167 concepts · 0 complete · 146 partly written · 21 not yet written
Concepts
Single ideas explained: the chemistry, physics and cell biology the rest of the atlas leans on.
-
"Quantum fingerprint": the metaphor, quantum-chemistry descriptors and quantum hardware, kept apartThree things share the word: a metaphor, descriptors from approximate electronic structure on ordinary computers, and quantum circuits on hardware; none beat ordinary fingerprints on an activity task. partly written
His idea, in his words. "conditional graph connections or entanglements, similar to quantum"; "It would be cool to invent a quantum fingerprint of some kind."
-
3' end cleavage, poly(A) tail control and mRNA exportThe transcript is cut at a read signal, given a tail that settles at about 50 to 110 nt in human cells, and exported; the tail tracks stability and predicts translation only in early embryos. partly written
-
AAV vectors not yet written
To write this, these must be answered:
- How does an AAV capsid get from blood to the nucleus, and what are the measured dose limits and toxicities?
-
Allostery in PSI-MI and the Allosteric DatabaseAllostery has a defined vocabulary inside PSI-MI and a database of allosteric proteins, sites and modulators; the PSI-MI vocabulary is the best-defined of the hard structural concepts. partly written
The vocabulary. PSI-MI, the molecular-interaction vocabulary, defines allostery, allosteric effector, allosteric change in structure, allosteric change in dynamics, homotropic and heterotropic types, pre-assembly and conformational status, among 1,660 terms in the Ontology Lookup Service. The 2013 paper that added cooperative interactions says such data were "lacking from the main bioinformatics resources". The vocabulary has text definitions, not logical ones.
-
Alpha helices and beta sheetsAlpha helices and beta sheets are the two regular ways a protein backbone pairs its N-H and C=O groups with each other; which one a stretch forms depends on side chains and context. partly written
The point. The backbone of every protein carries donors (N-H) and acceptors (C=O). Helix and sheet are the two ways to satisfy them regularly. The side chains decide which is favoured; the backbone only makes both possible.
-
Alternative start codons and two protein formsRibosomes can start a protein at a different place than the first AUG, so one message can give two N-terminal forms; this is real but usually minor, and differs from his idea. partly written
His idea, kept as he said it. He said you may get two start codons, and so two versions of the protein; he also mentioned a reverse truncation. Kept in his words and not used as evidence.
-
Animal-to-human translation figuresThe published translation figures count different things; this record gives each with its definition and denominator. partly written
The short answer. No single translation rate exists. Each published figure counts a different starting pool, so figures cannot be averaged or compared without their denominators.
-
Antibody-drug conjugates not yet written
To write this, these must be answered:
- What is the measured fraction of an antibody-drug conjugate dose that reaches tumour, and from which primary source?
-
Assay metadata as part of the representation: batch labels and what has been shown to helpMethods that use assay metadata use one field, a batch label, and usually subtract it; metadata such as operator, site, date or quality as typed content of the representation has no measured gain on record. partly written
His idea, in his words. "It would be cool to invent new ontologies that somehow take advantage of different assay metadata, because assays are done by different people in different places at different times and have different longitudinality and data quality."
-
Basic Formal Ontology (BFO)BFO is the upper ontology many OBO ontologies align to: not yet written here. not yet written
To write this, these must be answered:
- What does BFO commit an ontology to, and is there a measured benefit of aligning to it?
-
Binding kinetics: on rate, off rate and residence time partly written
Short answer. Kd equals koff divided by kon. kon is how fast partners find each other, koff how fast the complex falls apart, and 1 over koff is the average lifetime. A lifetime of one picosecond needs a Kd of hundreds of molar, so it is a collision, not a bond. A lifetime of seconds is ordinary for tight signalling binders.
-
Bioisostere not yet written
To write this, these must be answered:
- What is a bioisostere, and which published programme documents one with the before and after numbers? The Pfizer coronavirus-protease series changed warhead class several times, but the figures in the chain could not be re-opened and are marked unverified.
-
Biolink ModelThe Biolink Model is a schema, not a graph: a hierarchy of classes and predicates, with qualifiers, used to standardise knowledge graphs. partly written
Short answer. Biolink is a versioned schema of classes and predicates arranged under related_to, with qualifiers and a knowledge level on edges. It standardises what graphs like RTX-KG2 call things; it contains no data itself.
-
Biomedical knowledge graphs: shapes, edges and what has workedA biomedical knowledge graph joins genes, drugs and diseases by edges that are claims; what an edge is, and how its claim is checked, differs a great deal between graphs. partly written
Short answer. A biomedical knowledge graph is a network of things (genes, drugs, diseases) joined by edges that are claims. The deep difference is what an edge is: a bare relation copied from a database, a relation with its source, or a statement with its evidence and a belief score. Headline benchmark scores flatter; honest splits sit lower. No documented clinical decision was found that rested on a public graph.
-
Cell line misidentification and driftDocumented ways immortalised cell lines mislead: wrong identity, genetic drift between labs, contamination. partly written
The short answer. Cell lines can be the wrong cells, or drift away from what the label says. Both are documented and both change experimental results.
-
ChEBIChEBI is the ontology of small molecules of biological interest: not yet written here. not yet written
To write this, these must be answered:
- What does ChEBI cover, who maintains it, and what usage measure exists beyond citation counts?
-
Chemistry-trained reaction models applied to enzymes and metabolism: what was measuredModels pretrained on patent chemistry help on enzymatic and metabolite prediction when mixed with enzyme data, and most of the evidence is the authors' own benchmarks. partly written
Does chemistry-trained prediction transfer to enzymes?. Yes, partly. Patent-only models score about 10% on enzyme tasks, mixed training reaches about 50%, and enzyme text lifts it to 62.2%. A natural-product pathway model reached top-10 of 60.6% with general-chemistry pretraining and 0 without it on one subset. Nearly every benchmark was run by its own authors.
-
Chromatin accessibilityChromatin accessibility is how physically reachable stretches of DNA are to a probe such as a transposase, given the proteins and nucleosomes packaged on them. partly written
What it is physically. Accessibility is steric: DNA covered by nucleosomes or bound proteins is harder for an enzyme to reach. ATAC-seq counts where Tn5 got in; fragment lengths about 200 base pairs apart indicate nucleosome spacing.
-
Clearance and half-life not yet written
To write this, these must be answered:
- How are clearance, volume of distribution and half-life defined and measured, and how do they relate to the extraction ratio?
-
Cluster of differentiation (CD) namesA CD number names a group of antibodies that recognise the same cell-surface molecule, in the order the workshops assigned them, not a gene family. partly written
In short. CD numbers come from antibody workshops, so one CD number names a molecule as an antibody sees it. In gene names it survives as an alias or a previous symbol: CD20 is MS4A1, CD117 is KIT, CD45 is PTPRC.
-
Conformational ensembles and molecular dynamics: databases and an alpha ontologyEnsembles and dynamics have databases and one alpha-stage simulation ontology; no metadata standard for molecular dynamics was found. partly written
Ensembles. CoDNaS 2.0 defines a conformer as a chain at 95% sequence identity with RMSD and similarity scores per pair, covering about 70% of PDB protein structures in 2016. PED 2024 holds 461 entries and 538 ensembles of intrinsically disordered proteins, with a validation pipeline and a controlled vocabulary its authors say is organised into an ontology; the vocabulary page was not opened.
-
Conformational restriction not yet written
To write this, these must be answered:
- What does locking a flexible group into a ring do to potency and absorption, with a documented example? The nirmatrelvir bicyclic proline is the likely case, but its numbers come from a preprint that could not be re-opened.
-
Conformational state labels for kinases and GPCRsKinases and GPCRs have rule-based state labels from distance cutoffs, with the authors noting the criteria are not widely agreed. partly written
Kinases. KinCore labels kinase structures DFGin, DFGinter or DFGout (DFG is an Asp-Phe-Gly motif) and by dihedral clusters, from distance cutoffs and cluster centroids; it held 7,177 human chains from 278 genes in April 2021. Its authors write that the criteria for "active" and "DFGin/out" are "not widely agreed upon".
-
Conservation of the gene-expression machinery across bacteria, archaea and eukaryotesA shared core (RNA polymerase, the ribosome's catalytic centre, the Spt5/NusG clamp, the SecY/Sec61 channel) with mostly unrelated helpers, and no clean cross-domain rate comparison. partly written
In short. One core, two shells. The RNA polymerase core and the ribosome's catalytic centre are shared by structure across all three domains of life, as are the Spt5/NusG clamp factor and the SecY/Sec61 channel. Initiation control, termination, capping, cap-scanning and the nucleus are not. Archaea are a mosaic of bacterial and eukaryotic parts. 'LUCA had it' is always the authors' inference.
-
Context-dependent edges: storing the conditions under which a statement holdsA fact stored with the species or cell type it was seen in; every large schema has the slot, and what is open is how often it is filled and used. partly written
His idea, in his words. "Can we make data more dynamic so a certain piece of data is connected to all these other pieces of data, but only in this context? You have conditional graph connections or entanglements, similar to quantum."
-
Cost per usable data pointWhat an experiment really costs per result you can use, which depends on what the quoted price counts and what fails. partly written
-
DNA methylationDNA methylation is a methyl group added to cytosine, mostly where a cytosine is followed by a guanine (a CpG), which changes how readable the nearby DNA is without changing its letters. partly written
What it is. A methyl group on cytosine, mostly in a CpG (a cytosine followed by guanine). A first oxidised form, 5-hydroxymethylcytosine (5hmC), is made by TET enzymes. Because the tag is chemically invisible to a sequencer, assays convert it into a letter change.
-
Do cells move? not yet written
To write this, these must be answered:
- Do cells move around? How would a T cell reach and kill a cell in tissue?
-
Does the molecular representation or the data matter more? What chemistry comparisons showNo comparison was found that varies only the representation; near misses suggest evaluation design matters more than representation or algorithm. partly written
His idea, in his words. "You have innovation at the algorithm layer. I find you can innovate at the algorithm layer all you want, but that can only be as powerful as the data representation. The data representation, of course, can only be as powerful as the data, but at least there's more space to innovate there"
-
Drug formulation and carriers not yet written
To write this, these must be answered:
- What are tablets, salts, prodrugs, liposomes, antibody-drug conjugates and viral vectors, and when is each used? Is there a measured share of products using a carrier?
-
EC numbers (enzyme classification)An EC number classifies an enzyme by the reaction it catalyses, not by its sequence or gene. partly written
In short. EC 2.7.11.22 is 'cyclin-dependent kinase': four numbers, from class to serial number, fixed by the reaction. Unlike a CDK1/CDK2 gene number, an EC number is a real hierarchy, but of chemistry.
-
Efflux pumps and drug transporters not yet written
To write this, these must be answered:
- Which transporters (P-gp, BCRP, OATP and others) matter for drugs, what do they do, and how are they tested?
-
Epigenetic inheritance across generationsWhether non-DNA-sequence information passes through the germline to later generations in mammals is read in two ways by published reviews: a limited effect exists, or it has not been demonstrated. partly written
Two readings. Reading A: epimutations are inherited from yeast to humans, with a more limited effect in mammals. Reading B: for mammals the evidence is inconclusive. Both are published reviews or comments; neither reports a new experiment.
-
Epigenome editingEpigenome editing writes or removes a chromatin mark at one chosen gene with a programmable protein, which is the main interventional test of whether a mark can hold a gene state. partly written
What exists. Four studies are opened here: Amabile 2016 (three genes silenced in cells), Liu 2016 (methylation edited at named loci, including in mice), Nunez 2021 (CRISPRoff) and Cappelluti 2024 (mouse liver). All edit one gene or a few, in cells or mouse.
-
Equilibrium density-gradient centrifugationIn cesium chloride density-gradient centrifugation a salt solution spun for many hours forms its own density gradient, and DNA collects as a band where the solution density equals its own. partly written
-
Evidence and Conclusion Ontology (ECO)ECO names kinds of evidence behind a claim: not yet written here. not yet written
To write this, these must be answered:
- How does ECO relate to the Gene Ontology evidence codes, and which resources use it?
-
Experiment and assay ontologies (OBI, BAO, EFO)OBI, BAO and EFO describe investigations, bioassays and experimental factors: not yet written here. not yet written
To write this, these must be answered:
- What do OBI, BAO and EFO record about who ran an assay, where, when and with what quality, and how often are those fields used?
-
Family stem plus number (CDK1, CDK2)A stem such as CDK names a family; the number after it is mostly a label assigned in a proposal, not a similarity rank. partly written
In short. CDK means cyclin-dependent kinase, and the stem is a fair guide to the family. The number is a historical slot: CDK1 was called CDC2 until 2009, and nothing in the sources read says neighbouring numbers are the closest relatives.
His reading. He asked whether CDK1, CDK2 and the like are cyclin-dependent kinases, and said there are many such repeating-element names.
-
Food composition databases are not reaction databases: what public food data holdPublic food databases list what foods contain; they do not carry reaction rates, and company process data are not public. partly written
Can food databases supply reaction data to biology?. Not as far as the pages opened show. FoodData Central lists composition, and no paper was found using FooDB or Phenol-Explorer to parameterise a pharmacokinetic model. Company reaction data are not public, and the search stopped there.
-
Free-energy calculation (relative binding free energy)Physics-based simulation that estimates how much tighter one close analogue binds a target than another. partly written
In short. Free-energy calculation ranks close analogues by simulated binding energy. Its error is close to, not below, the noise between assays. Evidence is mostly vendor-authored or retrospective, and one blind challenge favoured machine learning for potency.
-
Gene and protein names: the decoderA decoder for gene and protein names: who assigns them, what each pattern encodes, and where a name misleads. partly written
The one rule. A name is a historical label under a priority rule, not a measurement. Join data on the stable ID (for human genes, the HGNC ID), and read the symbol as a readable layer on top.
His framing. He said that knowing a few naming patterns, like p53 or p47, and the repeating elements such as CDK1 and CDK2 would let a reader get insight faster, and that some very rare genes may not yet be understood or named.
-
Gene annotation, evidence codes and what 'unknown function' meansAn annotation ties a gene to a function with a graded record of how well backed it is, and a count of 'unknown' genes is the output of a labelling rule applied to that record. partly written
-
Gene duplication and paralogues not yet written
To write this, these must be answered:
- How does duplication produce family members such as CDK1 and CDK2, and why does that create redundancy?
-
Gene essentiality and the quasi-essential classHow a gene is called essential, quasi-essential or non-essential from where transposons land, and why each call depends on the method and the conditions. partly written
-
Gene Ontology enrichment analysis, and how it failsEnrichment analysis asks whether a gene list hits a GO term more often than chance, so its answer depends on the annotation table, the background and the release. partly written
How it works. Choose a gene list and a background (the genes that could have been in the list). Count how many list genes carry each term, test each term with a hypergeometric, chi-square or binomial test, and correct for the many terms tested. The statistics are simple; the answer depends on the annotation table behind the counts.
-
Genetic interaction and the combinatorial spaceA genetic interaction is a double mutant whose fitness differs from what its two single mutants predict; the number of possible combinations grows far faster than anyone can test. partly written
-
Germline and somatic variantsA germline variant is inherited and present in essentially every cell; a somatic variant arises later in some cells, and a blood sample can carry both. partly written
His sentence, kept as he said it. He described taking a blood sample, observing that each cell carries somatic mutations, and trying to sift which variants are present across all cells. Kept here in his words and not used as evidence.
-
Germline epigenetic reprogrammingIn mouse primordial germ cells DNA methylation is erased genome-wide by TET-driven oxidation and replication-coupled dilution, imprints last to go, and rare elements escape. partly written
What is erased. In mouse primordial germ cells genome-wide DNA demethylation resets the epigenome. Loss of methylation takes place during germ-cell expansion and migration, with evidence that much of it is passive.
-
Graft rejection and immunosuppression: why donor cells are attacked and how that is managedCells from a donor carry foreign HLA identity tags, so the recipient immune system attacks them unless it is suppressed, the donor is matched, or the cells are edited; a patient own cells avoid this. partly written
Rejection in brief. Every cell carries identity tags called HLA. If cells come from another person the recipient T cells, natural killer cells and antibodies can recognise them as foreign. Cells from the patient own body carry the patient tags and are not attacked for that reason.
-
HetionetHetionet is a 2017 heterogeneous network of 47,031 nodes of 11 types and 2,250,197 edges of 24 types, built to rank drugs for repurposing. partly written
Short answer. Hetionet integrates 29 resources into 47,031 nodes and 2,250,197 edges. The Rephetio model ranked drugs by degree-weighted path counts. Its authors printed its flaws: a degree-only prior scored 97.9% on training positives but 54.1% on external indications.
-
Histone modificationsHistone modifications are chemical groups on histone tails, added and removed by enzymes and bound by reader proteins; whether they cause gene states or record them is contested. partly written
The notation. A mark is named by histone, residue and group: H3K27me3 is histone H3, lysine 27, three methyl groups. H3K9me is a methyl on lysine 9 of H3.
-
Histone recycling and the PRC2 read-write loopAt replication parental histones are shared between the two daughter DNAs and mixed with new unmarked histones, so marks are diluted and then rewritten, in part by PRC2 reading its own product. partly written
Dilute, then rewrite. Replication builds nucleosomes from old modified histones and new unmodified ones. Where a mark was is reproduced with high accuracy; how much of it is there is restored afterwards, at mark-specific speeds.
-
How a link-prediction benchmark is built and how it leaksA link-prediction benchmark hides true edges and asks a model to rank them back; mirror edges, node degree and random splits inflate the score without any biology learned. partly written
Short answer. Hide some true edges, rank them back, score the rank. Three things inflate the score: test edges whose mirror image is in training, node degree, and random splits. Always ask for a degree-only baseline and a split by time or by disease.
-
How a multimodal single-cell or spatial data object ties its parts togetherAnnData, MuData and SpatialData tie parts by a name or an integer id and by named coordinate systems; none stores how sure that tie is. partly written
The tie in one paragraph. AnnData matches annotations to the matrix by a string index. MuData matches cells across modalities by cell name. SpatialData matches a table row to a mask pixel value or shape index by an integer id, and places images, masks and points in named coordinate systems. None of them checks that the match is true, and none stores an uncertainty for a mask or a transform.
-
How a pathway is encoded: BioPAX, SBML, GO-CAM and Reactome's own modelA pathway is not one thing stored two ways: BioPAX keeps entities, hierarchy and order, SBML a runnable skeleton, GO-CAM causal activities, and each conversion drops something. partly written
Summary. On one Reactome pathway of 16 reactions, the SBML export kept all 16 reactions but flattened nested complexes, dropped the pathway hierarchy and step order, and carried no rate laws, while the BioPAX export kept them. The one measured Reactome-to-GO-CAM conversion gave 3,593 complete activity units from 11,570 reactions, by the converter's own authors.
-
How a phenotype ontology turns features into a ranked diagnosis, and the evidencePhenotype similarity plus variant scoring ranks candidate diagnoses; the measured yields come from solved cases, and no outcome study was found. partly written
The four steps. A clinician writes findings as HPO terms. A similarity score weights rare shared terms more than common ones. A ranker such as Phenomizer applies it to the query and each disease. In genome work a phenotype score is combined with variant rarity and predicted damage (Exomiser) and a person reviews the ranked list.
-
How a reaction gets from a patent paragraph into a database record, and what goes wrongA patent paragraph becomes a reaction record through a chain of guesses (find, tag, convert names, assign roles, map atoms), and only pieces of that chain have been measured. partly written
How much can a patent-derived reaction record be trusted?. Treat these records as text-derived, not bench truth. The original pipeline's author checked 100 reactions and found 22% flawless. The atom mapper behind the public set got 38.47% of a hand-mapped benchmark right. Failed reactions are probably near absent.
-
How a reaction is written for a computer: reaction SMILES, atom maps and templatesA reaction is stored as text, an atom map says which atom goes where, and a template generalises the reaction centre; each loses something. partly written
What can each way of writing a reaction not say?. Reaction SMILES omits reagents and by-products that were not listed. Atom maps are imperfect even when made by good tools. Templates lose context beyond the reaction centre. Condition and yield models barely beat guessing the average.
-
How a task-environment scorer works and how it failsA scorer turns output and a hidden answer into one number, and audits show that number can rise without understanding. partly written
What kinds of scorer are there?. Exact match compares with a stored answer. An executable check runs the output against hidden tests. A rubric awards weighted points. A model judge compares output with a reference. A held-out experiment trains or tests on hidden data. Each fails differently.
-
How a variant benchmark leaks: circularity, gene-level labels and trivial baselinesA variant benchmark leaks when its labels reach the predictor, and measured repairs split by gene, remove training overlap or use an independent assay. partly written
What are the two kinds of circularity?. Type 1 circularity puts the same variants in training and test. Type 2 puts different variants of the same protein on both sides. Grimm and colleagues show type 2 matters because most proteins carry a single class.
-
How agentic science systems hold what they knowA comparison of what Robin, Kosmos, the Google AI co-scientist, the AI Scientist-v2, the Virtual Biotech and Biomni keep between steps, and what they do not keep. partly written
What is stored, system by system. In the papers opened, these systems keep prose and run files, not typed claims with a revision rule. The nearest to typed is the Virtual Biotech, which stores schema-validated JSON with a source tag per annotation. Kosmos says 'structured world model' and publishes no schema.
-
How DNA methylation is copied through replicationAfter DNA replication each daughter duplex is half-methylated, and a reader (UHRF1) and a writer (DNMT1) re-methylate the new strand, mostly within minutes but with a measurable failure rate. partly written
The half-methylated template. A CpG pairs with a CpG on the other strand, so methylation is symmetric. Replication leaves each daughter with one methylated old strand and one blank new strand, a hemimethylated site. That half-pattern is the template the copy machinery reads.
-
How entity grounding works and fails in knowledge graphsGrounding turns a string such as "ERK" into one identifier; most graphs publish no accuracy for this step, and the documented failures are shared aliases, wrong species and family versus member. partly written
Short answer. Grounding maps a text string to one database identifier, at reading time and again when databases are merged. Measured numbers exist only for the text-reading tools. PrimeKG and RTX-KG2 report no precision or recall for their identifier mapping.
-
How far pathway and reaction databases agree with each otherFive human metabolic databases agree on 1 to 3% of reactions; part of the gap is definitional, and database choice did not always change a prediction. partly written
Summary. In one 2011 audit of five human metabolic databases, all five agreed on 1% of 7,758 reactions, or 3% of 6,968 ignoring electrons, protons and water, and on 5 of 30 reactions of the tricarboxylic acid cycle. Part of the disagreement is what counts as the pathway. Choice of database did not significantly change a survival prediction in one test.
-
How Helicobacter pylori lives in the stomachHelicobacter pylori survives stomach acid by managing it: it makes urease, and the lining of the stomach carries a steep pH gradient in which the organism lives. partly written
-
How MOFA+, WNN and totalVI join two layers, and how they are benchmarkedAll three join by the shared cell or sample, keep each layer's own features, and have been compared mostly by their own authors; independent tests with a trivial baseline are patchy. partly written
The join and the difference. MOFA+, Seurat's weighted-nearest-neighbour method (WNN) and totalVI all join two layers by the cell barcode or sample id and keep each layer's own columns, so the gene-to-protein many-to-many mapping never enters. They differ after the join: a linear factor model, a per-cell weighted neighbour graph, and a deep generative model. All three are one group's choice; no standards body defines any of them.
-
How mutations change protein stabilityA mutation changes the free energy of folding by an amount (ddG) that depends on what the residue was doing, and truncations act by a different mechanism than substitutions. partly written
The three rules of thumb, checked. Polar-for-hydrophobic in the core: supported, the biggest signal in the data. Charge flip: qualified, it depends on burial and on whether a partner exists. Truncation: right that it is worst, but the mechanism is loss of the protein (or a dominant truncated product), not a stability change.
-
How omics layers are joined: by identifier, sample, cell, position, latent space and graphSix ways layers are joined, the key each uses, what it must assume, and where it is documented to break. partly written
The six joins. Layers are joined by identifier (a gene stands for its protein), by sample, by cell barcode, by position, by a statistical latent space, and by a knowledge graph. Each assumes something the data may not give, and measured results exist for only a few.
-
How one fact becomes an edge with evidenceIn INDRA a fact becomes a typed Statement with a list of Evidence objects and a belief score, so the edge can be ranked by trust without deleting anything. partly written
Short answer. A sentence is read, its names are grounded to identifiers, it is filled into a typed Statement, each supporting sentence or database row is kept as an Evidence object, identical Statements are merged and a belief score is computed. The edge is a summary; the evidence list is the record.
-
How PaperQA2 turns a paper into something an agent reasons overPaperQA2 cuts papers into text chunks, has an LLM summarise each against the question, and reasons over those cited summaries; nothing is written back to a store. partly written
The steps. Search by keyword for open-access papers, parse the PDF, cut into windows of about 2,250 tokens, embed, rank, then have an LLM write a 200 to 400 token summary of each chunk in light of the question and score it from 0 to 10. The unit of knowledge is that cited summary, built fresh per question.
-
Imageromics (Amit's word) and the technologies it coversAmit's coinage for spatial, imaging-based omics; no published source read uses it, and here it is read as omics measured by imaging that keeps position. partly written
His word. He uses the phrase "spatial imageromics" together with spatial proteomics and spatial metabolomics, as new technologies whose data and pathways need representing. This is his term, kept as he wrote it.
-
In vitro, cell-based and in vivo not yet written
To write this, these must be answered:
- What does each connote, what can each kill a project for, and what can each not tell you?
-
INDRAINDRA assembles mechanistic Statements with evidence and belief scores from text readers and databases, on demand rather than as one fixed graph. partly written
Short answer. INDRA stores a Statement (a typed, grounded mechanism) plus a list of Evidence objects, one per sentence or database entry that said it, and a belief score computed from how many Evidences and which sources. The 2023 corpus had about 9e5 unique Statements from 567,507 articles plus databases.
-
JCVI-syn3A, the minimal synthetic cellA bacterium built at the J. Craig Venter Institute with a genome cut down to little more than the genes that still allow growth, used as the simplest living cell to model. partly written
What it is and why it exists. A minimal cell is a bacterium with as few genes as can still grow. The first free-living genome sequenced, Mycoplasma genitalium (580,070 bp, about 470 coding regions), already had 152 genes (32%) with no assigned role. JCVI-syn3A descends from a stripped Mycoplasma mycoides genome.
-
Joining host genetics to pathogen genomics: paired genome studies and where the join breaksPaired host-variant and pathogen-variant studies exist for HIV, hepatitis B, tuberculosis and malaria; a shared schema joining them to interaction resources is open. partly written
His idea, in his words. "in terms of genetics, can we marry genetics and infectious disease data?"
-
Knowledge-graph shape: the hypergraph and the n-ary relationAn n-ary relation joins several nodes at once in one statement; no native biomedical graph of this kind was found. partly written
Short answer. A hyperedge ties together enzyme, substrate and residue in one statement instead of splitting them over several edges. No native hypergraph biomedical knowledge graph was found; INDRA's Modification and Complex Statements are the nearest, because one Statement holds several named arguments.
-
Knowledge-graph shape: the plain tripleA plain triple states one typed relation between two nodes and nothing else: no residue, paper, certainty or context. partly written
Short answer. A triple is (subject, predicate, object), for example MAP2K1 phosphorylates MAPK1. Hetionet-style typed edges are close to this. It is simple to build and to hold out for link prediction, and it says nothing about residue, evidence, certainty or cell type.
-
Knowledge-graph shape: the property graphA property graph is a triple whose edge carries free-form key-value properties, such as a source, a paper id or a score. partly written
Short answer. A property graph keeps the same edge but lets it carry properties, such as residues, a PubMed id or a source database. SPOKE runs on Neo4j, a property-graph database, and some of its edges carry p-values or scores. Properties are free-form, and a property cannot itself be pointed at by another claim.
-
Knowledge-graph shape: the reified statement with qualifiersA reified statement turns a claim into an object that can carry qualifiers, references and a rank, as Wikidata and the Biolink Model do. partly written
Short answer. The claim becomes a node of its own, so qualifiers (residue = T185), references and a rank can attach to it. Wikidata does this natively, and the Biolink Model adds qualifiers to a core triple.
-
Knowledge-graph shape: the signed mechanistic statement with evidenceA mechanistic Statement names the kind of interaction, its arguments and its modification, and carries a list of evidence and a belief score. partly written
Short answer. INDRA stores Phosphorylation(enz, sub, residue, position) plus a list of Evidence objects and a belief score; whatever the source did not say is left blank. It does not claim causation under intervention, and multi-step paths are not treated as explanations.
-
Kosmos and its 'structured world model'Kosmos is Edison Scientific's AI scientist; it says it keeps a structured world model between cycles and does not publish what is in it. partly written
What the paper says. Each cycle runs up to ten literature or analysis tasks, then updates the world model with summaries of the task outputs and queries it to plan the next tasks. The paper's methods are per-figure; no section describes the world model's fields or update rule.
-
Learned representations against the simple baseline, per omics layerIndependent tests mostly find single-cell, DNA and protein language models level with or behind a simple baseline, with gains where the data are diverse. partly written
Summary. Two independent tests found single-cell foundation models no better than variable-gene selection or an additive baseline. A test of DNA models found them matching or trailing a one-hot convolutional network, with exceptions. Protein models beat a site-independent baseline modestly, trail alignment-based methods, and clearly win only on diverse sequences. Which model wins also depends on who runs the test.
-
Lipid nanoparticlesA lipid nanoparticle is a fat bubble that stays nearly neutral in blood and turns positively charged inside the acidic endosome, and it is used to deliver nucleic acids such as siRNA and mRNA. partly written
The short version. A lipid nanoparticle carries RNA that is too large and charged to cross a membrane. Its ionisable lipid is neutral in blood and charged in the endosome. Blood ApoE steers it to the liver. Most RNA is still lost after uptake.
-
Logical definitions and reasoners in ontologiesA logical definition states what a term is as a kind of something with a property, so a reasoner can place the term in the hierarchy without anyone typing the link. partly written
How it works. A logical definition has the form: X is equivalent to a kind G that has relation R to Y. The Cell Ontology authors give "nucleate erythrocyte", an erythrocyte that has a nucleus. Because "nucleate cell" is defined as a cell that has a nucleus, a reasoner concludes the first is a kind of the second, with no one entering that link.
-
Lymphodepletion: clearing space before cells are infusedLymphodepletion is chemotherapy given days before an engineered cell infusion to reduce the patient own lymphocytes, which in the studies read improved expansion and persistence of the infused cells. partly written
Lymphodepletion in brief. Before CAR T cells are infused, the patient gets chemotherapy that lowers their own white cells. It is part of the treatment: the regimen changes how well the infused cells expand and how long they last.
-
Lysosomal trapping not yet written
To write this, these must be answered:
- How much do weak bases accumulate in lysosomes, measured, and when does it change volume of distribution or effect?
-
Machine learning and language models in building and mapping ontologiesLanguage models can propose ontology terms and recognise terms in text at measured, moderate accuracy, and benchmark leakage can halve apparent gains. partly written
Proposing new terms. DRAGON-AI, a retrieval-augmented GPT-4 system, proposed is-a relations across 10 ontologies with precision 0.894 and recall 0.5; GPT-3.5 scored 0.846 and 0.419; a reasoner baseline had precision 1.0 by definition and recall 0.337. Less confident raters scored the AI definitions as good as human ones.
-
Macromolecular crowding and diffusion in bacterial cytoplasmHow fast molecules move and meet inside a bacterium, what a physics-only account predicts, and where it has to be supplemented by measured or fitted inputs. partly written
-
Maillard chemistry and glycation: one reaction family in a food factory and in a cellA sugar and an amine react without any enzyme, which makes colour and flavour in food and damages proteins in cells, where enzymes defend against it. partly written
Is food Maillard chemistry the same as glycation in the body?. The first steps are the same chemistry. The conditions differ, and the cell has enzymes a factory lacks. Nobody in the sources opened put food-side and body-side rates on one axis, so the quantitative analogy is untested there.
-
Many representations of one object, chosen per task or per querySeveral encodings of the same molecule or entity are kept and one is picked for the question; per-task winners and concatenations are documented, a trained per-query chooser was not found. partly written
His idea, in his words. "The solution may be to represent data in many different ways, so it exists in the superposition of many different states. You sample the data for different representations, and you take the one that's most similar at a certain point."
-
Mass-based protein names (p53, p47)A p-number is the size a protein ran at on a gel when it was first seen, not its real mass or a unique identity. partly written
In short. In p53 the number is the apparent molecular weight on a gel (53,000), read in 1979. The sequence of the same protein weighs 43,653 Da. Treat a p-number as a 1970s-80s lab label, not a measurement of mass.
-
Matched molecular pairsTwo compounds that differ by one small defined change, pooled across many projects to see what that change usually does. partly written
In short. A pair is two molecules one edit apart. Pooling pairs across projects turns chemists' experience into a table of typical effects. AstraZeneca authors collected pairs for solubility, protein binding and oral exposure. Most single edits do not help, and a pair is only as good as the two measurements behind it.
-
Metabolite damage and repair: spontaneous chemistry inside cells that reaction resources mostly omitCells make damaged small molecules by spontaneous chemistry or enzyme errors, and repair enzymes undo it, but these loops are typically missing from metabolic reconstructions. partly written
Do metabolic models include damage chemistry?. Mostly not, by the account of practitioners who work on it. A 2021 review says damage and repair loops are typically not yet included in reconstructions. In infected red blood cells, 911 metabolites found matched only 41% of the predicted metabolome.
-
Models of modelsEvery biological number is a measurement of a model system standing in for a human, and each layer has its own error; clinical trials are the first real human data, and they are models too. partly written
The short answer. Every number is a measurement of a model of a human. Both stand-ins have measured errors for some layers. No source combines them, and the human trial is also a model, through its endpoints and its selected patients.
His framing (his words). He said: "these are all models of models. You have a measurement of a model, and the measurement itself is imperfect. Then you have the model of your true model, which is a human, and the model itself is imperfect because you've taken a cell out of its native environment, put it into this somewhat simulatory environment, and you've got problems."
His statement on organoids and H&E (his words). He said: "Yes, organoids will help, but they still don't simulate the circulatory system. H&E helps, but it's still not at the organ level."
His statement on in vitro (his words). He said: "To me, in vitro was basically useless. All it tells you is binding. It's a quick way to kill a project."
His statement on clinical trials (his words). He said: "Anyway, the whole thing looming over all of this is that everything is models of models until you get to clinical trials, and that's real human data."
-
Mondo, the merged disease ontologyMondo merges many disease resources into one ontology with exact mappings, and says it may deliberately disagree with any one source. partly written
What it is. Mondo merges 21 disease resources, giving more than 25,000 concepts from about 100,000 source concepts. It carries 9,195 exact mappings to Orphanet and 8,841 to OMIM, and its 2025 paper states that deployment at two health-data companies is in progress.
-
MPRAbox and CodonBox: the two worked sandboxes in the Rao and Lander paperMPRAbox scores designed DNA sequences with a damp oracle, and CodonBox hides an invented genetic code, so the two show how a score can rise without understanding. partly written
How does MPRAbox score an agent?. The agent writes 50,000 DNA sequences of 200 bases. An oracle model labels them. A fresh network is trained on those labels and tested on 14 hidden sets. The agent sees 14 Pearson correlations and their mean.
-
mRNA-protein correlationHow well mRNA level predicts protein level depends on the axis, the noise treatment and the statistic. partly written
The short answer. There is no single mRNA-protein correlation. It depends on whether you compare genes within a sample or one gene across samples, whether noise is corrected, and which correlation statistic is used.
-
Nanopublication: a statement bundled with its provenanceThe smallest publishable unit in linked data: one assertion with its provenance and publication information as separate named graphs, each carrying a verifiable identifier. partly written
What it is. A nanopublication has a head, an assertion graph, a provenance graph and a publication-information graph. The specification defines no dedicated field for conditions such as species or tissue; such context has to be written as additional statements in the assertion.
-
One-carbon folate dependency in leukemiaLeukemia cells that need the one-carbon folate pathway to make nucleotides, and the contested question of which enzyme to drug. partly written
SHMT1 and SHMT2 in T-ALL. The 2022 paper found that both SHMT1 and SHMT2 must be lost for the drug effect, the arrest was in S and G2, and the authors state arrest more than death and unexplained T-ALL specificity. One cell line was tested in mice, 7 per group, not blinded.
-
Ontology mappings, cross-references and SSSOMA mapping is a claim that a term in one vocabulary matches a term in another; SSSOM adds the kind of match, a confidence and where it came from. partly written
What a mapping is. A mapping is a triple: a subject term, a relationship (the predicate) and an object term, ideally with who made it and how sure they were. The older cross-reference (an "xref") says only that two terms are related, without saying how.
-
Oral absorption and the Biopharmaceutics Classification SystemA swallowed drug must dissolve and cross the gut wall, and the classification system sorts drugs by whether solubility or permeability is the limit. partly written
The short version. A tablet must dissolve, cross the gut wall and survive two sets of enzymes before reaching the circulation. The classification sorts drugs by solubility and permeability; carriers along the gut can change where in the gut a drug is absorbed.
-
Orthology resources: OMA, OrthoDB, PANTHER and how they agreeOMA, OrthoDB and PANTHER give group membership rather than a typed relation, and methods disagree on which gene pairs are orthologs. partly written
What they give. OMA gives hierarchical orthologous groups, descendants of an ancestral gene at a taxonomic level, with a completeness score but no pair confidence. OrthoDB gives orthologous groups at chosen species-tree levels with no confidence for delineation, and its identifiers are not stable between releases. PANTHER gives family trees.
-
PAN-GO: Gene Ontology annotation by evolutionary modellingPAN-GO annotates genes by reviewing gains and losses of function on family trees, reaching 82% of human protein-coding genes against 67% from primary annotations alone. partly written
What it is. In GO, orthology is a method: ISO annotations copy a human experimental annotation to a mouse ortholog and record the ortholog in the "With" field; IBA annotations rest on an explicit model of gain and loss of function on a phylogenetic tree and name the tree node. PAN-GO is the compendium built from such models.
-
Physical-property rules in drug designRules of thumb on size and lipophilicity that guide which analogues get made, and the argument about them. partly written
-
Physiologically based PK modelling not yet written
To write this, these must be answered:
- How do PBPK models work, how often have they failed in regulatory review, and why?
-
Pi-pi stacking and cation-pi interactionsPi-pi stacking and cation-pi contacts are attractions through the electron clouds of aromatic rings; their energies in water are small and the textbook explanation is under dispute. partly written
The picture. Aromatic rings attract through dispersion and electrostatics, but they rarely sit face-on. A cationic Lys or Arg on a ring face (cation-pi) is common in proteins. The in-water energy is small and varies from system to system.
-
Placeholder gene symbols (C#orf#, FAM#, KIAA#)Some human genes still carry a holding symbol that says only that no function-bearing name was available. partly written
In short. C#orf# means open reading frame on chromosome #, FAM# means family with sequence similarity, KIAA# means a Kazusa cDNA clone. HGNC reports 427 protein-coding genes still on such symbols.
-
Post-translational modifications not yet written
To write this, these must be answered:
- What are they, and how do they change what a protein does?
-
PrimeKGPrimeKG is a drug-disease-gene knowledge graph assembled from 20 curated resources, with 129,375 nodes and 4,050,249 relations. partly written
Short answer. PrimeKG (Chandak, Huang, Zitnik, 2023) has 129,375 nodes of 10 types and 4,050,249 relations of 30 types from 20 resources, built from curated databases without text mining. Zitnik is the last author. The paper reports no link-prediction benchmark, and no per-edge evidence is described.
-
Prodrugs not yet written
To write this, these must be answered:
- What is a prodrug, how many marketed drugs are prodrugs, and what is each used to fix?
-
Protein complexes: dimers, tetramers, oligomers not yet written
To write this, these must be answered:
- How do complexes form, how do they "know" when to form, and what regulates them?
-
Protein folding and chaperonesWhat meets a human protein after the ribosome: chaperone binding measured in cells, and the folding times and chaperone-dependent share that were not found. partly written
-
Protein interfaces and assemblies: databases and predictors, not an ontologyInterfaces and assemblies are held in databases and predicted by software; no ontology term for an interface was found, and assemblies are contested. partly written
What exists. ProtCID clusters interfaces across crystal forms at chain and Pfam-domain level; PISA and EPPIC predict which crystal contacts form a biological assembly; PDBe-KB aggregates interface annotations; the IHMCIF dictionary has an interface-residue feature for modelling inputs. None is an ontology with logical definitions.
-
Protein targeting, translocation, quality control at the ribosome and degradationHow a new protein is sent to the ER or mitochondria, how stalled chains are caught, and how long proteins last, with human numbers where measured. partly written
-
Proteomics standards and archives: mzML, mzTab, ProForma, SDRF and ProteomeXchangeHUPO-PSI defines the proteomics formats and ProteomeXchange and PRIDE hold the data; the archive reports growth and fast accessions but no compliance rate was found. partly written
Summary. HUPO-PSI writes the formats: mzML for spectra, the mzTab family for results, ProForma for proteoforms and SDRF-Proteomics for samples. ProteomeXchange had 64,330 datasets submitted through June 2025, 44,248 public, with PRIDE holding 77% (the consortium's own report). No measured compliance rate for any of the formats was found.
-
Public benchmarks and environments for science agentsPublic science-agent benchmarks differ in task, scorer and how far anyone other than the builders has checked them. partly written
Which benchmarks test biology and chemistry agents?. LAB-Bench asks multiple-choice biology questions. BixBench asks open-answer data-analysis questions scored by an LLM judge. ScienceAgentBench scores executed programs. CGBench asks models to assign ClinGen evidence codes from a paper.
-
Purpose as structure: scoring a target differently by indicationThe reason a question is asked decides which evidence counts, yet it lives in tool defaults and design documents rather than as a stored edge. partly written
His idea, in his words. "You also have meta: you have a meta edge depending on what incentive you're looking at the information for. For example, if you're trying to design against a cancer target, you're interested in what essential pathways you have. Versus if you're looking at a rare disease target, you're trying to limit the number of off targets you have."
-
Reaction data: what is public, what is proprietary, and what each holdsPublic reaction data are mostly text-mined patents and curated enzyme reactions; the larger, richer sets are proprietary. partly written
Which reaction data can a biologist actually reuse?. Public: patent-mined reactions (USPTO, CC0 per a blog), the Open Reaction Database (CC-BY-SA), Rhea (CC BY 4.0), BRENDA (free for academics). Proprietary: Reaxys, Pistachio and company notebooks. Each public set has known errors, and only pieces of them have been measured.
-
Reaction-prediction benchmarks: what a headline accuracy measures, and how splits inflate itA reaction model is scored by reproducing the one recorded product, and the score falls sharply when the test reactions are made unlike the training ones. partly written
How much does a reaction model's score fall when the test is made harder?. Molecular Transformer top-1 fell from 90.4% on a random USPTO split to 74.6% when test products were made dissimilar to training. On templates never seen in training, template-free models reached top-10 of only 0.27 to 0.79%. A no-training nearest-precedent lookup scored 37.3% top-1 on USPTO-50k.
-
Receptor and ligand namesLigands and receptors often share a stem family (TNFSF10 and TNFRSF10A), but no general rule was found that links a receptor's name to its ligand's. partly written
In short. A ligand and its receptor can share a family stem (TNFSF10, TNFRSF10A) or a root with an R (IL6, IL6R). HGNC says it consults specialist groups for receptors, so conventions are set per family; there is no single rule.
-
Recharge rate (what a university core charges)The price a university core facility bills its users, set by rule to recover the core's costs and no more. partly written
-
RNA and protein half-lives in human cells: what has been measuredHow long human mRNAs and proteins last, how the numbers were obtained, and why no single median exists in the sources opened. partly written
In short. Human mRNA half-lives run from minutes for regulatory RNAs to many hours; one HeLa study gives a mean of 6.9 hours. Protein half-lives in non-dividing human primary cells run from about 10 to over 1,000 hours. Method, abundance cut-offs and cell division change the numbers, so none is a constant.
-
RNA capping and the poly(A) tailThe 5' cap is a methylated guanosine joined to the first nucleotide by an inverted 5'-5' triphosphate, and the poly(A) tail is a run of about 200 A added at the 3' end of mammalian mRNA. partly written
The cap, in three steps. The cap is an inverted guanosine on a 5'-5' triphosphate, methylated at N7. It is made co-transcriptionally by three reactions: triphosphatase, guanylyltransferase, methyltransferase.
-
RNA interferenceRNA interference is the cell machinery in which a short RNA guide strand finds a matching messenger RNA and causes it to be cut, which is the mechanism siRNA medicines borrow. partly written
RNAi in brief. A short double-stranded RNA is taken up by a protein complex called RISC. One strand is kept as a guide; it pairs with a matching messenger RNA and the complex cuts that message so the protein is not made.
-
Robin, FutureHouse's multi-agent discovery systemRobin proposes disease mechanisms and drugs from the literature, ranks them with an LLM judge, analyses lab data, and reported ripasudil for dry age-related macular degeneration. partly written
How it holds knowledge. What passes between steps is free-text reports and ranked lists. The public repository writes timestamped run folders; no persistent store was found. Ranking is a Bradley-Terry-Luce fit from an LLM judge's pairwise comparisons.
-
RTX-KG2 and the NCATS Translator graphsRTX-KG2 is a Biolink-standardised graph of about 39 million edges built from 70 sources; it is one of several graphs of the NCATS Biomedical Data Translator. partly written
Short answer. RTX-KG2 (Wood et al., 2022) canonical build KG2.7.3 has 6.4 million nodes and 39.3 million edges in 56 categories with 77 Biolink predicates, from 70 sources; about a third of edges come from SemMedDB text mining. Each edge carries its source and publications.
-
Sakana's AI Scientist-v2 and its experiment treeThe AI Scientist-v2 stores each experiment as a node in a tree and had one of three generated papers accepted at a workshop; an outside test of version 1 found frequent coding failures. partly written
How it holds knowledge. Each node holds a script, a plan, results and a vision-language-model review. Nodes flagged 'buggy' get debugging children, so a negative result is just a node that is not the best.
-
Salt bridgeA salt bridge is a close pair of opposite charges in a protein; unlike an ionic bond in a crystal, both ions would be hydrated in water, so its worth depends on burial and is contested. partly written
Is it just an ionic bond?. No. A crystal ionic bond has no water to compete. In a protein the charges were hydrated before the pair formed, and burying both costs desolvation energy. The net may be favourable or unfavourable, which is why the same pair type is worth different amounts in different places.
-
Scaffold hopping not yet written
To write this, these must be answered:
- What is scaffold hopping, and which published programme moved to a new core with its numbers? No source opened covers it.
-
Science sandbox: a task environment where an agent runs experiments against a hidden oracleA science sandbox is a packaged world in which an agent submits experiments, a hidden oracle returns reports, and the agent is judged on what it learned. partly written
What does 'sandbox' mean in the Rao and Lander paper?. The paper defines a science sandbox as a closed-loop setting for evaluating how an agent learns through experiment (Methods, 'Definition of a science sandbox'). Results, paragraph 1, calls it a controlled environment in which AI agents can conduct experiments to learn the rules governing a phenomenon. It is a task environment, not a confinement boundary.
-
Semi-conservative replicationIn semi-conservative replication each daughter DNA double helix keeps one chain of the parent and gets one newly made chain. partly written
-
Sequence Ontology (SO)The Sequence Ontology names genome features and variant consequences: not yet written here. not yet written
To write this, these must be answered:
- What does SO cover, how large is it, and how is it used in variant annotation?
-
Solute carrier (SLC) namesSLC plus a family number, a letter and a member number: the family is shared homology, not shared cargo, and the member number is a label. partly written
In short. SLC6A4 is family 6, subfamily A, member 4. Treat the family as a homology group and the member number as an identifier; HGNC itself warns that family members do not all transport the same things.
-
SplicingSplicing cuts introns out of pre-mRNA and joins the exons by two RNA-catalysed reactions, run by a machine of more than 100 proteins rebuilt on every intron. partly written
-
SPOKESPOKE is a UCSF Neo4j knowledge graph of about 27 million nodes and 53 million edges, rebuilt weekly and not available for bulk download. partly written
Short answer. SPOKE has 27,056,367 nodes of 21 types and 53,264,489 edges of 55 types from 41 databases, but about 92% of the nodes are proteins, so its size says little. It is rebuilt weekly on Neo4j and, by its authors' own statement, cannot be bulk-downloaded.
-
Spreadsheet gene-name errors and the symbol renamingExcel turns some gene symbols into dates; HGNC renamed the affected human symbols in 2019. partly written
In short. Symbols like SEPT2 can be read by a spreadsheet as dates and the original is lost. It was reported in 2004 and measured in 2016 and 2021; HGNC renamed several symbol families in 2019.
-
Subunit and isoform suffixesA trailing letter or number can mean a subunit, a duplicated gene or an arbitrary label; HGNC does not routinely give isoforms their own symbols. partly written
In short. A trailing letter such as the A in PRKACA marks a Greek-letter subunit written in Latin letters; CDK11A and CDK11B are separate genes. Splice isoforms of one gene normally share one symbol.
-
The 5' and 3' endsThe 5' and 3' ends are chemical: a strand has a free 5' phosphate at one end and a free 3' hydroxyl at the other, and new nucleotides are joined by the 3'-OH attacking the next triphosphate. partly written
What the ends are. The 5' end of a strand carries a phosphate on the 5' carbon of its last sugar; the 3' end carries a free hydroxyl on the 3' carbon. A strand therefore reads in one direction, and the two strands of a double helix run in opposite directions.
-
The assay-error ceilingThe noise in measured potency values caps how well any model, and any chemist, can rank close compounds. partly written
In short. Two labs measuring the same compound on the same target often disagree by far more than the difference between two analogues. Same-lab repeats are about three to four times tighter. A model that beats the replicate error of its own test labels is probably fitting noise.
-
The central dogma as Crick stated it, against how it is retoldCrick's dogma is a negative rule about sequence information leaving protein, not a one-way pipeline from DNA to RNA to protein. partly written
What the dogma says. Crick 1970 states the dogma as a purely negative claim, and the distinction matters to him: the positive statement that information does pass from nucleic acid to protein belongs to the sequence hypothesis, while the dogma says only that once information has passed into protein it cannot get out again (p. 562). The three transfers it says never occur are protein to protein, protein to DNA and protein to RNA.
-
The Gene Ontology (GO)The Gene Ontology is a shared vocabulary of terms for gene-product function, process and location, with annotations kept apart from the ontology. partly written
What it is. The Gene Ontology (GO) is a controlled vocabulary in three aspects: molecular function, biological process and cellular component. Its terms form a directed acyclic graph, so one term can have several parents. Annotations (a gene product, a term, an evidence code, a reference) are stored in separate files from the ontology.
-
The Google AI co-scientist and its Elo tournamentGoogle's AI co-scientist generates hypotheses, ranks them in pairwise LLM debates with an Elo rating, and keeps hypotheses and reviews in a text memory. partly written
How it holds knowledge. A persistent text context memory holds hypotheses, reviews, Elo scores, a research plan and feedback appended to prompts. The only graph is a hypothesis-to-hypothesis similarity graph used to schedule tournament matches. Source code is closed, so the real format is unknown.
-
The Human Phenotype Ontology (HPO)HPO is a vocabulary of abnormal human features, linked to diseases, that lets software compare a patient with disease profiles. partly written
What it is. HPO was launched in 2008 with over 8,000 terms and annotated all clinical entries in OMIM (Online Mendelian Inheritance in Man, a catalogue of genetic diseases). A specific term such as "seizure" implies all its ancestors in an is-a hierarchy, and diseases carry curated lists of terms.
-
The hydrogen bond (in proteins and water)A hydrogen bond is a donor N-H or O-H sharing its hydrogen with an acceptor O or N; in a protein the energy that matters is not the bond but the difference from the same groups bonding to water. partly written
The picture. In a protein every backbone residue has an N-H (donor) and a C=O (acceptor). Helices and sheets are the two regular ways to pair them. Each pairing is worth about 1 kcal/mol net, because the groups were already hydrogen-bonded to water before folding.
-
The hydrophobic effectThe hydrophobic effect is the free-energy gain from removing oily groups from water, a rearrangement of the water rather than a bond between the oily groups. partly written
Not a force. Nothing attracts two oily groups to each other. Water, which bonds to itself, loses less freedom when the oily groups are together than when each sits apart. The protein core is oily because burying those groups lowers the free energy of the system.
-
The ontology families of biology, and how widely each is usedThe families that followed the Gene Ontology cover phenotype, disease, chemistry, anatomy, cells, sequence, experiments, evidence and relations, with usage measured unevenly. partly written
Reading the numbers. Sizes differ in date and counting method, and "usage" is usually a maintainer statement unless a source measures it. Citation counts measure reach, not quality. Where no usage measure was found, the entry says so.
-
The Protein Conformation Ontology (PRC)PRC is an early OWL ontology that models protein conformations and conformational changes, tested on one case. partly written
What it is. The Protein Conformation Ontology models conformations as dispositions in the Basic Formal Ontology, at secondary, tertiary and quaternary levels, with conformational change as a process. It is described as a work in progress and was tested on the activation of the epidermal growth factor receptor (EGFR, HGNC:3236).
-
The Protein Ontology (PRO)PRO is an OWL ontology of protein families, gene-specific proteins, proteoforms and complexes; it has no conformation classes. partly written
What it is. PRO represents protein families, gene-specific proteins, proteoforms (a sequence plus its modifications) and complexes. In 2017, reasoning over it added more than 80,000 axioms with no inconsistencies (release 50.0). It held 332,990 terms in the Ontology Lookup Service on 15 November 2026.
-
The QIAGEN Knowledge Base (behind Ingenuity Pathway Analysis)QIAGEN's proprietary, manually curated knowledge base of literature findings with experimental context, sold inside IPA and as the QIAGEN Biomedical Knowledge Base. partly written
Short answer. Two products share one curated core: the QIAGEN Knowledge Base inside Ingenuity Pathway Analysis, and the QIAGEN Biomedical Knowledge Base sold as data in two tiers, KB-HD (human-curated) and KB-AI (AI-extracted). Every size and curation figure is QIAGEN's own claim, and no outside accuracy audit was found.
-
The Virtual BiotechA Stanford-led system of agents that turns ClinicalTrials.gov records into schema-validated JSON joined to targets and cell types, then tests whether cell-type-specific targets succeed more often. partly written
How it holds knowledge. Agents produce JSON validated against a Pydantic schema with a source tag per annotation, joined by NCT trial IDs, Open Targets IDs and Cell Ontology terms. Trials without a mapped target are dropped. Mapping accuracy is not reported in the part read.
-
Therapeutic window and therapeutic indexThe gap between the exposure of a drug that gives its effect and the exposure that causes harm, judged on free drug. partly written
In short. The window is arithmetic on exposure: the concentration that harms divided by the concentration that helps, both on free drug. Medicinal chemists raise the numerator by editing the molecule; clinicians and formulators move the denominator by dose and schedule. A margin is a screen with a known error rate, not a safety threshold.
-
Thermal energy (kT) as a yardstickOne unit of thermal energy, kT (or RT per mole), is the energy that random jostling at body temperature hands to every contact in a molecule; it is the ruler for deciding whether an interaction can hold. partly written
The yardstick. RT is about 0.6 kcal/mol at body temperature. Compare any interaction with that: about 1 RT gets undone constantly; several RT survive; and a whole protein is held by the small net of many such terms.
-
Three-dimensional genome foldingThe genome is folded in the nucleus into compartments, contact domains and loops, mapped by Hi-C, with cohesin and CTCF forming loops. partly written
Layers of folding. Compartments: spatial separation of open and closed chromatin. Domains and loops: segments with enhanced internal contact, with loops anchored at CTCF-bound sites. Hi-C maps all of this as contact frequency.
-
Transcription by RNA polymerase II: initiation, pausing, elongation, terminationHow human cells copy a gene into RNA, how fast and how accurately, and which of those numbers were seen and which were fitted. partly written
In short. RNA polymerase II (Pol II) is the enzyme that copies protein-coding genes into RNA. It starts at the promoter, usually halts for a while shortly after the start (promoter-proximal pausing), is released to run along the gene, and lets go past the end. Where Pol II sits is measured well. How fast it moves and how long it waits are fitted from snapshots and drug time-courses, and the fits disagree.
-
Translation initiation: scanning and start-codon choiceHow the small ribosomal subunit loads at the cap, scans and picks a start codon, how that choice is measured, and what it does to the protein. partly written
In short. A small ribosomal subunit loaded with the first transfer RNA binds the cap, slides 5' to 3', and starts at the first AUG in decent context. A weak context lets some subunits skip it (leaky scanning) and start further on, which can give a second, N-terminally shorter protein. This is real for named genes, and the second form is the minor one in every case that has any number.
His idea, in his words. "There's also a reverse truncation where, I just realized, you may get two start codons. Now you have two versions of the protein: one that's transcribing from the front to the end, and one that's transcribing from the middle to the end."
-
Uberon and the Cell OntologyUberon is a cross-species anatomy ontology built by merging four species anatomy ontologies; the Cell Ontology covers cell types. partly written
Uberon. Uberon was published in 2012 with over 6,500 classes and has 589 citing records in Europe PMC (2026-11-18, not audited). Four vertebrate anatomy ontologies were folded into it, for reasons the 2014 paper gives.
-
Variant mechanism labels: loss of function, gain of function and dominant negativePublic mechanism labels for variants are mostly gene-level, curated from text, and dominated by loss of function. partly written
What does the mechanism of a variant mean as a target?. Loss of function (LOF) means the protein does less. Gain of function (GOF) means it does something new or more. Dominant negative (DN) means the mutant interferes with the activity of a wild-type protein. The class is a label on a variant, but the public sets assign it to a gene.
-
Viral vector tropism: which cells a gene delivery virus can enterTropism is the set of cell types a vector enters and expresses its gene in; for AAV it is set mainly by the capsid and by the route of giving it. partly written
Tropism in brief. A viral vector is a virus shell with the disease genes removed and a therapeutic gene inside. Which cells it can enter depends on the protein shell (the capsid). Where the vector is put (into a vein or directly into the eye) decides which of those cells it meets first.
-
Vitamin C, collagen, and the lost GULO geneVitamin C is needed to keep collagen-building enzymes working, and humans, other anthropoid primates and guinea pigs must eat it because they lost the gene for its last synthetic enzyme. partly written
-
What a missing value means in each omics layerThe same blank is a measured zero, an unmeasured feature, a missing layer or an imputed value depending on the layer and tool, and the format rarely says which. partly written
Summary. In AnnData an outer join writes a never-measured feature as 0; in MuData a cell absent from a modality is a 0 in the map; MOFA+ treats a missing layer as unobserved in the likelihood; totalVI zero-fills a missing protein and ignores it in the loss; in sparse ATAC a zero mostly means not detected. The same-looking blank carries several meanings.
-
What biology used before the Gene OntologyBefore 1998 biology classified function with enzyme numbers, organism-specific catalogues and clinical or literature vocabularies that did not share terms across organisms. partly written
The systems in use. Enzyme numbers (1955 to 1961) gave one name per enzyme. MeSH (1954) indexed the literature. UMLS (1986) integrated clinical vocabularies by mapping. The first genome-scale function catalogue was for E. coli (1993). Each solved a local problem; none described gene function across organisms.
-
What controlled vocabularies exist for protein structure, and what is missingMost structural concepts have a database, a file dictionary or a flat vocabulary; one early logical ontology covers conformation, and pocket, interface and an MD metadata standard remain empty as ontology terms. partly written
The verdict after a proper search. Report 34 inferred that state, residue-pocket-function relations, interfaces and ensembles were unrepresented. After searching the dictionaries and resources, that is wrong as stated: each has something, nearly all of it a database, a file dictionary or a flat vocabulary. Still empty: pocket and interface as typed entities in any ontology searched, and any metadata standard for molecular dynamics.
-
What the data object is, layer by layerThe file formats and tables that each omics layer lands in, who defines each, and which are standards and which are one group's choice that became practice. partly written
Standard, practice or one group's choice. Most data objects in daily use began as one project's design and became practice: FASTQ, SAM and BAM, VCF and AnnData. A smaller set is written by a standards body or consortium: mzML and mzTab-M for mass spectrometry, SBML for models, OME-NGFF for images. No source opened measures how often any of them is followed.
-
Where the spatial tie is weakest: segmentation, registration and resolutionImaging platforms leave 20% or more of transcripts unassigned by default segmentation, foreign-cell signal is common, and no micrometre registration error was found for cross-platform alignment. partly written
Summary. Default segmentation leaves 20% or more of transcripts unassigned on each platform in the method authors' own benchmark. One lung preprint found 47.0 to 73.4% of some cell types with more than 5% foreign-marker signal. The only measured registration error found, 4.5 micrometres, is for mass-spectrometry imaging; none was found for cross-platform alignment. These are authors benchmarking their own tools.
-
Where the textbook arrow fails or is incompleteReverse transcription, RNA editing, RNAs that are the product, and quality-control pathways, and how little of each was measured in the sources opened. partly written
In short. Crick's dogma forbids only transfer out of protein, so RNA products, RNA editing and reverse transcription do not break it by his own account. They do break the simple picture of a protein-coding pipeline. Each non-standard transfer has to beat enzyme and sequencing artefacts before it counts.
-
Whole-cell modelA program that represents the genes of one cell and the reactions of their products and steps them forward through a cell cycle. partly written
-
Wikidata's biomedical graphWikidata holds a community-edited knowledge graph of genes, proteins, drugs and diseases, with references and qualifiers on statements. partly written
Short answer. Waagmeester, Stupp and colleagues (2020) describe over 750 million statements on 61 million items (2019), including over 1.1 million genes and about 940,000 proteins. Statements carry references and qualifiers, and conflicting claims coexist.