Today's digest has 27 articles: 8 medchem, 8 pharmacology, 9 computational and 2 industry. Highlights include a first-in-class METTL1 inhibitor in Nature, audits showing leakage can inflate AI ADMET and peptide-prediction claims, and a partial clinical hold on Kite's KITE-753 CAR-T trials.
2026-10-31 · ⚗️TAK1 (TGFβ-activated kinase 1); covalent small-molecule kinase inhibitor · Journal of Medicinal Chemistry
- An amino-substituted furo[2,3-c]pyridine scaffold was equipped with an enal warhead, yielding an irreversible, covalent TAK1 inhibitor (compound 26).
- In cells it blocked TAK1-driven signaling, and in mice it maintained target inhibition over time.
- It curbed growth of lymphoma xenografts, though off-target kinase contributions are possible; selectivity and pharmacokinetics still need optimization.
💡 PointsBased on the abstractAn enal instead of the usual acrylamide is a deliberate choice: a more reactive but tunable electrophile aimed at a cysteine in TAK1's ATP pocket, the same logic as 5Z-7-oxozeaenol. A furopyridine hinge binder is a fresh alternative to that resorcylic lactone natural product. Sustained in vivo engagement is encouraging; until a kinome scan rules out off-target hits, I wouldn't credit the xenograft effect to TAK1.
Discovery of 7-Aminofuro[2,3- c ]pyridine-Based Irreversible Inhibitors of TGFβ-Activated Kinase 1 (TAK1) with In Vivo Antitumor Activity ↗
2026-10-31 · ⚗️MDM2 / stabilized helical peptides · Organic Letters
- Short peptides seldom keep a stable helix, hampering intracellular protein–protein interaction targeting; the authors cap the N-terminus with 4-mercaptoproline.
- A covalent bridge to an internal cysteine preorganizes the backbone and encourages helix propagation.
- Adjusting residue stereochemistry and linker composition improved helicity, uptake, protease stability and MDM2 binding, with reduced cancer cell viability and migration.
💡 PointsBased on the abstractAll-hydrocarbon staples tend to add greasy bulk that hurts solubility; here the N-terminal cap does double duty as helix nucleator and cross-link anchor, which is neater. For an MDM2 p53-mimetic I'd still want the thiol bridge's stability in cytosol and resistance to reduction checked, but it's a cheaper handle than another staple screen.
N-Terminal 4-Mercaptoproline Capping for Peptide α-Helix Stabilization and Intracellular Protein–Protein Interaction Modulation ↗
2026-10-31 · ⚗️Self-assembling therapeutic peptide (CQ7) for pulmonary fibrosis · Journal of Medicinal Chemistry
- A therapeutic peptide was fused with a self-assembly motif to give CQ7, which forms supramolecular nanofibers under physiological conditions, both inside and outside cells.
- The fibrous form improves lung retention and delivery, and showed anti-fibrotic and anti-inflammatory activity in a mouse fibrosis model dosed by injection or inhalation.
- Transcriptomics suggested the mechanism involves mitochondrial function, lower oxidative stress and modulated apoptosis.
💡 PointsBased on the abstractIs this a better peptide or just a better formulation? The fibre doesn't change the pharmacophore; it changes exposure, and in the lung that's the real liability for a peptide. Appending an assembly motif is a cheap fix for clearance, and it's tractable. But I'd want the monomer-versus-fibre comparison at matched dose before crediting the mitochondrial story, and a look at how reproducible the assembly is batch to batch.
Enhanced Pulmonary Fibrosis Treatment by Self-Assembling Peptide Nanofibers via Improved Pulmonary Delivery and Retention ↗
2026-10-28 · ⚗️METTL1 tRNA m7G methyltransferase; selective catalytic inhibitor · Nature
- Developed STM9005, a selective inhibitor of the tRNA methyltransferase METTL1, testing whether a tRNA-modifying enzyme can be drugged in cancer.
- In diverse cancer cell models, it slowed growth, shifted differentiation and disturbed the cell cycle, while lowering m7G on selected tRNAs and their levels, altering translation of oncogenic mRNAs.
- In several mouse models from different tissues, catalytic inhibition curbed tumours and extended survival.
💡 PointsBased on the abstractMETTL1 is a SAM-dependent RNA methyltransferase, so my first worry is selectivity against the wider methyltransferase family; the paper's "selective catalytic inhibitor" claim is what I'd test first. A catalytic-dead phenocopy would matter more to me than efficacy across tissues, since it separates on-target from off-target toxicity.
Pharmacological inhibition of METTL1 as a therapy against solid and blood cancers ↗
2026-10-26 · ⚗️Thrombin (serine protease) / synthetic peptides · Angewandte Chemie International Edition
- Two sets of short thiol-bearing peptides are joined combinatorially through disulfide bonds, so m fragments and n fragments give m × n distinct peptide sequences.
- Acoustic liquid transfer in 1536-well plates formed several pairs per well, enabling functional screening of over 800,000 compounds against thrombin.
- The screen delivered potent thrombin inhibitors; at under 1 kDa, such pairs may suit oral or intracellular targeting.
💡 PointsBased on the abstractWould I trust a disulfide hit? Only after resynthesis. Plasma thiols and glutathione will scramble that bond, so the real deliverable is the pair's pharmacophore: I'd lock it as a thioether or single-chain linker and see if potency holds. Thrombin suits this well, since the S1 and S2/S3 pockets give an obvious two-fragment split to compare against Phe-Pro-Arg-type chemotypes.
Chemical Synthesis and Functional Screening of an 800 000‐Member Library Using a Peptide‐Pairing Strategy ↗
2026-10-26 · ⚗️Akt2/Akt3 kinases; covalent-allosteric small-molecule inhibitors · Angewandte Chemie International Edition
- Structural and modeling work exposed isoform differences in Akt allosteric pockets, enabling covalent inhibitors that favor Akt2 or Akt3 over Akt1.
- Holding the kinase inactive sidesteps ATP-site selectivity limits; pyrazole compounds favored Akt2, pyridine ones mostly Akt3.
- Colorectal cancer organoids and co-crystal structures supported the selectivity rationale.
💡 PointsBased on the abstractThe pocket differences between Akt isoforms are tiny, so I'd want to know how much of the Akt2/Akt3 discrimination comes from residue contacts and how much from warhead reactivity; ATP-site compounds never managed it. The pyrazole-for-Akt2, pyridine-for-Akt3 split is a usable SAR handle. Proteome-wide reactivity profiling comes before I believe any of it.
Isoform‐Selective Targeting of Akt Through Covalent Allosteric Inhibition ↗
2026-10-26 · ⚗️Natural products and drug molecules (small-molecule synthesis methods) · Angewandte Chemie International Edition
- This review surveys methods for directly modifying complex molecules, mainly natural products and drug molecules, at a late stage.
- Strategies are grouped into skeletal editing and peripheral editing, each illustrated with recent representative examples.
- Biomimetic, nature-inspired editing gets a brief discussion; the authors aim to guide site-selective diversification of complex scaffolds.
💡 PointsBased on the abstractThe real pitch here: once a scaffold is precious and gram-limited, editing the core or periphery of the material in hand beats a 15-step re-synthesis of each analog. I'd still check site selectivity and functional-group tolerance on my actual lead before counting on any of it.
Direct Molecular Editing of Complex Molecules for Drug Discovery ↗
2026-10-26 · ⚗️BPTF bromodomain (small molecule inhibitor) · Angewandte Chemie International Edition
- Comparative structural analysis and modeling yielded BZ2, a regioisomer of an earlier BPTF bromodomain inhibitor with off-target activity.
- Crystal structures and calculations attribute improved selectivity against other bromodomains, including BRD7 and BRD9, to differing water networks and polar contacts.
- Both genetic bromodomain disruption and BZ2 treatment curbed neuroblastoma cell growth, positioning it as a probe and lead.
💡 PointsBased on the abstractA regioisomer swap buying selectivity against BRD7 and BRD9 is the kind of cheap, tractable win I'd happily take: no new scaffold, just moving a substituent so it exploits a water network the off-targets can't match. That's a transferable trick for any bromodomain series. I'd still want the BZ2 selectivity panel beyond those two before calling it a probe.
Differential Water Networks Guide Selectivity Optimization of a Cell Active BPTF Inhibitor in Neuroblastoma ↗
2026-10-30 · 🧬Cytochrome P450 2E1 (CYP2E1) probe substrate chlorzoxazone · Drug Metabolism and Disposition
- A rebuttal to a 2026 Supersomes study that attributed chlorzoxazone 6-hydroxylation mainly to hepatic CYP1A1 rather than CYP2E1.
- The authors revisit older literature they say the critics overlooked, which already dealt with these selectivity objections.
- Their conclusion is that the reaction remains a dependable CYP2E1 activity readout for assessing drug interaction risk.
💡 PointsBased on the abstractRebutting a Supersomes result with older literature is the right instinct: recombinant systems tell you what an enzyme can do, not what it does in liver. Abundance and kinetics scaling decide that. Still, a probe is only as good as its last inhibitor-based validation, so I'd want 6-hydroxylation confirmed with CYP2E1 inhibition in microsomes before trusting it in any new DDI design.
Review of the selectivity of chlorzoxazone 6-hydroxylation as a probe for human cytochrome P450 2E1 ↗
2026-10-29 · 🧬TMEM63B (mechanosensitive ion channel / lipid scramblase); allosteric modulators and agonistic antibody YN9303-24 · Biochemical Pharmacology
- TMEM63B, a TCS-family protein, works as both a force- and osmolarity-gated ion channel and a lipid scramblase.
- Structures capture separate channel and scramblase states, and patient mutations connect it to epilepsy, hemolytic anemia, and placental, lung and gut defects.
- Drug discovery routes include allosteric gating modulators, isoform- or tissue-selective approaches to limit on-target toxicity, and agonist antibody YN9303-24 for loss-of-function disease.
💡 PointsBased on the abstractDual-function targets are an assay headache: before trusting any SAR, I'd want to know whether a hit shifts gating, scrambling, or both, so pressure-clamp patch recording plus annexin V scrambling readouts are the minimum. The patient mutations give genuine human genetics, but with expression this broad, an agonist antibody like YN9303-24 only makes sense if tissue selectivity holds up.
TMEM63B as a dual-function mechanosensitive channel and lipid scramblase: structural insights, physiological roles, and pharmacological potential ↗
2026-10-28 · 🧬cGAS and STING small-molecule inhibitors (RU.521, STING-IN-2) · Biochemical Pharmacology
- Single-cell data from mouse and human aneurysm tissue place cGAS-STING signaling mainly in endothelial cells, with higher expression in ruptured than stable human lesions.
- In angiotensin II-driven mouse aneurysm, the pathway was switched on; either a cGAS inhibitor or a STING antagonist lowered aneurysm occurrence and tissue damage and improved endothelial mitochondria.
- Patient aneurysm samples also showed elevated pathway proteins, and DHFR loss emerged as a possible link between DNA damage and activation.
💡 PointsBased on the abstractTwo inhibitors, one mouse model: that's thin. Angiotensin II aneurysm is acute and hypertension-driven, a poor mimic of slow human AAA, and RU.521 is far more potent on mouse than human cGAS. Human evidence is correlative expression, not genetics. Next study needs dosing after aneurysm is established, exposure data, and a second model such as elastase.
cGAS-STING activation promotes abdominal aortic aneurysm formation in mice and humans: Therapeutic potential of inhibition ↗
2026-10-28 · 🧬SRC kinase; β-carboline alkaloid derivative (natural-product-derived small molecule) · Biochemical Pharmacology
- Starting from a plant alkaloid with known antitumor activity, the authors built a scalable total synthesis and made a series of nineteen analogues; D14 was the most cytotoxic across glioma lines, including patient-derived primary cells.
- Computational screening, dynamics and thermal-shift assays pointed to SRC as the target, with binding near Gln147 and reduced SRC phosphorylation, dampening PI3K/AKT and STAT3 signaling.
- In an orthotopic mouse glioma model, D14 slowed tumor growth without evident organ toxicity.
💡 PointsBased on the abstractNineteen analogues, one winner, and a target call resting mostly on docking, dynamics and a thermal shift: I wouldn't call SRC the mechanism yet. Without kinome selectivity and a Gln147 mutant, this could be polypharmacology. And orthotopic GL261 is forgiving; show brain exposure across the BBB and a real margin over the cytotoxic dose.
D14, a β -carboline alkaloid derivative of dehydrocrenatidine, exerts anti-glioma effect by targeting SRC ↗
2026-10-28 · 🧬Dengue virus; attenuated-virus human challenge; anti-NS1 antibodies · Science Translational Medicine
- Forty-five healthy adults with differing dengue histories received a live attenuated virus of a different serotype in a controlled phase 1 challenge.
- Prior exposure boosted viral load and immune responses, yet baseline antibodies that enhanced infection in vitro predicted the weakest control and fading T and B cell memory.
- Antibodies against NS1 across serotypes, or strong neutralizing and T cell immunity at baseline, tracked with early interferon signaling and little replication, pointing vaccine design toward NS1.
💡 PointsBased on the abstractAny dengue vaccine program now has to treat baseline serostatus as a covariate, not a footnote: in vitro enhancing antibodies predicting poor control in humans is the readout we've lacked. NS1 antibodies tracking with suppressed replication is attractive, but with 45 volunteers and one attenuated strain, I'd want it validated as a true correlate before designing around it.
Dengue virus controlled human challenge reveals that preexisting immunity shapes viral replication and broadly protective responses ↗
2026-10-28 · 🧬NOX1 (with RAC1) small-molecule modulator; gramine · Biochemical Pharmacology
- Gramine was nominated as a NOX1-binding compound; docking and a thermal shift assay pointed to direct engagement, and it hastened NOX1 protein turnover.
- In ethanol-stressed hepatocytes, knocking down NOX1 or RAC1 reproduced much of the benefit, easing oxidative, mitochondrial and ferroptosis-type damage.
- Gramine improved cell health even when given after ethanol, but forcing NOX1 expression weakened its effect.
💡 PointsBased on the abstractKnockdown phenocopying the drug is reassuring, but it is cell-only; I'd want Nox1-null or hepatocyte-specific mice under chronic-plus-binge ethanol before going further. Also, docking plus a thermal shift on a ~160 Da indole is thin proof of binding, and with no free-exposure or selectivity data, a plant alkaloid with known toxicity could easily be working off-target.
Gramine limits alcohol-associated liver injury by suppressing NOX1-linked ferroptotic responses and mitochondrial dysfunction ↗
2026-10-28 · 🧬MCT1 (Slc16a1) mitochondrial lactate transporter · Nature
- Mouse hearts still oxidize glucose carbon without the pyruvate carrier, via a parallel lactate route through MCT1.
- MCT1 sits in the inner mitochondrial membrane of mouse and human cardiomyocytes, pairing with matrix LDHB to convert lactate to pyruvate and NADH for respiration.
- Removing MCT1 impairs metabolic flexibility and worsens heart failure and survival under neurohormonal or ischemia-reperfusion stress.
💡 PointsBased on the abstractTreat this as a cardiac safety flag, not a target. The phenotype only surfaces under stress, so a clean baseline echo would miss it, and MCT1 inhibitors already in oncology block plasma-membrane and mitochondrial lactate flux alike. I'd want adult, cardiomyocyte-restricted deletion, human loss-of-function data and cardiac exposure margins before trusting any of it.
MCT1-dependent mitochondrial lactate import sustains cardiac metabolism ↗
2026-10-28 · 🧬Withaferin A (natural product) acting on the ROS/mitochondrial apoptosis and Keap1/Nrf2/HO-1/GPX4 ferroptosis pathways in gastric cancer · Biochemical Pharmacology
- A withanolide from Withania somnifera was tested in gastric cancer cells, where it curbed growth and migration.
- It triggered ROS-driven mitochondrial injury and shifted apoptosis regulators toward cell death, while also promoting ubiquitin-mediated loss of GPX4 and engaging Keap1/Nrf2/HO-1 signalling to cause lipid peroxidation and iron buildup.
- In a mouse xenograft the compound slowed tumours without obvious systemic toxicity.
💡 PointsBased on the abstractWithaferin A is a notoriously promiscuous electrophile, so "hits Keap1/Nrf2 and GPX4" mostly reads as covalent reactivity, not a validated target. ROS-driven killing in cell lines and a xenograft with "no obvious toxicity" say nothing about free exposure at the efficacious dose or the safety margin. Show tumour drug levels and a thiol-reactivity control before any follow-up.
Withaferin A induces apoptosis and ferroptosis in gastric cancer cells via ROS-mediated mitochondrial damage and the Keap1/Nrf2/HO-1/GPX4 axis ↗
2026-10-30 · 💻Structure prediction / binding-site analysis and docking · bioRxiv
- Each predicted AlphaFold pocket across a large set of proteins was positioned between the unbound (apo) and ligand-bound (holo) states seen in experimental structures of the same site.
- Most pockets resembled the holo state, mirroring the balance of experimental structures in the PDB.
- In a retrospective analysis, binding-site predictors and cognate-ligand docking performed better on holo-like pockets, especially where apo and holo states differ strongly. The confidence score reflects pocket variability, not state.
- Per-pocket classifications are released to guide docking readiness.
💡 PointsBased on the abstractAlphaFold-based docking benchmarks may partly reward PDB bias rather than prediction skill, since most predicted pockets sit on the holo side. That makes the released per-pocket labels a handy triage step before docking into a model. The analysis is retrospective and cognate-ligand only, so behaviour with novel chemotypes remains an open question.
A tale of two states: AlphaFold binding pockets on the apo-holo continuum ↗
2026-10-30 · 💻Generative design / conditional diffusion · Journal of Cheminformatics (Springer Nature)
- GCMDiff generates molecules as graphs, with atom-pair and bond-centred encoding, rather than SMILES strings, so topology is represented explicitly.
- A diffusion process is steered by interpretable conditions: atom count, Rule-of-Five properties and substructure features from Checkmol, PubChem and Rings in Drugs descriptors.
- Benchmark experiments are reported for the generated molecules; the feed snippet is truncated, so validation beyond benchmarks and comparisons with other generators are not stated.
💡 PointsBased on the abstractConditioning on Rule-of-Five properties is the least interesting knob; my projects fail on potency and synthesis, not logP. Substructure conditioning matters only if it reliably enforces fragment or pharmacophore constraints. With no prospective synthesis or assay data and unclear baselines against existing graph diffusion models, I can't tell it beats a SMILES generator plus a filter. Released code decides whether I try it.
GCMDiff: a condition-guided diffusion framework for pharmacophore-fragment drug discovery and target-specific molecular design ↗
2026-10-30 · 💻ADMET property prediction / model benchmarking · bioRxiv
- Graph neural networks, frozen pretrained encoders and fingerprint baselines were compared on six Therapeutics Data Commons ADMET tasks, using scaffold splits, an identical tuning budget per family, validation-only selection and multi-seed medians.
- Differences between families mostly fell within seed variability, so no approach dominated. Fingerprints were competitive on several endpoints, and the GNN's lone clear edge traced to added descriptors, not the graph itself.
- Selecting optimizers on test data or reporting single best runs inflated apparent gains, supporting endpoint-specific model choice and leakage-controlled evaluation.
💡 PointsBased on the abstractOnce descriptors are added, the GNN's one clear win looks like featurization rather than graph structure, which makes me doubt many reported GNN gains. Six public TDC tasks and nothing prospective say little about my own series. The real takeaway is the checklist: matched tuning budgets, validation-only selection, multi-seed medians. Handy for vetting vendor claims.
Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction ↗
2026-10-30 · 💻Peptide property prediction / leakage-aware benchmarking · Journal of Cheminformatics (Springer Nature)
- The authors present a reproducible framework that audits tumor-homing peptide predictors for leakage, motif-driven scoring, chemical redundancy and sensitivity to hard negatives, rather than judging them on score alone.
- It combines motif-rule comparisons, Morgan/ECFP chemical-space analysis, strict external-recognition filtering and a conservative scheme for ranking peptide analogs.
- Evaluation is retrospective on a curated public benchmark of a few thousand peptides, using a simple k-mer logistic-regression baseline under standard stratified cross-validation; the abstract is truncated, so stricter-split results are not stated.
💡 PointsBased on the abstractThe audit is worth more than the predictor it audits. A k-mer logistic regression under stratified CV says little about novel scaffolds, so headline scores likely overstate real performance, though without stricter-split numbers or wet-lab follow-up I can't say by how much. I'd borrow the checklist for any peptide model I inherit.
THP-NanoTarget: leakage-aware cheminformatics for tumor-homing peptide evaluation and conservative analog prioritization ↗
2026-10-30 · 💻LLM-orchestrated de novo protein binder design · bioRxiv
- LPT chains fixed LLM stages, each running a bounded tool-using loop, so binder design starts from one plain-language request while keeping token use and compute low.
- The stages can query a curated database of about 14,000 digested papers; a model with access answered most held-out questions drawn from it, whereas one without access answered none.
- Three end-to-end campaigns targeted CALCRL/RAMP1, PD-1/PD-L1 and PVR-TIGIT. Support is in silico only, from agreement across several folding models, with no experimental testing reported.
💡 PointsBased on the abstractThe novelty in LPT is a bounded orchestration layer, not a new binder-design method, and its literature test only shows retrieval beats no retrieval, a low bar. Folding-model agreement is weak evidence without wet-lab data, since those models share training biases. On my project, hit rates against plain RFdiffusion pipelines would settle it.
LPT: Orchestrating bounded agentic loops for automated protein binder design ↗
2026-10-30 · 💻Transcriptomics-based virtual screening / spatial transcriptomics · Journal of Chemical Information and Modeling (ACS)
- Bulk RNA signatures average away intratumoral differences, so the authors score compounds against localized spatial expression signatures instead, focusing on heterogeneous triple-negative breast cancer.
- They compared bulk and spatial signatures on a set of about 1,300 diverse compounds, evaluating how well each ranked known anticancer drugs.
- Spatial signatures ranked known anticancer drugs better than bulk ones and also flagged unclassified and non-oncology compounds as candidates; evidence is computational, with no experimental validation described.
💡 PointsBased on the abstractRanking known anticancer drugs higher is a retrospective enrichment test in one cancer type, and it doesn't show the spatial signature will find new ones. There's no wet-lab follow-up, and nothing on leakage or baselines. The non-oncology hits are the interesting part; I'll reserve judgment until a few are actually tested.
Spatially Resolved Transcriptomic Correlation Improves the Prediction of Anticancer Drug Candidates ↗
2026-10-29 · 💻Drug-target interaction prediction / linear embedding alignment · Molecular Informatics (Wiley)
- A linear model projects ChemBERTa molecule embeddings and ESM-2 protein embeddings into one shared space, where predicted affinity is simply the cosine similarity between drug and target.
- It was pretrained on a large curated set of interactions and evaluated on the KIBA and Davis kinase benchmarks, where the authors report it surpassing existing methods.
- The learned space reportedly recovers kinome phylogeny without supervision and retrieves known targets of approved drugs, though validation appears retrospective with no new experiments.
💡 PointsBased on the abstractTwo frozen embeddings and a cosine is a refreshingly simple baseline, which is exactly why the split matters. Random splits let near-identical kinase inhibitors straddle train and test, and the pretraining corpus may overlap the benchmark pairs. Before believing the margin over deep models, I'd score it on my own scaffolds.
Interpretable Drug–Target Prediction via Geometric Alignment: A Linear Alternative to Deep Architectural Complexity ↗
2026-10-28 · 💻Integrated pharmacophore / QSAR / docking virtual screening · Molecular Informatics (Wiley)
- Structure-based and ligand-based pharmacophore models were built from the H3R crystal structure and known inhibitors, then checked retrospectively against decoys alongside a QSAR model.
- These models screened a commercial CNS-focused, BBB-oriented library; docking, consensus scoring and QSAR filtering narrowed it to a short list of candidates.
- Radioligand displacement testing confirmed one submicromolar binder and several low micromolar ones, showing the combined workflow yields real, experimentally confirmed hits.
💡 PointsBased on the abstractNothing methodologically new here; it's a standard pharmacophore-QSAR-docking funnel, so the interesting question is the hit rate against a plain docking-only baseline, which isn't shown. Decoy-based retrospective checks are easy to flatter. Still, wet-lab confirmation from a purchasable, BBB-aware library is useful, though I'd want novelty versus known H3 chemotypes first.
Novel Histamine H3 Receptor Ligands by Integrated Ligand‐ and Structure‐Based Computational Modeling ↗
2026-10-25 · 💻Proteochemometric bioactivity prediction / uncertainty quantification · Journal of Cheminformatics (Springer Nature)
- The authors extend proteochemometric models with assay descriptors, encoded either as categorical identifiers or as text embeddings of assay descriptions, so that experimental context is modelled alongside compound and protein features.
- They benchmarked on a curated ChEMBL-derived GPCR set under random, scaffold and assay-held-out splits, judging both point accuracy and probabilistic quality, retrospectively.
- Split choice drove difficulty most. Categorical identifiers helped consistently when assays recurred; on held-out assays the representations converged, with text embeddings holding a small, steady edge.
💡 PointsBased on the abstractCategorical assay IDs paying off when assays recur is close to leakage by design, so I discount that half of the result. What matters for a new campaign is the held-out-assay split, where the text-embedding edge is small and hardly justifies retooling. Whether any of it survives a prospective campaign is untested.
PCAM: proteochemometric assay-aware modeling for robust bioactivity estimation and uncertainty quantification ↗
