goldkorn A, Duddalwar V. (2024). "Artificial intelligence and radiomics applied to prostate cancer bone metastasis imaging: a review." IRADIOLOGY.

Comprehensive systematic review of biomarkers in metastatic renal cell carcinoma: predictors, prognostics, and therapeutic monitoring

A systematic review (PubMed, 2017-2022) cataloging biomarkers in metastatic renal cell carcinoma that predict treatment response, indicate prognosis, or enable therapeutic monitoring across immune-checkpoint, targeted, and VEGF-inhibitor therapies, making the case for biomarker-guided personalized treatment.

Venue: Cancers

Dani KA*, Rich JM*, Kumar SS, Cen H, Duddalwar VA, D'Souza A. (2023). "Comprehensive systematic review of biomarkers in metastatic renal cell carcinoma: predictors, prognostics, and therapeutic monitoring." Cancers.

Deep learning image segmentation approaches for malignant bone lesions: a systematic review and meta-analysis

A PRISMA systematic review and meta-analysis of 41 studies on deep-learning segmentation of malignant bone lesions across CT, MRI, and PET/CT, finding that mostly U-Net-based models achieve strong accuracy (median Dice ~0.85-0.9) while dataset homogeneity and clinical generalization remain the key challenges.

Venue: Frontiers in Radiology

Rich JM, Bhardwaj LN, Shah A, Gangal K, Rapaka MS, Oberai AA, et al. (2023). "Deep learning image segmentation approaches for malignant bone lesions: a systematic review and meta-analysis." Frontiers in Radiology.

Artificial intelligence in pathomics and genomics of renal cell carcinoma

A review of how artificial intelligence applied to histopathology images (pathomics) and gene-expression data (genomics) is advancing renal cell carcinoma diagnosis, subtyping, and survival prediction, including identifying subtype-specific gene patterns that could guide targeted therapy.

Venue: Urologic Clinics of North America

Knudsen JE, Rich JM, Ma R. (2023). "Artificial intelligence in pathomics and genomics of renal cell carcinoma." Urologic Clinics of North America.

Localized multifocal retroperitoneal ganglioneuroma with an infiltrative appearance on imaging: a case report

A case report of a rare multifocal retroperitoneal ganglioneuroma that appeared aggressively infiltrative and fatty on CT yet proved benign on histology after resection, arguing it belongs in the differential for fatty retroperitoneal or mediastinal tumors and distinguishing it from lipomatous ganglioneuroma.

Venue: Case Reports in Oncology

Rich JM, Duddalwar VA, Ter-Oganesyan R, Hu P, Chopra S, Cheng PM, Aron M. (2023). "Localized multifocal retroperitoneal ganglioneuroma with an infiltrative appearance on imaging: a case report." Case Reports in Oncology.

Feminizing adrenocortical tumor with multiple recurrences: a case report

A case report of a 35-year-old man with a rare feminizing adrenocortical tumor (high estrogen, low testosterone, and gynecomastia) that recurred locally twice within a few years despite complete low-grade resection with clear margins, highlighting the high recurrence risk and need for close post-surgical surveillance.

Venue: Case Reports in Oncology

Rich JM, Duddalwar V, Cheng PM, Aron M, Daneshmand S. (2023). "Feminizing adrenocortical tumor with multiple recurrences: a case report." Case Reports in Oncology.

Editorial Commentary on our AI workflow for small renal mass diagnosis

An invited editorial commentary by urologists at Wake Forest University School of Medicine accompanying our Urology Practice study on the comparative effectiveness and cost-effectiveness of an AI workflow for diagnosing small renal masses on CT.

Venue: Urology Practice

Jabalameli N, Balaji S, Rais-Bahrami S. (2026). "Editorial Commentary." Urology Practice, 13(5), 491-492.

Package selection and versioning matter

A Nature Methods Research Highlight of our Cell Systems study showing that the choice of single-cell RNA-seq package (Seurat vs. Scanpy) - and even the package version - can substantially alter analysis results.

Venue: Nature Methods

Tang L. (2026). "Package selection and versioning matter." Nature Methods, 23(5), 872.

Lior Pachter on package selection and versioning in single-cell RNA-seq

Tweet about my manuscript 'The impact of package selection and versioning on single-cell RNA-seq analysis' comparing Seurat and Scanpy received 580,000+ views on X (formerly Twitter), shared by Lior Pachter.

Venue: X (formerly Twitter)

Pachter L [@lpachter]. (2024, April 5). Post on X (formerly Twitter).