Biology subjects

Ramesh, S.

Publications and source records attributed to Ramesh, S..

At least 19 recordsLinked to original sources

bioRxiv2026

FEATMAP: Targeted Correction of Acquisition Signatures Harmonizes Medical Foundation Model Embeddings and Enables Robust Task Generalization

AO_SCPLOWBSTRACTC_SCPLOWMedical foundation models compress biomedical data into embeddings that support diverse downstream clinical tasks. However, successful model deployment is hampered by performance degradation on external data. It is recognized that embeddings capture acquisition signatures, such as hardware and technical differences, in addition to biology. Effective harmonization must remove the acquisition signature while preserving biological signals, a trade-off that current methods fail to balance adequately. Input-level normalization fails to eliminate acquisition signatures from embeddings, whereas embedding-level methods adjust features in an untargeted manner. We present FEATMAP, a harmonization approach that models acquisition signatures as geometric distortions between manifolds of similarly arranged embeddings. Using paired data that isolate the effect of acquisition signatures, FEATMAP fits a single global affine transformation per foundation model to correct acquisition signatures directly in the embedding space. This targeted, reusable correction aims to preserve biological and demographic variation while harmonizing across acquisition signatures. Across scanner and foundation-model harmonization in digital pathology and field-strength harmonization in brain MRI, FEATMAP improves cross-condition embedding similarity, reduces performance gaps without retraining, and suggests potential for the alignment of disparate embedding spaces.

bioinformatics↗

bioRxiv2026

Pervasive cryptic selection in the human noncoding genome

The prevailing dogma in evolutionary genetics holds that mutations within sequences that are conserved across a phylogeny are deleterious in those species, and mutations outside are neutrally evolving. Indeed, such comparative genomic approaches have estimated that mutations in approximately 5% of the human genome experience negative selection. However, sites that have biological function in certain lineages but not in others, i.e. functional turnover, may violate this assumption since these sites may be invisible to comparative genomic approaches. Thus, the extent of such cryptic, or hidden, negative selection remains elusive. Here, we developed a statistical test to detect cryptic selection in human polymorphism data. Applying our approach to simulated data shows that cryptic selection shapes the site frequency spectrum (SFS) and the statistical detection power depends on the proportion of mutations experiencing cryptic selection, the amount of sequence tested, and the sample size. We applied our method to polymorphism data from the 1000 Genomes Project, comparing variants in putatively functional noncoding regions to those in putatively neutral regions. We detected pervasive signals of cryptic selection in putatively functional regions, even after filtering out the top 70% of conserved sites. Using simulations with varying levels of cryptic selection, we estimated the extent of genome-wide constraint in the human genome. Our approximation suggests that mutations in at least 7% of the human genome are under negative selection, which is greater than the estimates from conservation-based methods, and that many of these mutations have escaped detection by comparative genomic methods. In sum, our results highlight the evolutionary dynamic nature of the noncoding genome and suggest the need to account for functional turnover when identifying putatively neutral variants for evolutionary analyses.

evolutionary biology↗

bioRxiv2026

Doxycycline inhibits both apicoplast and mitochondrial translation in apicomplexan parasites

Doxycycline is a tetracycline-class antibiotic used for malarial prophylaxis and as an occasional partner drug in malaria treatment. Doxycyclines antimalarial mechanism of action is attributed to inhibiting the prokaryotic 70S ribosomes of the apicoplast, a non-photosynthetic plastid present in most apicomplexan species including Plasmodium and Toxoplasma. At lower concentrations (<5 M) doxycycline exhibits a delayed death phenotype, typical of inhibitors of apicoplast housekeeping processes. Unlike other delayed death drugs (e.g. clindamycin), doxycycline has rapid schizonticidal activity at higher concentrations (>10 M) via an unknown and apicoplast-independent mechanism. In other eukaryotes, and plausibly in Plasmodium and Toxoplasma, doxycycline inhibits mitochondrial 70S ribosomes. Here we use a mass spectrometry approach to assess steady state and turnover for apicoplast DNA encoded proteins in Plasmodium falciparum. We directly show that these proteins decrease in both abundance and synthesis following treatment with doxycycline and clindamycin. High concentrations of doxycycline, but not clindamycin, also reduced the abundance of a mitochondrial-encoded protein. Proteins encoded by the mitochondrial genome are required for the formation of complexes III and IV in the electron transport chain. We show a reduction in abundance and activity of complex IV following doxycycline treatment in Toxoplasma gondii. High doxycycline concentrations also disrupt oxidative phosphorylation in both P. falciparum and T. gondii, likely as a consequence of impaired complex III and IV formation. For the first time we directly measure apicoplast translation inhibition and characterise doxycycline as a mitochondrial translation inhibitor of P. falciparum and T. gondii.

microbiology↗

bioRxiv2026

Structural basis of caveolin-driven membrane bending

Caveolins are monotopic membrane proteins essential for caveolae formation and have key roles in signaling and lipid regulation. Caveolins assemble into amphipathic discs with a central {beta}-barrel, an architecture distinct from other membrane-remodeling proteins. These discs embed in the membrane inducing membrane curvature. However, the mechanism of disc-driven bending remains unresolved. Using cryo-electron tomography, structure-guided mutagenesis, and mammalian cell studies, we show that evolutionarily distinct caveolins differ dramatically in their ability to curve membranes despite their conserved architecture. Through computational and theoretical analyses, we demonstrate that patterning of hydrophobic residues along the rim of the disc of human Caveolin-1 induces the deformation of the surrounding leaflet, dictating membrane bending. Finally, we determine a 4.1[A] resolution structure of Caveolin-1 within heterologous caveolae in situ, showing the disc adopts a funnel-like conformation, further shaping membrane architecture. Together, these findings reveal fundamental structural principles that empower caveolins to sculpt and remodel cellular membranes.

cell biology↗

bioRxiv2025

Comparative circadian transcriptome analysis reveals dampened and phase-advanced rhythms in sun-exposed human skin

BackgroundDaily molecular rhythms modulate skin physiology. However, the effects of chronic sun exposure on these rhythms remain unstudied. ObjectivesThis study aimed to identify and compare rhythmic genes and pathways in photoprotected and chronically photoexposed human skin in vivo. MethodsTwenty healthy White women, aged 51-63, with moderate-severe photoageing of the dorsal forearm were recruited. Skin biopsies (3mm) were taken from photoprotected (upper buttock) and photoexposed (dorsal forearm) skin of each individual at noon, 6PM, midnight, and 6AM, across a 24-hour cycle. Skin biopsies were analysed by RNA sequencing. Cosinor analysis identified cycling genes along with their amplitudes and peak expression phases. Nested models were used to identify genes that were differentially rhythmic between the photoprotected and photoexposed sites. Phase set and gene set analyses identified pathways enriched among rhythmic transcripts or altered between the two sites. ResultsIn the photoprotected buttock skin, 1546 genes (12%) met the criteria for cycling. In photoexposed forearm skin, the number was reduced to 959 (8%). As a group, transcripts that cycled in both sites had overall higher amplitude in photoprotected skin (p < 2.2e-16). Peak expression times for these transcripts showed a pronounced bimodal distribution and were clustered in the early morning and mid-afternoon. Distributions of peak times were significantly different between photoexposed and photoprotected skin (p < 0.00025), with peak times advanced in photoexposed skin. We identified 480 genes with significantly different rhythmic properties between the skin sites. Genes involved in DNA repair, MYC targets, E2F and G2M checkpoint pathways were enriched among those that showed higher amplitude oscillations in photoprotected skin. Genes involved in epithelial mesenchymal transition and apical junction pathways showed higher amplitude oscillations in photoexposed skin. ConclusionsTemporal rhythms have a marked influence on skin molecular physiology and are altered in photoaged skin. Temporally advanced cycling patterns and a reduced number of rhythmic genes in photoexposed as compared to photoprotected skin suggest that chronic UV exposure may disrupt and/or reprogram circadian output rhythms to further alter skin physiology.

cell biology↗

bioRxiv2025

Shared and distinct sequence-function signatures define different modes of human TpoR activation

The human thrombopoietin receptor (hTpoR) exists primarily as JAK2-associated monomers that become activated when converted to dimeric forms that support JAK trans-phosphorylation. This can be achieved by several different modes of stimuli, including the natural ligand Tpo, biologic agonists that bind the same site as Tpo, small-molecule drugs that bind the transmembrane (TM) domain, oncogenic mutations in and near the TM domain, and by association with constitutively active JAK V617F or a mutant form of the chaperone protein calreticulin. It is unclear how the dimeric structures induced by synthetic agonists and mutations relate to one another, and whether any of these induce the same active structure as the native ligand Tpo, yet this has important implications both for fundamental cytokine receptor biology and for development of targeted interventions for hTpoR-driven myeloproliferative diseases. Here we used deep mutational scanning (DMS) across the TM and juxtamembrane (JM) regions of hTpoR to extract feature-rich sequence-function signatures across a variety of different activating contexts. While each displayed some unique features, synthetic agonists and activating mutations all exhibited strong dependence on a common TM interface that is consistent with previous models of a left-handed, near-parallel helix dimer with H499 facing lipid. In contrast, Tpo-mediated activation was broadly insensitive to TM-JM substitutions, indicating that it does not rely on the same interface. Modeling with AlphaFold 3 (AF3) consistently yielded a right-handed, "splayed" helix dimer that is close at the extracellular face, contains H499 in the interface and diverges toward the cytosolic face, resting on an intracellular amphipathic JM helix that lies parallel to the membrane, which is also observed in a DMS/AF3 analysis of human erythropoietin receptor. This splayed Tpo-bound dimer could be stably inserted into a lipid bilayer with associated JAK2 using molecular dynamics and is supported by experiments showing that most or all of the TM domain can be replaced by poly-valine, with little effect on Tpo-driven activation but catastrophic effects on responses to synthetic ligands. Our data support at least two different structural modes of hTpoR activation that reconcile prior biochemical models, rationalize patient variants, and inform mechanism-based agonist and antagonist design.

biochemistry↗

bioRxiv2025

The landscape of fitness effects of putatively functional noncoding mutations in humans

While annotations of noncoding regions in the human genome are increasing, the fitness effects of mutations in these regions remain unclear. Here, we leverage these functional genomic annotations and human polymorphism data to infer the distributions of fitness effects of new noncoding mutations in humans. Our novel approach controls for mutation rate variation and linked selection along the genome. We find distinct patterns of selection in putative enhancers, promoters, and conserved noncoding regions. While mutations in enhancers are often neutral, approximately 30% of mutations in promoters are deleterious. The most conserved noncoding regions, showing reduced divergence across mammals and primates, have the highest proportion of deleterious mutations. Notably, while we infer the most conserved sites across mammals and primates are enriched for deleterious mutations, such conserved sites only account for a minority of the deleterious mutations in noncoding regions. For example, the top 5% of conserved noncoding sites encompass fewer than 20% of deleterious mutations, indicating that functional noncoding regions vary widely in the distribution of their evolutionary constraint. Our findings highlight the dynamic evolution of gene regulation and shifting selection pressures over deep evolutionary timescales. Consistent with this finding, we infer mutations in [~]7-9% of the noncoding genome are deleterious. These insights have broad implications for using comparative genomics to identify non-neutrally evolving sequences in the human genome.

evolutionary biology↗

bioRxiv2024

Vitrification and rapid rewarming of precision-cut liver slices for pharmacological and biomedical research

Background and AimsHigh-throughput in vitro pharmacological toxicity testing is essential for drug discovery. Precision-cut liver slices (PCLS) provide a robust system for screening that is more representative of the complex 3D structure of the whole liver than isolated hepatocytes. However, PCLS are not available as off-the-shelf products, significantly limiting their translational potential. Cryopreservation could solve this bottleneck by effectively preserving PCLS indefinitely until their time of use. Conventional cryopreservation (slow cooling in DMSO-forming ice) results in poor PCLS viability and function and, therefore, has proven unsuitable. Here, we explore an "ice-free" cryopreservation approach called vitrification and focus on culturing and assessing PCLS for 3 days post-vitrification and rewarming, given that most acute drug toxicity tests are conducted over 24h. MethodsRat liver slices were diffusively loaded with a cryoprotective agent (CPA) cocktail consisting of EG and Sucrose. The CPA-loaded PCLS were placed on a polymer cryomesh, vitrified in liquid nitrogen (LN2), and rapidly rewarmed in CPA. The vitrified and rewarmed PCLS were subsequently cultured in a controlled volume of serum-free, chemically defined media for 3 days. ResultsThe cryopreserved PCLS maintained high viability, morphology, function, enzymatic activity, and drug toxicity response. Results show that the vitrified PCLS perform comparably to untreated controls and significantly outperform conventionally cryopreserved PCLS in all assessments (p < 0.05). ConclusionsRapid vitrification and rewarming of PCLS using cryomesh enabled successful preservation and culture. This approach maintained high viability, function, enzymatic activity, and drug response for 3 days in culture, similar to controls. Impact and ImplicationsThe implications of using vitrification to store PCLS are extensive. This technology provides an exciting opportunity for the development of an "off-the-shelf" cold supply chain of human PCLS from organs declined for transplant, which are sliced, cryopreserved, and stored in a repository and available for on-demand shipping for industrial and academic biomedical research. This would also allow PCLS to become a scalable, reproducible, wide-ranging, and population-representative source of tissue that can accurately mimic in-vivo conditions of the human liver. These transformative technologies could revolutionize our practice in studying not just the metabolism of drugs but also increase our capacity to study the zonal progression of many liver diseases and conduct other exciting biomedical research.

bioengineering↗

bioRxiv2024

TrkA+ sensory neurons regulate osteosarcoma proliferation and vascularization to promote disease progression

Bone pain is a presenting feature of bone cancers such as osteosarcoma (OS), relayed by skeletal-innervating peripheral afferent neurons. Potential functions of tumor-associated sensory neurons in bone cancers beyond pain sensation are unknown. To uncover neural regulatory functions, a chemical-genetic approach in mice with a knock-in allele for TrkA was used to functionally perturb sensory nerve innervation during OS growth and disease progression. TrkA inhibition in transgenic mice led to significant reductions in sarcoma-associated sensory innervation and vascularization, tumor growth and metastasis, and prolonged overall survival. Single-cell transcriptomics revealed that sarcoma denervation was associated with phenotypic alterations in both OS tumor cells and cells within the tumor microenvironment, and with reduced calcitonin gene-related peptide (CGRP) and vascular endothelial growth factor (VEGF) signaling. Multimodal and multi-omics analyses of human OS bone samples and human dorsal root ganglia neurons further implicated peripheral innervation and neurotrophin signaling in OS tumor biology. In order to curb tumor-associated axonal ingrowth, we next leveraged FDA-approved bupivacaine liposomes leading to significant reductions in sarcoma growth, vascularity, as well as alleviation of pain. In sum, TrkA-expressing peripheral neurons positively regulate key aspects of OS progression and sensory neural inhibition appears to disrupt calcitonin receptor signaling (CALCR) and VEGF signaling within the sarcoma microenvironment leading to significantly reduced tumor growth and improved survival. These data suggest that interventions to prevent pathological innervation of osteosarcoma represent a novel adjunctive therapy to improve clinical outcomes and survival.

cancer biology↗

bioRxiv2024

Generative Adversarial Networks Accurately Reconstruct Pan-Cancer Histology from Pathologic, Genomic, and Radiographic Latent Features

Artificial intelligence models have been increasingly used in the analysis of tumor histology to perform tasks ranging from routine classification to identification of novel molecular features. These approaches distill cancer histologic images into high-level features which are used in predictions, but understanding the biologic meaning of such features remains challenging. We present and validate a custom generative adversarial network - HistoXGAN - capable of reconstructing representative histology using feature vectors produced by common feature extractors. We evaluate HistoXGAN across 29 cancer subtypes and demonstrate that reconstructed images retain information regarding tumor grade, histologic subtype, and gene expression patterns. We leverage HistoXGAN to illustrate the underlying histologic features for deep learning models for actionable mutations, identify model reliance on histologic batch effect in predictions, and demonstrate accurate reconstruction of tumor histology from radiographic imaging for a virtual biopsy.

cancer biology↗

bioRxiv2023

Bioenergetic mapping of 'healthy microbiomes' via compound processing potential imprinted in gut and soil metagenomes

Despite mounting evidence of their importance in human health and ecosystem functioning, the definition and measurement of healthy microbiomes remain unclear. More advanced knowledge exists on health associations for compounds used or produced by microbes. Environmental microbiome exposures (especially via soils) also help shape, and may supplement, the functional capacity of human microbiomes. Given the synchronous interaction between microbes, their feedstocks, and micro-environments, with functional genes facilitating chemical transformations, there exists an intriguing opportunity to examine microbiomes in terms of their capacity to process compounds relevant to human health. Here we integrate functional genomics and biochemistry frameworks to derive new quantitative measures of the potential for human gut and environmental soil metagenomes to process major compound classes (e.g., lipids, carbohydrates) and selected biomolecules (e.g., vitamins, short-chain fatty acids) linked to human health. Metagenome functional potential profile data were translated into a universal compound mapping landscape based on bioenergetic van Krevelen mapping of function-level meta-compounds and corresponding functional relative abundances, reflecting imprinted genetic capacity of microbiomes to metabolize an array of different compounds. We show that measures of compound processing potential associated with human health and disease (examining atherosclerotic cardiovascular disease, colorectal cancer, type 2 diabetes and anxious-depressive behavior case studies), and displayed seemingly predictable shifts along gradients of ecological disturbance in plant-soil ecosystems (three case studies). Ecosystem quality explained 60-92% of variation in soil metagenome compound processing potential measures in a post-mining restoration case study dataset. With growing knowledge of the varying proficiency of environmental microbiota to process human health associated compounds, we might design environmental interventions or nature prescriptions to modulate our exposures, thereby advancing microbiota-oriented approaches to human health. Compound processing potential offers a simplified, integrative approach for applying metagenomics in ongoing efforts to understand and quantify the role and linkages of microbiota in environmental- and human-health.

bioinformatics↗

bioRxiv2023

MET Inhibitor Capmatinib Radiosensitizes MET Exon 14-Mutated and MET-Amplified Non-Small Cell Lung Cancer

PurposeThe objective of this study was to investigate the effects of inhibiting the MET receptor with capmatinib, a potent and clinically relevant ATP-competitive tyrosine kinase inhibitor, in combination with radiation in MET exon 14-mutated and MET-amplified non-small cell lung (NSCLC) cancer models. Methods and MaterialsIn vitro effects of capmatinib and radiation on cell proliferation, colony formation, MET signaling, apoptosis, and DNA damage repair were evaluated. In vivo tumor responses were assessed in cell line xenograft and patient-derived xenograft models. Immunohistochemistry (IHC) was used to confirm in vitro results. ResultsIn vitro clonogenic survival assays demonstrated radiosensitization with capmatinib in both MET exon 14-mutated and MET-amplified NSCLC cell lines. No radiation-enhancing effect was observed in MET wild-type NSCLC and human bronchial epithelial cell line. Minimal apoptosis was detected with the combination of capmatinib and radiation. Capmatinib plus radiation compared to radiation alone resulted in inhibition of DNA double-strand break repair as measured by prolonged expression of {gamma}H2AX. In vivo, the combination of capmatinib and radiation significantly delayed tumor growth compared to vehicle control, capmatinib alone, or radiation alone. IHC indicated inhibition of phospho-MET and phospho-S6 and a decrease in Ki67 with inhibition of MET. ConclusionsInhibition of MET with capmatinib enhanced the effect of radiation in both MET exon 14-mutated and MET-amplified NSCLC models.

cancer biology↗

bioRxiv2023

Reference Sequence Browser: An R application with a User-Friendly GUI to rapidly query sequence databases

Land managers, researchers, and regulators increasingly utilize environmental DNA (eDNA) techniques to monitor species richness, presence, and absence. In order to properly develop a biological assay for eDNA metabarcoding or quantitative PCR, scientists must be able to find not only reference sequences (previously identified sequences in a genomics database) that match their target taxa but also reference sequences that match non-target taxa. Determining which taxa have publicly available sequences in a time-efficient and accurate manner currently requires computational skills to search, manipulate, and parse multiple unconnected DNA sequence databases. Our team iteratively designed a Graphic User Interface (GUI) Shiny application called the Reference Sequence Browser (RSB) that provides users efficient and intuitive access to multiple genetic databases regardless of computer programming expertise. The application returns the number of publicly accessible barcode markers per organism in the NCBI Nucleotide, BOLD, or CALeDNA CRUX Metabarcoding Reference Databases. Depending on the database, we offer various search filters such as min and max sequence length or country of origin. Users can then download the FASTA/GenBank files from the RSB web tool, view statistics about the data, and explore results to determine details about the availability or absence of reference sequences.

bioinformatics↗

bioRxiv2023

Immune Responses in Checkpoint Myocarditis Across Heart, Blood, and Tumor

Immune checkpoint inhibitors (ICIs) are widely used anti-cancer therapies that can cause morbid and potentially fatal immune-related adverse events (irAEs). ICI-related myocarditis (irMyocarditis) is uncommon but has the highest mortality of any irAE. The pathogenesis of irMyocarditis and its relationship to anti-tumor immunity remain poorly understood. We sought to define immune responses in heart, tumor, and blood during irMyocarditis and identify biomarkers of clinical severity by leveraging single-cell (sc)RNA-seq coupled with T cell receptor (TCR) sequencing, microscopy, and proteomics analysis of 28 irMyocarditis patients and 23 controls. Our analysis of 284,360 cells from heart and blood specimens identified cytotoxic T cells, inflammatory macrophages, conventional dendritic cells (cDCs), and fibroblasts enriched in irMyocarditis heart tissue. Additionally, potentially targetable, pro-inflammatory transcriptional programs were upregulated across multiple cell types. TCR clones enriched in heart and paired tumor tissue were largely non-overlapping, suggesting distinct T cell responses within these tissues. We also identify the presence of cardiac-expanded TCRs in a circulating, cycling CD8 T cell population as a novel peripheral biomarker of fatality. Collectively, these findings highlight critical biology driving irMyocarditis and putative biomarkers for therapeutic intervention.

immunology↗

bioRxiv2023

Validating a low-cost, open-source, locally manufactured workstation and computational pipeline for automated histopathology evaluation using deep learning

Deployment and access to state-of-the-art diagnostic technologies remains a fundamental challenge in providing equitable global cancer care to low-resource settings. The expansion of digital pathology in recent years and its interface with computational biomarkers provides an opportunity to democratize access to personalized medicine. Here we describe a low-cost platform for digital side capture and computational analysis composed of open-source components. The platform provides low-cost ($200) digital image capture from glass slides and is capable of real-time computational image analysis using an open-source deep learning (DL) algorithm and Raspberry Pi ($35) computer. We validate the performance of deep learning models performance using images captured from the open-source workstation and show similar model performance when compared against significantly more expensive standard institutional hardware.

pathology↗

bioRxiv2023

McrD binds asymmetrically to methyl-coenzyme M reductase improving active site accessibility during assembly

Methyl-coenzyme M reductase (MCR) catalyzes the formation of methane and its activity accounts for nearly all biologically produced methane released into the atmosphere. The assembly of MCR is an intricate process involving the installation of a complex set of post-translational modifications and the unique Ni porphyrin cofactor F430. Despite decades of research, details of MCR assembly remain largely unresolved. Here, we report the structural characterization of MCR in two intermediate states of assembly. These intermediate states lack one or both F430 cofactors and form complexes with the previously uncharacterized McrD protein. McrD is found to bind asymmetrically to MCR, displacing large regions of the alpha subunit and increasing active site accessibility for the installation of F430--shedding light on the assembly of MCR and the role of McrD therein. This work offers crucial information for the expression of MCR in a heterologous host and provides new targets for the design of MCR inhibitors. One-sentence summaryStructural characterization of methyl-coenzyme M reductase assembly intermediates.

biochemistry↗

bioRxiv2022

Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers

BackgroundLarge language models such as ChatGPT can produce increasingly realistic text, with unknown information on the accuracy and integrity of using these models in scientific writing. MethodsWe gathered ten research abstracts from five high impact factor medical journals (n=50) and asked ChatGPT to generate research abstracts based on their titles and journals. We evaluated the abstracts using an artificial intelligence (AI) output detector, plagiarism detector, and had blinded human reviewers try to distinguish whether abstracts were original or generated. ResultsAll ChatGPT-generated abstracts were written clearly but only 8% correctly followed the specific journals formatting requirements. Most generated abstracts were detected using the AI output detector, with scores (higher meaning more likely to be generated) of median [interquartile range] of 99.98% [12.73, 99.98] compared with very low probability of AI-generated output in the original abstracts of 0.02% [0.02, 0.09]. The AUROC of the AI output detector was 0.94. Generated abstracts scored very high on originality using the plagiarism detector (100% [100, 100] originality). Generated abstracts had a similar patient cohort size as original abstracts, though the exact numbers were fabricated. When given a mixture of original and general abstracts, blinded human reviewers correctly identified 68% of generated abstracts as being generated by ChatGPT, but incorrectly identified 14% of original abstracts as being generated. Reviewers indicated that it was surprisingly difficult to differentiate between the two, but that the generated abstracts were vaguer and had a formulaic feel to the writing. ConclusionChatGPT writes believable scientific abstracts, though with completely generated data. These are original without any plagiarism detected but are often identifiable using an AI output detector and skeptical human reviewers. Abstract evaluation for journals and medical conferences must adapt policy and practice to maintain rigorous scientific standards; we suggest inclusion of AI output detectors in the editorial process and clear disclosure if these technologies are used. The boundaries of ethical and acceptable use of large language models to help scientific writing remain to be determined.

scientific communication and education↗

bioRxiv2022

Self-supervised deep learning for pan-cancer mutation prediction from histopathology

The histopathological phenotype of tumors reflects the underlying genetic makeup. Deep learning can predict genetic alterations from tissue morphology, but it is unclear how well these predictions generalize to external datasets. Here, we present a deep learning pipeline based on self-supervised feature extraction which achieves a robust predictability of genetic alterations in two large multicentric datasets of seven tumor types.

cancer biology↗