UK Biotech Boost
The UK is launching a large-scale shift from animal testing to models created from real human cells, including mini-organs, organoids, and "organ-on-a-chip" systems. This is being done as a national program with significant funding and a clear goal: to make drug development faster, more accurate, and safer. Scientists will grow organoids from NHS patient cells - tiny structures smaller than a millimeter that replicate key functions of real organs.
The goal is to address the issue that more than 90% of drugs that pass animal tests fail in human clinical trials. Mice do not get sick like humans, and their reactions often do not match human responses. Organoids solve this problem by allowing drugs to be tested directly on human tissue, specifically on the tissue of a particular patient. The project is funded by the UK Medical Research Council with £20 million.
The project aims to create a library of standardized, validated organoids available to universities and pharmaceutical companies. This will enable ineffective or toxic drugs to be rejected earlier, reducing the risk to patients and decreasing the number of animals used in research. In 2025, 2.54 million animal procedures were conducted in the UK, and the government intends to accelerate the reduction of this number with new technologies. Meanwhile, Innovate UK is allocating an additional £2 million to projects that aim to reduce the use of dogs, monkeys, and other animals in safety tests, as reported in Nature Aging, July 2026.
🔗 Read original →
The UK is launching a large-scale shift from animal testing to models created from real human cells, including mini-organs, organoids, and "organ-on-a-chip" systems. This is being done as a national program with significant funding and a clear goal: to make drug development faster, more accurate, and safer. Scientists will grow organoids from NHS patient cells - tiny structures smaller than a millimeter that replicate key functions of real organs.
The goal is to address the issue that more than 90% of drugs that pass animal tests fail in human clinical trials. Mice do not get sick like humans, and their reactions often do not match human responses. Organoids solve this problem by allowing drugs to be tested directly on human tissue, specifically on the tissue of a particular patient. The project is funded by the UK Medical Research Council with £20 million.
The project aims to create a library of standardized, validated organoids available to universities and pharmaceutical companies. This will enable ineffective or toxic drugs to be rejected earlier, reducing the risk to patients and decreasing the number of animals used in research. In 2025, 2.54 million animal procedures were conducted in the UK, and the government intends to accelerate the reduction of this number with new technologies. Meanwhile, Innovate UK is allocating an additional £2 million to projects that aim to reduce the use of dogs, monkeys, and other animals in safety tests, as reported in Nature Aging, July 2026.
🔗 Read original →
GOV.UK
Patients to benefit from faster access to medicines as UK moves away from animal testing in science
New projects to move UK science away from animal testing
Language and Logic
Researchers compared the results of two individuals with severe aphasia to those of healthy participants in an fMRI study. The study, published in PNAS on July 6, found that language areas of the brain were not engaged when searching for rules and checking deductive conclusions. The team, led by neurobiologist Evelina Fedorenko from the Massachusetts Institute of Technology (MIT), investigated the long-standing debate about whether words serve as an internal means of formal reasoning.
The study involved 29 adults who underwent functional MRI scans, which indirectly track brain activity by monitoring changes in blood flow. The participants first identified their language network, which responded more strongly to sentences than to sets of nonsense words. Then, different subgroups of this sample searched for a rule that transformed one list of numbers into another, applied a given rule to new examples, checked syllogisms with "if-then" conditions, and chose a missing figure in geometric matrices.
The comparison between rule search and rule application separated the search process from instruction execution. In deductive tasks, the language network did not provide a significant response; when searching for a rule, its response was more than four times weaker than when reading sentences. The rule search involved a multiple-demand network, a distributed set of frontal and parietal areas active during complex goal-directed tasks. Complex deductive inferences involved a different set of frontal and parietal areas.
🔗 Read original →
Researchers compared the results of two individuals with severe aphasia to those of healthy participants in an fMRI study. The study, published in PNAS on July 6, found that language areas of the brain were not engaged when searching for rules and checking deductive conclusions. The team, led by neurobiologist Evelina Fedorenko from the Massachusetts Institute of Technology (MIT), investigated the long-standing debate about whether words serve as an internal means of formal reasoning.
The study involved 29 adults who underwent functional MRI scans, which indirectly track brain activity by monitoring changes in blood flow. The participants first identified their language network, which responded more strongly to sentences than to sets of nonsense words. Then, different subgroups of this sample searched for a rule that transformed one list of numbers into another, applied a given rule to new examples, checked syllogisms with "if-then" conditions, and chose a missing figure in geometric matrices.
The comparison between rule search and rule application separated the search process from instruction execution. In deductive tasks, the language network did not provide a significant response; when searching for a rule, its response was more than four times weaker than when reading sentences. The rule search involved a multiple-demand network, a distributed set of frontal and parietal areas active during complex goal-directed tasks. Complex deductive inferences involved a different set of frontal and parietal areas.
🔗 Read original →
bioRxiv
Evidence from Formal Logical Reasoning Reveals that the Language of Thought is not Natural Language
Humans are endowed with a powerful capacity for both inductive and deductive logical thought: we easily form generalizations based on a few examples and draw conclusions from known premises. Humans also arguably have the most sophisticated communication system…
Tissue Age Models
Researchers trained separate models to evaluate age based on 40 types of tissue samples from 983 deceased donors. The study, published in Nature Medicine on August 14, used 25,713 images of stained tissue sections to compare model predictions with pathology and telomere length, as well as gene expression data from blood samples.
The team trained a model for each tissue type to predict the donor's chronological age based on the tissue section's structure. The average absolute error was 4.88 years. The authors then calculated the tissue age gap, which is the difference between the model's prediction and the donor's chronological age. If a 55-year-old person's tissue section resembles those of older individuals, the age gap increases.
The study found that large age gaps in the GTEx project coincided with shorter telomeres and pathological signs on the same tissue sections. The authors also compared 19 histological clocks with 28 DNA methylation clocks, which are chemical marks on DNA that change with age. The age gaps showed a low correlation coefficient of 0.09 in the GTEx project, indicating that methylation and tissue structure reflect partially different properties of tissue state.
🔗 Read original →
Researchers trained separate models to evaluate age based on 40 types of tissue samples from 983 deceased donors. The study, published in Nature Medicine on August 14, used 25,713 images of stained tissue sections to compare model predictions with pathology and telomere length, as well as gene expression data from blood samples.
The team trained a model for each tissue type to predict the donor's chronological age based on the tissue section's structure. The average absolute error was 4.88 years. The authors then calculated the tissue age gap, which is the difference between the model's prediction and the donor's chronological age. If a 55-year-old person's tissue section resembles those of older individuals, the age gap increases.
The study found that large age gaps in the GTEx project coincided with shorter telomeres and pathological signs on the same tissue sections. The authors also compared 19 histological clocks with 28 DNA methylation clocks, which are chemical marks on DNA that change with age. The age gaps showed a low correlation coefficient of 0.09 in the GTEx project, indicating that methylation and tissue structure reflect partially different properties of tissue state.
🔗 Read original →
Nature
Histological aging signatures for monitoring tissue-specific aging and disease
Nature Medicine - Whole-slide histopathological images from 40 tissue types reveal morphological changes associated with aging, which, when paired with transcriptomic data from blood, are used to...
Primate mRNA Delivery
Researchers compared six variants of lipid nanoparticles for mRNA delivery in primates using 30 microliters of serum on August 11 in Nature Biotechnology, describing the snapCodes method. This method allows for the comparison of six lipid nanoparticle variants in a single non-human primate using a 30 microliter serum sample. The composition of the lipid nanoparticle determines the entire delivery chain, from the mRNA reaching the cell from the blood to its release and the cell's ability to assemble the protein.
The study builds upon previous research from the same team in 2025, where researchers introduced a mixture of 45 lipid nanoparticle variants to a primate. To read the barcodes, they had to dissect cells from the liver, spleen, bone marrow, and blood. The authors noted that comparing carriers in primates becomes a separate stage of development, with a high cost and potential harm to animals.
The snapCodes method transfers the reading of the result to the serum, allowing for a more efficient comparison of variants. Each of the six carrier variants contained mRNA with instructions for a protein with a SNAP tag and its own DNA code snapCode. The authors first tested this chain in control experiments and then intravenously administered the six variants to mice and non-human primates.
🔗 Read original →
Researchers compared six variants of lipid nanoparticles for mRNA delivery in primates using 30 microliters of serum on August 11 in Nature Biotechnology, describing the snapCodes method. This method allows for the comparison of six lipid nanoparticle variants in a single non-human primate using a 30 microliter serum sample. The composition of the lipid nanoparticle determines the entire delivery chain, from the mRNA reaching the cell from the blood to its release and the cell's ability to assemble the protein.
The study builds upon previous research from the same team in 2025, where researchers introduced a mixture of 45 lipid nanoparticle variants to a primate. To read the barcodes, they had to dissect cells from the liver, spleen, bone marrow, and blood. The authors noted that comparing carriers in primates becomes a separate stage of development, with a high cost and potential harm to animals.
The snapCodes method transfers the reading of the result to the serum, allowing for a more efficient comparison of variants. Each of the six carrier variants contained mRNA with instructions for a protein with a SNAP tag and its own DNA code snapCode. The authors first tested this chain in control experiments and then intravenously administered the six variants to mice and non-human primates.
🔗 Read original →
Nature
Non-invasive nanoparticle barcoding in nonhuman primates
Nature Biotechnology - Simultaneous delivery of several lipid nanoparticles is quantified using snapCodes in a single nonhuman primate.
Science Debate
The debate over the "decline of breakthrough science" has narrowed down to how bibliographic databases store references. On August 12, Nature published a commentary on a 2023 study about the decline of "breakthrough" science and the authors' response. At the center of the debate is the CD-index, a metric that attempts to determine if a work has changed the direction of research based on future citations.
The 2023 study analyzed almost 45 million articles and 3.9 million patents. A high CD-index is given to a work that is cited by later publications without simultaneously referencing its predecessors. The authors interpreted this pattern as a subsequent work deviating from the previous scientific path. The problem is visible in the five-year version of the metric, CD5. If a record in the catalog lacks a list of sources, a later article cannot cite both the record and its predecessors.
Vincent Holst and co-authors manually checked random articles and patents from this category and found that 93% of articles and 98% of patents had references in the original documents. In the open catalog SciSciNet, where publications and connections between their references are collected, they obtained a change in the average CD5 from −0.31 to −0.01 from 1944 to 2011 after excluding records with CD5 = 1. According to their version, the early years appear more "breakthrough" because the catalog did not preserve the entire bibliography.
In response, in Nature, the authors of the 2023 study, Michael Park, Erin Leichty, and Russell Funk, associate zero records in Holst's set with editorial materials, books, conference publications, and reviews, which often lack a list of sources. They also report that alternative metrics that immediately exclude documents without references preserve the trend of decline.
🔗 Read original →
The debate over the "decline of breakthrough science" has narrowed down to how bibliographic databases store references. On August 12, Nature published a commentary on a 2023 study about the decline of "breakthrough" science and the authors' response. At the center of the debate is the CD-index, a metric that attempts to determine if a work has changed the direction of research based on future citations.
The 2023 study analyzed almost 45 million articles and 3.9 million patents. A high CD-index is given to a work that is cited by later publications without simultaneously referencing its predecessors. The authors interpreted this pattern as a subsequent work deviating from the previous scientific path. The problem is visible in the five-year version of the metric, CD5. If a record in the catalog lacks a list of sources, a later article cannot cite both the record and its predecessors.
Vincent Holst and co-authors manually checked random articles and patents from this category and found that 93% of articles and 98% of patents had references in the original documents. In the open catalog SciSciNet, where publications and connections between their references are collected, they obtained a change in the average CD5 from −0.31 to −0.01 from 1944 to 2011 after excluding records with CD5 = 1. According to their version, the early years appear more "breakthrough" because the catalog did not preserve the entire bibliography.
In response, in Nature, the authors of the 2023 study, Michael Park, Erin Leichty, and Russell Funk, associate zero records in Holst's set with editorial materials, books, conference publications, and reviews, which often lack a list of sources. They also report that alternative metrics that immediately exclude documents without references preserve the trend of decline.
🔗 Read original →
Nature
Dataset artefacts can partially drive the measured decline in disruption
Nature - Dataset artefacts can partially drive the measured decline in disruption
VOICE Model
The VOICE model evaluates gene activity in individual cells based on tissue images and a library of already measured cells. On August 8, a team from the University of Michigan released the VOICE model preprint. The authors trained the model on 23.2 million cells from 75 human tissue sections across 15 tissues, where each cell has a standard H&E-stained tissue image and gene activity measurements using Xenium technology.
The model takes cell shape from the image and, for genes that cannot be recognized by shape, refers to a library of similar cells where gene activity has already been measured. H&E staining shows cell shape and tissue structure, while Xenium technology measures the activity of pre-selected genes in individual cells of the same section and retains their coordinates. The model was trained on sections where these two types of data were already matched.
The model has two paths to a single estimate: a direct path using H&E cell features and its surroundings to estimate gene activity, and a second path that searches the library for similar cells of the same tissue and averages their measured Xenium values. For each gene, VOICE adjusts the weight between these paths on the training sections. The authors describe this division as: "some genes are closely related to cell morphology and are predicted directly from the image; others are better evaluated by cells with a similar transcriptomic profile". The transcriptomic profile is a set of gene activity measurements in a single cell, as described in Nature Aging, July 2026.
🔗 Read original →
The VOICE model evaluates gene activity in individual cells based on tissue images and a library of already measured cells. On August 8, a team from the University of Michigan released the VOICE model preprint. The authors trained the model on 23.2 million cells from 75 human tissue sections across 15 tissues, where each cell has a standard H&E-stained tissue image and gene activity measurements using Xenium technology.
The model takes cell shape from the image and, for genes that cannot be recognized by shape, refers to a library of similar cells where gene activity has already been measured. H&E staining shows cell shape and tissue structure, while Xenium technology measures the activity of pre-selected genes in individual cells of the same section and retains their coordinates. The model was trained on sections where these two types of data were already matched.
The model has two paths to a single estimate: a direct path using H&E cell features and its surroundings to estimate gene activity, and a second path that searches the library for similar cells of the same tissue and averages their measured Xenium values. For each gene, VOICE adjusts the weight between these paths on the training sections. The authors describe this division as: "some genes are closely related to cell morphology and are predicted directly from the image; others are better evaluated by cells with a similar transcriptomic profile". The transcriptomic profile is a set of gene activity measurements in a single cell, as described in Nature Aging, July 2026.
🔗 Read original →
PubMed Central (PMC)
A visual–omics foundation model to bridge histopathology with spatial transcriptomics
Artificial intelligence has revolutionized computational biology. Recent developments in omics technologies, including single-cell RNA sequencing and spatial transcriptomics, provide detailed genomic data alongside tissue histology. However, current ...
Matrix Proof Found
The GPT-5.6 model helped find a path to proving the 22-year-old Kroizë hypothesis on matrices. On July 27, Shangmu Jin posted a manuscript with an exact boundary for matrix functions. On August 14, mathematicians Alex Townsend and Anne Greenbaum reported that, along with the hypothesis author Michel Kroizë, they had thoroughly checked the proof.
The Kroizë hypothesis states that if a polynomial p, a formula of additions and multiplications, nowhere exceeds a magnitude M on the numerical range of a matrix A, then the matrix p(A) has a norm no higher than 2M. In 2004, Michel Kroizë formulated a general variant of the question, and his 2007 estimate gave a multiplier of 11.08; in 2017, Kroizë and Palenzia reduced it to 1 + √2, approximately 2.414.
To get exactly two, a different approach was needed in the proof. In the previous approach, the target value was estimated along with an additional expression, which kept the universal bound above two. The current manuscript by Jin cites a specific contribution from ChatGPT: the model suggested where to take the values of the auxiliary function and add a point at zero. This choice removes the additional expression, and comparing two weighted sums of squares gives a multiplier of two, as reported in SIAM News.
🔗 Read original →
The GPT-5.6 model helped find a path to proving the 22-year-old Kroizë hypothesis on matrices. On July 27, Shangmu Jin posted a manuscript with an exact boundary for matrix functions. On August 14, mathematicians Alex Townsend and Anne Greenbaum reported that, along with the hypothesis author Michel Kroizë, they had thoroughly checked the proof.
The Kroizë hypothesis states that if a polynomial p, a formula of additions and multiplications, nowhere exceeds a magnitude M on the numerical range of a matrix A, then the matrix p(A) has a norm no higher than 2M. In 2004, Michel Kroizë formulated a general variant of the question, and his 2007 estimate gave a multiplier of 11.08; in 2017, Kroizë and Palenzia reduced it to 1 + √2, approximately 2.414.
To get exactly two, a different approach was needed in the proof. In the previous approach, the target value was estimated along with an additional expression, which kept the universal bound above two. The current manuscript by Jin cites a specific contribution from ChatGPT: the model suggested where to take the values of the auxiliary function and add a point at zero. This choice removes the additional expression, and comparing two weighted sums of squares gives a multiplier of two, as reported in SIAM News.
🔗 Read original →
Discovered Materials Test
Discovered Materials launched the Material Discovery Bench, a test for AI models searching for materials for multilayered microchips, on August 12. Seven models proposed 526 candidates, but only one recipe received a "worth trying" evaluation; the team is attempting to synthesize this material.
In such microchips, memory is placed on top of logical circuits to reduce signal travel distance. A dielectric is needed between layers to isolate electrical circuits and dissipate heat. Otherwise, accumulated heat within the assembly limits the number of layers.
The AI model searches for articles, works with materials science programs, and proposes a crystalline structure - a way to arrange atoms in a substance. Computer models select variants that, according to calculations, effectively dissipate heat, isolate electrical circuits, withstand load, and maintain structure, resulting in a calculated candidate for an interlayer material.
As reported in Nature Aging, July 2026, the model must then write a recipe for a thin film - a layer of substance deposited on a plate. The recipe must specify the starting substances, their application method, and the conditions under which the desired atomic arrangement is obtained. Specialists in thin film deposition evaluated recipe examples and developed rules; during the test, a separate language model applies these rules.
The team is now attempting to synthesize the single material that received a "worth trying" evaluation, following the Material Discovery Bench protocol.
🔗 Read original →
Discovered Materials launched the Material Discovery Bench, a test for AI models searching for materials for multilayered microchips, on August 12. Seven models proposed 526 candidates, but only one recipe received a "worth trying" evaluation; the team is attempting to synthesize this material.
In such microchips, memory is placed on top of logical circuits to reduce signal travel distance. A dielectric is needed between layers to isolate electrical circuits and dissipate heat. Otherwise, accumulated heat within the assembly limits the number of layers.
The AI model searches for articles, works with materials science programs, and proposes a crystalline structure - a way to arrange atoms in a substance. Computer models select variants that, according to calculations, effectively dissipate heat, isolate electrical circuits, withstand load, and maintain structure, resulting in a calculated candidate for an interlayer material.
As reported in Nature Aging, July 2026, the model must then write a recipe for a thin film - a layer of substance deposited on a plate. The recipe must specify the starting substances, their application method, and the conditions under which the desired atomic arrangement is obtained. Specialists in thin film deposition evaluated recipe examples and developed rules; during the test, a separate language model applies these rules.
The team is now attempting to synthesize the single material that received a "worth trying" evaluation, following the Material Discovery Bench protocol.
🔗 Read original →
TechCrunch
Discovered Materials is playing AI whack-a-mole to hunt cooler chips | TechCrunch
Discovered Materials raised $9 million to fund the hunt for more novel materials to build more efficient chips.
Biostasis Quality Criteria
The Biostasis Roadmap co-author, Emile Kendziorra, stated that two years after its publication, most tasks remain open. One of the questions is what degree of brain preservation is needed for future methods to restore a person. Biostasis is an attempt to preserve a person in anticipation of future medical advancements that could restore them. The focus is on brain structures related to memory, personality, and identity.
Developers need a criterion that shows which damages are still compatible with preserving this information. The Biostasis Roadmap 2024 authors described several levels of testing, including light microscopy, electron microscopy, and the arrangement of proteins, lipids, and nucleic acids, including DNA and RNA. After thawing, researchers can check cell viability and electrical activity of nerve tissue.
Computer tomography (CT) and measurements during the procedure provide indirect data, showing how the cryoprotectant - a solution that reduces ice formation - is distributed in tissues. These indicators are useful when their results are compared with the state of cells and tissue structure. In a new post, Kendziorra formulated the purpose of the roadmap as "a to-do list," and a general criterion would allow comparing different protocol improvements.
🔗 Read original →
The Biostasis Roadmap co-author, Emile Kendziorra, stated that two years after its publication, most tasks remain open. One of the questions is what degree of brain preservation is needed for future methods to restore a person. Biostasis is an attempt to preserve a person in anticipation of future medical advancements that could restore them. The focus is on brain structures related to memory, personality, and identity.
Developers need a criterion that shows which damages are still compatible with preserving this information. The Biostasis Roadmap 2024 authors described several levels of testing, including light microscopy, electron microscopy, and the arrangement of proteins, lipids, and nucleic acids, including DNA and RNA. After thawing, researchers can check cell viability and electrical activity of nerve tissue.
Computer tomography (CT) and measurements during the procedure provide indirect data, showing how the cryoprotectant - a solution that reduces ice formation - is distributed in tissues. These indicators are useful when their results are compared with the state of cells and tissue structure. In a new post, Kendziorra formulated the purpose of the roadmap as "a to-do list," and a general criterion would allow comparing different protocol improvements.
🔗 Read original →
PubMed Central (PMC)
Structural brain preservation: a potential bridge to future medical technologies
When faced with the prospect of death, some people would prefer a form of long-term preservation that may allow them to be restored to healthy life in the future, if technology ever develops to the point that this is feasible and humane. Some ...
Mice Study Reveals DLK1 Source
Researchers found that in mice with shortened telomeres, microglia is a source of DLK1, a protein that inhibits the maturation of myelin-producing cells. In an article published in Neuron on August 11, the authors separately investigated the source and action of the soluble form of DLK1, sDLK1.
In a telomere model, excess sDLK1 in the spinal fluid primarily came from microglia, and in other experiments, the protein itself delayed the maturation of myelin-producing cells. The authors bred mice without the Terc gene, necessary for telomerase enzyme function, for three generations.
In the third-generation mice, the telomere signal in microglia was lower, and microglia acquired signs of cellular aging. At the same time, there were changes in the gene function of oligodendrocytes, the cells that produce the myelin sheath around nerve fibers. A 2023 study had already shown that DLK1 of neuronal origin can delay the maturation of oligodendrocyte precursor cells.
The current authors investigated whether microglia with signs of cellular aging could be a source of the same protein. A July study had found inflamed microglia with signs of cellular aging near myelinated fibers in old mice. The current experiments tested one of the signals through which such microglia may affect myelin cells.
🔗 Read original →
Researchers found that in mice with shortened telomeres, microglia is a source of DLK1, a protein that inhibits the maturation of myelin-producing cells. In an article published in Neuron on August 11, the authors separately investigated the source and action of the soluble form of DLK1, sDLK1.
In a telomere model, excess sDLK1 in the spinal fluid primarily came from microglia, and in other experiments, the protein itself delayed the maturation of myelin-producing cells. The authors bred mice without the Terc gene, necessary for telomerase enzyme function, for three generations.
In the third-generation mice, the telomere signal in microglia was lower, and microglia acquired signs of cellular aging. At the same time, there were changes in the gene function of oligodendrocytes, the cells that produce the myelin sheath around nerve fibers. A 2023 study had already shown that DLK1 of neuronal origin can delay the maturation of oligodendrocyte precursor cells.
The current authors investigated whether microglia with signs of cellular aging could be a source of the same protein. A July study had found inflamed microglia with signs of cellular aging near myelinated fibers in old mice. The current experiments tested one of the signals through which such microglia may affect myelin cells.
🔗 Read original →
Neuron
Senescent microglia with shortened telomeres secrete soluble DLK1 to induce aging-associated hypomyelination and neuronal dysfunction
Liu et al. investigate mechanisms of brain aging using a telomere-shortening model.
They show that telomere shortening induces microglial senescence and identify soluble
DLK1 as a previously unrecognized microglial SASP factor that has detrimental effects…
They show that telomere shortening induces microglial senescence and identify soluble
DLK1 as a previously unrecognized microglial SASP factor that has detrimental effects…
Biological Age Linked to Gene Activity
A study published on July 20 in npj Aging analyzed 3,227 participants from the Health and Retirement Study, a long-term American study of people over 50 years old. The researchers measured both methylation, or chemical marks on DNA, and RNA levels, which indicate active genes in blood cells. Epigenetic clocks are formulas that convert DNA marks into an estimate of biological age, and they were trained to solve different tasks: Horvath and Hannum predict calendar age, PhenoAge and GrimAge predict health indicators and mortality risk, and DunedinPACE measures the pace of age-related changes in multiple systems.
The authors found that 49 genes were statistically associated with high Horvath values, while 3,204 genes were associated with DunedinPACE. No single gene was common to all five clocks, and the same result was obtained when comparing pathways in Reactome, a catalog of biological processes. On a broader level, the estimates converge, all affecting metabolism, development, immune function, and cellular regulation.
The researchers created transcriptomic aging signatures, TAGS, based on the profiles of genes associated with each clock. They were calculated based on RNA levels of genes associated with the corresponding clocks and were validated on 2,584 participants. The connection between TAGS and the corresponding clocks was moderate, reflecting current gene activity in the blood, while the original clocks reflected chemical marks on DNA. In a group of 645 people, TAGS were more strongly associated with several outcomes than the corresponding DNA methylation clocks.
🔗 Read original →
A study published on July 20 in npj Aging analyzed 3,227 participants from the Health and Retirement Study, a long-term American study of people over 50 years old. The researchers measured both methylation, or chemical marks on DNA, and RNA levels, which indicate active genes in blood cells. Epigenetic clocks are formulas that convert DNA marks into an estimate of biological age, and they were trained to solve different tasks: Horvath and Hannum predict calendar age, PhenoAge and GrimAge predict health indicators and mortality risk, and DunedinPACE measures the pace of age-related changes in multiple systems.
The authors found that 49 genes were statistically associated with high Horvath values, while 3,204 genes were associated with DunedinPACE. No single gene was common to all five clocks, and the same result was obtained when comparing pathways in Reactome, a catalog of biological processes. On a broader level, the estimates converge, all affecting metabolism, development, immune function, and cellular regulation.
The researchers created transcriptomic aging signatures, TAGS, based on the profiles of genes associated with each clock. They were calculated based on RNA levels of genes associated with the corresponding clocks and were validated on 2,584 participants. The connection between TAGS and the corresponding clocks was moderate, reflecting current gene activity in the blood, while the original clocks reflected chemical marks on DNA. In a group of 645 people, TAGS were more strongly associated with several outcomes than the corresponding DNA methylation clocks.
🔗 Read original →
Nature
How epigenetic clocks tick: unpacking the black box by deciphering biological pathways and transcriptomic signatures of accelerated…
npj Aging - How epigenetic clocks tick: unpacking the black box by deciphering biological pathways and transcriptomic signatures of accelerated aging
Mouse Aging Atlas
The mouse aging atlases show how tissues change the composition of cellular populations at different age windows. On August 14, Quanta Magazine published an interview with cellular biologist Junyue Cao about two mouse aging atlases. They allow tracking of how the proportion of different cellular populations in tissues changes with age and what signals accompany these transitions.
Tissue consists of cells with different tasks: some support vessels and tendons, others participate in the immune response, and others restore damage. With age, their ratio changes. Cao studies which populations shrink or expand in each age window and what signals are associated with these transitions.
He came to this task when searching for molecular targets to slow down aging: many molecular pathways are associated with age, and their action depends on the cell type. His group developed EasySci, a method that reads active genes in the nuclei of individual cells. By this set of genes, one can recognize the cell type and count how its proportion changes in the tissue. The PanSci atlas includes 21,786,931 nuclear profiles from over 600 samples of 14 tissues taken from mice of both sexes aged from three to 23 months.
All samples were processed according to the same scheme so as not to take differences between data batches for age-related changes. The authors found more than 200 cellular populations whose proportion in tissues changed significantly with age. Some groups shrank, others expanded; shifts occurred at different age windows. Nature Aging, July 2026, also explored this topic.
🔗 Read original →
The mouse aging atlases show how tissues change the composition of cellular populations at different age windows. On August 14, Quanta Magazine published an interview with cellular biologist Junyue Cao about two mouse aging atlases. They allow tracking of how the proportion of different cellular populations in tissues changes with age and what signals accompany these transitions.
Tissue consists of cells with different tasks: some support vessels and tendons, others participate in the immune response, and others restore damage. With age, their ratio changes. Cao studies which populations shrink or expand in each age window and what signals are associated with these transitions.
He came to this task when searching for molecular targets to slow down aging: many molecular pathways are associated with age, and their action depends on the cell type. His group developed EasySci, a method that reads active genes in the nuclei of individual cells. By this set of genes, one can recognize the cell type and count how its proportion changes in the tissue. The PanSci atlas includes 21,786,931 nuclear profiles from over 600 samples of 14 tissues taken from mice of both sexes aged from three to 23 months.
All samples were processed according to the same scheme so as not to take differences between data batches for age-related changes. The authors found more than 200 cellular populations whose proportion in tissues changed significantly with age. Some groups shrank, others expanded; shifts occurred at different age windows. Nature Aging, July 2026, also explored this topic.
🔗 Read original →
PubMed Central (PMC)
A panoramic view of cell population dynamics in mammalian aging
To elucidate aging-associated cellular population dynamics, we present PanSci, a single-cell transcriptome atlas profiling >20 million cells from 623 mouse tissues across different life stages, sexes, and genotypes. This comprehensive dataset ...
New Sight Restoration
A new approach to "turning on" a blind retina without genetic operations or implants uses special light-sensitive molecules that work as a molecular prosthesis for dead photoreceptors. The most impressive aspect is that in animal experiments, it worked not only after injections but also in the form of ordinary eye drops, under room-level lighting. This technology targets diseases that destroy photoreceptors, such as age-related macular degeneration and retinitis pigmentosa. The new technology utilizes photopharmacology - prosthe6 molecules that change shape under visible light.
They attach to the mGlu6 protein on ON-bipolar cells, which normally receive signals from photoreceptors. When light enters the eye, prosthe6 is activated and triggers a signal chain, as if the photoreceptor were still working. This is maximally close to natural physiology. In animal experiments, the results were noticeable: blind fish regained rapid eye movements, which require complex image processing. Mice with models of macular degeneration and retinitis pigmentosa regained light-perception behavior, avoiding bright areas again.
The prosthe6-12 and prosthe6-15 versions worked especially well, with the effect appearing after injection and after drops, under ordinary white light, without special lamps. The molecules act at an early stage of the visual chain - on bipolar cells, not on deeper structures like electronic implants. This gives a chance to preserve natural image processing within the retina. However, it is currently impossible to say how detailed human vision will be: animal behavior shows the presence of light sensitivity but not the restoration of sharpness.
There are also first steps towards the clinic: in Australia, a small test of another photo-switchable drug, KIO-301, has already been conducted on 6 patients with severe retinitis pigmentosa, as reported in Nature Aging, July 2026. It was safe but not intended to test efficacy. This shows that the technology is already moving out of laboratories. The most promising aspect is the drop form: if it can be proven that the drug passes through the human cornea in the necessary concentration, treatment may become simple, without requiring surgery, implants, or special glasses.
🔗 Read original →
A new approach to "turning on" a blind retina without genetic operations or implants uses special light-sensitive molecules that work as a molecular prosthesis for dead photoreceptors. The most impressive aspect is that in animal experiments, it worked not only after injections but also in the form of ordinary eye drops, under room-level lighting. This technology targets diseases that destroy photoreceptors, such as age-related macular degeneration and retinitis pigmentosa. The new technology utilizes photopharmacology - prosthe6 molecules that change shape under visible light.
They attach to the mGlu6 protein on ON-bipolar cells, which normally receive signals from photoreceptors. When light enters the eye, prosthe6 is activated and triggers a signal chain, as if the photoreceptor were still working. This is maximally close to natural physiology. In animal experiments, the results were noticeable: blind fish regained rapid eye movements, which require complex image processing. Mice with models of macular degeneration and retinitis pigmentosa regained light-perception behavior, avoiding bright areas again.
The prosthe6-12 and prosthe6-15 versions worked especially well, with the effect appearing after injection and after drops, under ordinary white light, without special lamps. The molecules act at an early stage of the visual chain - on bipolar cells, not on deeper structures like electronic implants. This gives a chance to preserve natural image processing within the retina. However, it is currently impossible to say how detailed human vision will be: animal behavior shows the presence of light sensitivity but not the restoration of sharpness.
There are also first steps towards the clinic: in Australia, a small test of another photo-switchable drug, KIO-301, has already been conducted on 6 patients with severe retinitis pigmentosa, as reported in Nature Aging, July 2026. It was safe but not intended to test efficacy. This shows that the technology is already moving out of laboratories. The most promising aspect is the drop form: if it can be proven that the drug passes through the human cornea in the necessary concentration, treatment may become simple, without requiring surgery, implants, or special glasses.
🔗 Read original →
ACS Publications
Restoration
of Saccadic Eye Movements and Visually
Guided Behavior in Ambient White Light with Photoswitchable Small
Molecules
of Saccadic Eye Movements and Visually
Guided Behavior in Ambient White Light with Photoswitchable Small
Molecules
Abstract. Blinding diseases
due to the degeneration of photoreceptors (PhRs),
such as geographic atrophy (GA) secondary to dry age-related macular
degenera
due to the degeneration of photoreceptors (PhRs),
such as geographic atrophy (GA) secondary to dry age-related macular
degenera
Neurons Revitalized
Researchers found that enhancing cellular cleanup in neurons derived from elderly donors' skin cells improved some mitochondrial metrics. On August 16, 2026, Eva Kliman and colleagues published a study on cortical neurons directly converted from skin cells of people of different ages. The neurons from elderly donors had weaker acidic stages of autophagy, a cellular cleanup process.
The authors increased the formation of autophagosomes, vesicles for this cleanup, and some mitochondrial metrics in the neurons from elderly donors approached those of young cells. The study was published in Nature Aging, July 2026. The researchers used G2-115 to increase autophagosome formation, resulting in a 60% increase in their number in all age groups.
The authors compared young and old cell lines, as well as six pairs of samples taken from the same men approximately 15 years apart. In the old cells, 16.64% of autophagosomes were acidic, compared to 47.06% in young cells. The results suggest that the lack of lysosomes in the neurites hinders autophagosome acidification, leading to more frequent mitochondrial division and fusion.
🔗 Read original →
Researchers found that enhancing cellular cleanup in neurons derived from elderly donors' skin cells improved some mitochondrial metrics. On August 16, 2026, Eva Kliman and colleagues published a study on cortical neurons directly converted from skin cells of people of different ages. The neurons from elderly donors had weaker acidic stages of autophagy, a cellular cleanup process.
The authors increased the formation of autophagosomes, vesicles for this cleanup, and some mitochondrial metrics in the neurons from elderly donors approached those of young cells. The study was published in Nature Aging, July 2026. The researchers used G2-115 to increase autophagosome formation, resulting in a 60% increase in their number in all age groups.
The authors compared young and old cell lines, as well as six pairs of samples taken from the same men approximately 15 years apart. In the old cells, 16.64% of autophagosomes were acidic, compared to 47.06% in young cells. The results suggest that the lack of lysosomes in the neurites hinders autophagosome acidification, leading to more frequent mitochondrial division and fusion.
🔗 Read original →
PubMed Central (PMC)
Sequelae and reversal of age-dependent alterations in mitochondrial dynamics via autophagy enhancement in reprogrammed human neurons
How aging of human neurons affects dynamics of essential organelle such as mitochondria and autophagosomes remains largely unknown. MicroRNA-induced directly reprogrammed neurons (miNs) derived from adult fibroblasts retain age-associated signatures ...
Implant Generates Electricity
The implant is a thin, biodegradable film that dissolves in the body over time, but while it works, it can be powered by external ultrasound and glow under ultraviolet light, showing its status. The technology is based on the triboelectric effect, where two different materials come into contact and rub against each other, generating a static charge. Researchers embedded a molecule called Ir1b in a biopolymer, a luminescent material based on iridium, which they found and refined using machine learning.
The Ir1b molecule was found to be approximately 1.8 times more efficient than regular luminescent materials in generating a triboelectric charge. The size of one molecule is just 1.3 nanometers. When an ultrasound of 20 kHz frequency is directed at the implant, it starts generating electricity: up to 3.6 V voltage and around 16 μA current. This is enough to power simple biomedical functions.
The implant is almost invisible under the skin, but when exposed to a UV lamp, it starts glowing. In experiments on mice, the device's contour was clearly visible, and if the film was damaged, the glow disappeared, indicating that the implant was broken or displaced. This technology has potential, especially for devices like biosensors or drug delivery systems, where a battery can increase size, complicate design, and pose risks, as reported in Nature Aging, July 2026.
🔗 Read original →
The implant is a thin, biodegradable film that dissolves in the body over time, but while it works, it can be powered by external ultrasound and glow under ultraviolet light, showing its status. The technology is based on the triboelectric effect, where two different materials come into contact and rub against each other, generating a static charge. Researchers embedded a molecule called Ir1b in a biopolymer, a luminescent material based on iridium, which they found and refined using machine learning.
The Ir1b molecule was found to be approximately 1.8 times more efficient than regular luminescent materials in generating a triboelectric charge. The size of one molecule is just 1.3 nanometers. When an ultrasound of 20 kHz frequency is directed at the implant, it starts generating electricity: up to 3.6 V voltage and around 16 μA current. This is enough to power simple biomedical functions.
The implant is almost invisible under the skin, but when exposed to a UV lamp, it starts glowing. In experiments on mice, the device's contour was clearly visible, and if the film was damaged, the glow disappeared, indicating that the implant was broken or displaced. This technology has potential, especially for devices like biosensors or drug delivery systems, where a battery can increase size, complicate design, and pose risks, as reported in Nature Aging, July 2026.
🔗 Read original →
Nature
Bioresorbable and optically readable triboelectric implants enabled by nanoscale iridophosphors
Nature Nanotechnology - A battery-free, bioresorbable triboelectric implant combines ultrasound energy harvesting with visible phosphorescent self-reporting, enabling non-invasive, long-term in...
DARPA Firefox Program
The DARPA agency has a program called Firefox, which is not related to the famous internet browser, but rather named after a cult film about a mentally controlled fighter jet. The program aims to create a non-invasive brain interface that can match the accuracy of implants without surgery or electrodes, using only optical methods. DARPA has set specific targets: 100 channels, spatial resolution of 50 microns, and temporal resolution of 100 microseconds.
This level of precision can read the activity of neural ensembles, not just general fluctuations. Currently, the agency has an optical system that can register signals at the depth of the human cortex, but it works on a limited scale. The Firefox program needs to solve two problems: quickly "scanning" multiple areas of the brain simultaneously and processing the massive data streams that will exceed the capabilities of classical digital electronics at such frequency and resolution.
DARPA explicitly states that it will be necessary to "exit the digital domain", which means searching for computational architectures that can work with neural streams in real-time, such as quantum circuits, neuromorphic processors, or even biocomputing. The idea is simple: if the interface must work at brain speed, then the calculations must be closer to its principles. Registration for the program is open until August 21, 2026, and the first day for developers is September 1.
🔗 Read original →
The DARPA agency has a program called Firefox, which is not related to the famous internet browser, but rather named after a cult film about a mentally controlled fighter jet. The program aims to create a non-invasive brain interface that can match the accuracy of implants without surgery or electrodes, using only optical methods. DARPA has set specific targets: 100 channels, spatial resolution of 50 microns, and temporal resolution of 100 microseconds.
This level of precision can read the activity of neural ensembles, not just general fluctuations. Currently, the agency has an optical system that can register signals at the depth of the human cortex, but it works on a limited scale. The Firefox program needs to solve two problems: quickly "scanning" multiple areas of the brain simultaneously and processing the massive data streams that will exceed the capabilities of classical digital electronics at such frequency and resolution.
DARPA explicitly states that it will be necessary to "exit the digital domain", which means searching for computational architectures that can work with neural streams in real-time, such as quantum circuits, neuromorphic processors, or even biocomputing. The idea is simple: if the interface must work at brain speed, then the calculations must be closer to its principles. Registration for the program is open until August 21, 2026, and the first day for developers is September 1.
🔗 Read original →
Cortisol Shifts Cell Clocks
The hormone cortisol sends a signal to cells to adjust their internal clocks, leading to a rapid increase in the production of the Per1 gene and PER1 protein. This process is crucial for synchronizing the cell's internal clock with the body's natural rhythm. Researchers found that cortisol, a glucocorticoid hormone, plays a key role in this process, and their findings were published in a preprint on August 13.
The study revealed that the glucocorticoid receptor, a protein that binds to cortisol and transmits its signal to the cell, is essential for this process. The researchers used a selective activator of the receptor to shift the cell's clock, and two receptor-blocking compounds to cancel this effect. They also found that disabling the receptor gene prevented the cell from responding to cortisol, but introducing an artificial variant of the receptor that is constantly present in the nucleus restored the response.
The researchers then focused on the Per1 gene, a key component of the molecular clock. They found that after a hormonal pulse, the level of mature Per1 RNA increased rapidly, and the level of PER1 protein increased within 12 minutes in human cells. The study also showed that this rapid response is due to enhanced RNA processing, rather than new RNA synthesis, and that it is mediated by the glucocorticoid receptor. The findings were published in preprint, August 2026.
🔗 Read original →
The hormone cortisol sends a signal to cells to adjust their internal clocks, leading to a rapid increase in the production of the Per1 gene and PER1 protein. This process is crucial for synchronizing the cell's internal clock with the body's natural rhythm. Researchers found that cortisol, a glucocorticoid hormone, plays a key role in this process, and their findings were published in a preprint on August 13.
The study revealed that the glucocorticoid receptor, a protein that binds to cortisol and transmits its signal to the cell, is essential for this process. The researchers used a selective activator of the receptor to shift the cell's clock, and two receptor-blocking compounds to cancel this effect. They also found that disabling the receptor gene prevented the cell from responding to cortisol, but introducing an artificial variant of the receptor that is constantly present in the nucleus restored the response.
The researchers then focused on the Per1 gene, a key component of the molecular clock. They found that after a hormonal pulse, the level of mature Per1 RNA increased rapidly, and the level of PER1 protein increased within 12 minutes in human cells. The study also showed that this rapid response is due to enhanced RNA processing, rather than new RNA synthesis, and that it is mediated by the glucocorticoid receptor. The findings were published in preprint, August 2026.
🔗 Read original →
bioRxiv
Post-transcriptional glucocorticoid receptor signalling synchronises circadian rhythms
Circadian rhythms are driven by cellular clocks throughout the body that maintain optimal health. Glucocorticoid (GC) hormones are vital for synchronising cellular clocks, aligning them with external day/night cycles, but the mechanism was unresolved. We…
Insilico Launches Virtual Aging Cell
Insilico Medicine launched the Virtual Aging Cell page on August 14, providing a preview of the project the company is developing. The page features a demo screen with artificially generated data, where users can input the organism type, tissue, biological age, and substance to simulate a molecular response.
The Virtual Aging Cell is a computer model that builds a predicted molecular response based on the given conditions. The biological age is estimated from the cell's state data and may differ from the organism's calendar age. Insilico suggests using the Virtual Aging Cell to compare the predicted cellular response to a single intervention at different age states.
The company claims that the future model will generate three types of profiles: gene activity, protein composition, and DNA methylation. These calculations will help compare the predicted effect of a substance on tissue at different age states and select a hypothesis for the next laboratory experiment, as described in Nature Aging, July 2026.
🔗 Read original →
Insilico Medicine launched the Virtual Aging Cell page on August 14, providing a preview of the project the company is developing. The page features a demo screen with artificially generated data, where users can input the organism type, tissue, biological age, and substance to simulate a molecular response.
The Virtual Aging Cell is a computer model that builds a predicted molecular response based on the given conditions. The biological age is estimated from the cell's state data and may differ from the organism's calendar age. Insilico suggests using the Virtual Aging Cell to compare the predicted cellular response to a single intervention at different age states.
The company claims that the future model will generate three types of profiles: gene activity, protein composition, and DNA methylation. These calculations will help compare the predicted effect of a substance on tissue at different age states and select a hypothesis for the next laboratory experiment, as described in Nature Aging, July 2026.
🔗 Read original →
PubMed Central (PMC)
Precious1GPT: multimodal transformer-based transfer learning for aging clock development and feature importance analysis for aging…
Aging is a complex and multifactorial process that increases the risk of various age-related diseases and there are many aging clocks that can accurately predict chronological age, mortality, and health status. These clocks are disconnected and are ...
Alzheimer's Protein Tau
Researchers applied the STARFISH method to primary mouse neurons, detecting tau protein synthesis in dendrites, the branches that receive signals from other cells. A membrane-bound complex called the proteasome quickly degraded about a third of the new tau protein. When this complex was blocked in neurons and a mouse model with human tau, aggregates accumulated, depending on new protein synthesis.
In a healthy brain, tau is mainly concentrated in axons, the long extensions through which a neuron transmits signals. In Alzheimer's disease, some tau appears in the cell body and dendrites, where protein clusters form. The authors asked if some of this protein is born directly in the dendrites. A map of mRNA shows where the cellular instructions for assembling a protein are located. STARFISH shows the moment when a ribosome reads the instructions and assembles the protein.
In mouse neurons, the mRNA map of tau covered the cell body and extensions, while STARFISH detected active tau synthesis only in dendrites. A three-dimensional reconstruction separated the dendrites from intersecting axons. The authors labeled newly assembled tau for 30 seconds and tracked its fate, finding that almost half of the label disappeared within 30-60 seconds. The substance iBEp, which blocks the proteasome, stopped this rapid degradation, with the authors estimating that the proteasome degrades about a third of new tau.
🔗 Read original →
Researchers applied the STARFISH method to primary mouse neurons, detecting tau protein synthesis in dendrites, the branches that receive signals from other cells. A membrane-bound complex called the proteasome quickly degraded about a third of the new tau protein. When this complex was blocked in neurons and a mouse model with human tau, aggregates accumulated, depending on new protein synthesis.
In a healthy brain, tau is mainly concentrated in axons, the long extensions through which a neuron transmits signals. In Alzheimer's disease, some tau appears in the cell body and dendrites, where protein clusters form. The authors asked if some of this protein is born directly in the dendrites. A map of mRNA shows where the cellular instructions for assembling a protein are located. STARFISH shows the moment when a ribosome reads the instructions and assembles the protein.
In mouse neurons, the mRNA map of tau covered the cell body and extensions, while STARFISH detected active tau synthesis only in dendrites. A three-dimensional reconstruction separated the dendrites from intersecting axons. The authors labeled newly assembled tau for 30 seconds and tracked its fate, finding that almost half of the label disappeared within 30-60 seconds. The substance iBEp, which blocks the proteasome, stopped this rapid degradation, with the authors estimating that the proteasome degrades about a third of new tau.
🔗 Read original →
PubMed Central (PMC)
Dendritic translation and neuroproteasome-mediated degradation of endogenous Tau revealed by STARFISH
In Alzheimer’s disease, the protein tau is thought to redistribute from axons to the somatodendritic compartment and form fibrillar aggregates. Although tau aggregation is a hallmark of Alzheimer’s disease, the dynamics of its synthesis and ...
China Unifies Lab Test Pricing
The National Medical Insurance Administration of China released a trial guide on laboratory service prices on August 14. The document consolidates previous regional positions into 662 main tariffs, 114 surcharges, and 7 extensions. Provinces must determine a uniform base price level, and municipalities with pricing authority must set actual rates around it.
Previously, in most regions, the same test was paid for at different rates depending on the methodology. The new guide introduces a single price for a measurable indicator, regardless of technology: for example, a blood glucose test now has a single tariff instead of separate rates for different measurement methods. The payment amount now also depends on data preservation. In an official clarification, the agency introduced a rule: clinics are required to preserve and transmit analysis data and results to the system.
If a laboratory does not preserve and upload this information, the insurance payment for each indicator is reduced by 10%, but no more than 5 yuan. When researching multiple indicators in one analysis, the total deduction is also limited to 5 yuan. The regulator introduced this requirement for mutual recognition of analyses between different clinics and regions. To enable another doctor to use an already prepared result, a single number on the form is not enough: a reference interval - a range of normal values - is needed to compare the measurement.
🔗 Read original →
The National Medical Insurance Administration of China released a trial guide on laboratory service prices on August 14. The document consolidates previous regional positions into 662 main tariffs, 114 surcharges, and 7 extensions. Provinces must determine a uniform base price level, and municipalities with pricing authority must set actual rates around it.
Previously, in most regions, the same test was paid for at different rates depending on the methodology. The new guide introduces a single price for a measurable indicator, regardless of technology: for example, a blood glucose test now has a single tariff instead of separate rates for different measurement methods. The payment amount now also depends on data preservation. In an official clarification, the agency introduced a rule: clinics are required to preserve and transmit analysis data and results to the system.
If a laboratory does not preserve and upload this information, the insurance payment for each indicator is reduced by 10%, but no more than 5 yuan. When researching multiple indicators in one analysis, the total deduction is also limited to 5 yuan. The regulator introduced this requirement for mutual recognition of analyses between different clinics and regions. To enable another doctor to use an already prepared result, a single number on the form is not enough: a reference interval - a range of normal values - is needed to compare the measurement.
🔗 Read original →
www.nhsa.gov.cn
国家医疗保障局 医保动态 国家医保局印发《检验类医疗服务价格项目立项指南（试行）》
Model ESM-2 Accuracy
The compressed version of the ESM-2 model maintained average accuracy but severely distorted the ranking of substitutions in the UBR5 protein. On August 15, Cen Shao compared six computation modes for three ESM-2 versions in a preprint. The evaluation covered 201 laboratory measurements from the ProteinGym dataset on the effect of amino acid substitutions in proteins.
The ESM-2 model reads the amino acid sequence and ranks possible substitutions by expected effect. The laboratory can choose the top lines of such a ranking for the next experiment, so it needs an accurate order of variants for its protein. Shao ran 2.41 million variants through three ESM-2 sizes and six precision modes. In the 3 billion parameter model, 8-bit compression almost did not change the average correlation with laboratory measurements: the difference with full precision was −0.0021.
At the human UBR5 protein, the same correlation fell from 0.591 to 0.223. The average metric described the entire set of experiments, and the ranking of variants for one protein changed radically. The reason lies in the scoring method. The model receives a substitution score by subtracting the close logarithms of the probability of the original and new amino acid. A small compression error in each of these numbers after subtraction can become a noticeable part of the final difference and change the order of candidates. As reported in the ProteinGym dataset, such errors can significantly impact the results.
🔗 Read original →
The compressed version of the ESM-2 model maintained average accuracy but severely distorted the ranking of substitutions in the UBR5 protein. On August 15, Cen Shao compared six computation modes for three ESM-2 versions in a preprint. The evaluation covered 201 laboratory measurements from the ProteinGym dataset on the effect of amino acid substitutions in proteins.
The ESM-2 model reads the amino acid sequence and ranks possible substitutions by expected effect. The laboratory can choose the top lines of such a ranking for the next experiment, so it needs an accurate order of variants for its protein. Shao ran 2.41 million variants through three ESM-2 sizes and six precision modes. In the 3 billion parameter model, 8-bit compression almost did not change the average correlation with laboratory measurements: the difference with full precision was −0.0021.
At the human UBR5 protein, the same correlation fell from 0.591 to 0.223. The average metric described the entire set of experiments, and the ranking of variants for one protein changed radically. The reason lies in the scoring method. The model receives a substitution score by subtracting the close logarithms of the probability of the original and new amino acid. A small compression error in each of these numbers after subtraction can become a noticeable part of the final difference and change the order of candidates. As reported in the ProteinGym dataset, such errors can significantly impact the results.
🔗 Read original →