Osteoarthritis has no approved drug that slows joint destruction. Patients get painkillers and steroid injections until the joint fails and is replaced. Public data could in principle point to existing drugs that push diseased cartilage back toward health. One source is gene-expression atlases of diseased joints. Another is LINCS L1000, a library of over a million profiles recording how cultured cells change their gene expression after each of thousands of compounds. Recent osteoarthritis studies take this route. Each mines one or a few small public datasets, nominates a target or compound, and then validates a single hit. None of them can say how often such computational nominations are right.
We propose to measure that error rate directly and then reduce it. We will profile how a few hundred compounds change gene expression in stressed primary human chondrocytes, the cells that build cartilage. That gives the first ground truth for whether a computational ranking predicts real reversal of the disease program in joint cells. We will then test a specific claim. Two well-documented artifacts of fragmented data generate most false nominations: disease signatures drawn from single small cohorts, and a known statistical bias in L1000's averaged profiles. Correcting both should raise the rate of true reversal among top-ranked compounds.
Core Claim
Computational repurposing for osteoarthritis fails for reasons we can name and measure, not just because "biology is hard." The first source of error is the disease signature. It is the list of genes up or down in diseased versus healthy joint tissue that a drug should invert. Current studies build it from one cohort of a few dozen samples, so study-specific batch effects and end-stage damage dominate it. The second source is the compound profile. L1000 averages replicate experiments into one consensus profile per compound. Smith et al. showed that this averaging alone makes unrelated compounds correlate. The null correlation reaches 0.4925 for compounds with 15 or more replicates, against 0.2919 for those with fewer. Heavily replicated compounds are 1.19 to 2.2 times more likely to land in the top 1% of any query.
We predict three things. Uncorrected matching will over-nominate heavily replicated compounds that do nothing in chondrocytes. A signature built across studies and tissues, and stable when any single study is left out, will predict measured reversal better than any single-cohort signature. Combining both corrections will give a measurable gain in hit rate under evaluation on donors and chemical scaffolds the method has not seen. The new contribution is the chondrocyte-native test that makes these claims falsifiable. The corrections themselves are mostly borrowed.
Relation to Prior Work
Three 2025 studies motivate the proposal by showing the current practice and its blind spots.
- Ma et al. combine blood genetic data with knee-OA genome-wide association data through Mendelian randomization, a method that uses inherited variants as natural experiments to separate cause from consequence. They intersect the result with synovial cohorts of 43 and 40 samples and nominate MEG3 and MAPK3. They stop at targets and do no compound screening.
- Yao et al. combine bulk and single-cell synovial data to implicate PYCARD in inflammatory macrophages. A database lookup in DGIdb returns 57 candidate compounds, of which only butein was tested.
- Mei et al. dock over 18,000 ZINC15 compounds against a single target, DDR2, using one single-cell dataset. Their best hit binds with a CETSA (cellular thermal shift assay) IC50 of 45.73 μM in chondrocytes.
None of the three reports how many nominated compounds fail. We do not try to find a better single target. We build the yardstick these studies lack.
Two papers constrain the design.
- Çelebi et al. and Lavelle et al. show that random train/test splits flatter drug models. In Lavelle et al., drug-blind R² in cancer response prediction was 0.03 on the original benchmark and rose only to 0.22 after harmonizing 5.4 million experiments.
- Jarada et al. note that repurposing models for degenerative disease are limited by the lack of cell-type-specific perturbation data.
Together these force two choices: evaluation that holds out donors and chemical scaffolds, and an honest test of whether L1000's mostly cancer-derived cell lines transfer to chondrocytes at all.
Smith et al. supply the replicate-bias correction. Huffman et al. supply a model for recording study, sample and assay metadata, which we reuse to keep provenance across studies. We do not propose a general osteoarthritis data model. That would be infrastructure without a falsifiable claim.
Research Program
Building the disease signature. We will harmonize every osteoarthritis expression dataset named in the file:
- the synovial bulk cohorts used by Ma et al. (GSE12021, GSE1919, GSE206848, GSE55235, GSE55457);
- the synovial bulk RNA-seq cohort GSE89408;
- synovial single-cell data GSE152805;
- cartilage single-cell data GSE200843.
Each sample will be recorded with its study, tissue, cell type, disease stage and platform, following the Study-Experiment-Assay layout of Huffman et al. We will then build two signatures, one for chondrocytes and one for synovial macrophages. Each is estimated with study as a random effect. Genes are kept only if their direction holds in every leave-one-study-out refit. A further variant up-weights genes with Mendelian-randomization support, reusing the instruments of Ma et al. Those instruments come from blood, not joint tissue, so this variant is treated as an ablation, not as part of the core method.
Correcting the compound profiles. Each L1000 compound profile is scored against the signature. Each score is compared to a null distribution built from unrelated compounds with the same replicate count, following Smith et al. A compound's score is also discounted by how inconsistent its effect is across L1000 cell lines. The idea is that a response shared across many unrelated cell types is more likely to carry over to chondrocytes.
Building the ground truth. We will profile about 200 L1000 compounds by 3′ RNA sequencing in primary human chondrocytes from at least six donors, stimulated with IL-1β. The compounds are chosen by stratified sampling across three factors:
- replicate-count bins in L1000;
- chemical scaffolds;
- the full range of predicted reversal scores, including compounds every method ranks low.
The panel includes compounds with prior chondroprotective reports (naringenin from Wang et al., butein) as positive anchors. For each compound and donor, the measured reversal is the correlation between the compound's induced change and the inverted chondrocyte signature. A secondary readout uses the canonical markers MMP13, ADAMTS5, COL2A1 and SOX9.
A smaller arm of about 40 compounds runs in synovial-macrophage-like cells. It tests whether the macrophage signature adds information.
Baselines
Naive reference. The do-nothing reference ranks compounds at random, and alternatively by their number of L1000 replicates. The replicate-count ranking tests whether uncorrected methods are merely recovering well-studied compounds.
Current osteoarthritis practice. The first literature baseline reproduces the workflow of Yao et al. and Ma et al. It takes top genes from differential expression in a single cohort (GSE89408, the largest synovial set) and ranks compounds by DGIdb gene–drug interactions with those genes. Yao et al. report 57 compounds from this lookup, with one tested positive. That is not a hit rate, so there is no reported result found on Lacuna for its precision.
Strongest competitor. The second, and strongest, baseline is standard connectivity scoring of the single-cohort signature against L1000 consensus profiles, with the Smith et al. replicate-matched null applied. It is the strongest competitor because it already contains the best-documented correction. Our remaining contribution must show up as gains over it. The only prior evidence for its correction is Smith et al.'s null-correlation figures (0.4925 versus 0.2919) and their 1.19–2.2× top-1% enrichment. There is no reported result found on Lacuna for its accuracy against measured chondrocyte responses.
Structure-based screening. A docking-style single-target screen against DDR2, following Mei et al., will be run on the same 200 compounds. It represents the single-target view.
Fair comparison. All methods see the same L1000 release and the same public datasets. Any method with tunable settings gets the same budget, about fifty configurations, chosen on held-out donors. Final numbers are reported on two other donors and on scaffolds unseen during tuning.
Ablations. Three ablations isolate the key ingredients:
- the cross-study stable signature with uncorrected L1000 scoring;
- the single-cohort signature with full correction;
- the full method with and without the Mendelian-randomization weighting.
What counts as winning. Our method beats the strongest baseline if two conditions hold. Its precision among the top 20 compounds, defined as measured reversal above the 90th percentile of the panel, must be higher with a bootstrap confidence interval over compounds and donors that excludes zero. And its Spearman correlation with measured reversal must be higher on held-out donors and scaffolds.
Why not just…?
Why not just use differential expression in one good cohort plus a database lookup? That is the cheapest workflow and exactly what current papers do. It is a baseline here. Ma et al.'s own Mendelian-randomization filter suggests that most differentially expressed genes are consequences of damage rather than causes. That is indirect evidence and does not show the workflow nominates bad drugs. The 200-compound panel answers the question directly. If the lookup matches our method's top-20 precision, the harmonization effort is unjustified for this use, and we would report that.
Why not just apply the Smith et al. correction and stop? This is the real threat. It is plausible that replicate bias explains most false positives and that cross-study signatures add little. The ablation pairing the single-cohort signature with full correction measures this directly. We designed the panel to stratify across replicate-count bins so the replicate effect and the signature effect can be separated rather than confounded. If correction alone captures the gain, the paper's conclusion narrows to "always correct for replicate bias in joint repurposing." That is still useful, and the chondrocyte dataset stands either way.
Why not just dock compounds against a validated target such as DDR2? Docking answers a different question, binding to one protein, and Mei et al.'s best hit needed 45.73 μM to engage its target in cells. The literature on epitranscriptomic regulators also suggests osteoarthritis runs through parallel programs that a single target may not control: WTAP in Lin et al., FTO in Liu et al., and macrophage pyroptosis in Yao et al. Including docking scores on the same panel tests whether binding to DDR2 predicts reversal of the whole program.
Why not just train a deep model on L1000 to predict responses? Cancer drug-response work shows such models barely generalize to unseen compounds without very large training sets (Lavelle et al.). No chondrocyte perturbation set exists to train one. Our panel is, if anything, the first resource that could later fine-tune such a model. Even so, 200 compounds is a test set, not a training set.
Decisive Checks
The first check is transfer. Across the 200 compounds, does an L1000 reversal score predict chondrocyte reversal at all, beyond the random and replicate-count references? If the best method's held-out Spearman correlation is near zero, the honest conclusion is that cancer-line perturbation data cannot guide osteoarthritis repurposing. That result would itself redirect the field toward joint-cell perturbation atlases.
The second check is the replicate bias. Among compounds ranked highly by uncorrected scoring, those from the high-replicate bin should show lower measured reversal than equally ranked low-replicate compounds. The corrected ranking should remove this gap.
The third check is the head-to-head comparison against the strongest competitor, using top-20 precision on held-out donors and scaffolds, followed by the ablations that assign any gain to signature construction or profile correction.
Finally, the top five compounds from our method that were not on the panel, along with five from the strongest baseline, will be tested in chondrocytes. The secondary marker readout (MMP13, ADAMTS5, COL2A1, SOX9) will be measured blind to the method that nominated each compound. This gives a small but genuinely forward-looking test of hit rate.
Animal efficacy is deliberately out of scope. The rodent models used throughout the field, the surgical meniscus-destabilization (DMM) model and the chemically induced MIA model, are themselves a known translation weakness. Their place is after compound nomination has been shown to work.
References
- Similarity bias from consensus perturbational signatures from the L1000 Connectivity Map
- Identification of MEG3 and MAPK3 as potential therapeutic targets for osteoarthritis through multiomics integration and machine learning
- Comprehensive multiomics analysis identifies PYCARD as a key pyroptosis-related gene in osteoarthritis synovial macrophages
- Conjoint analysis of single-cell sequencing and high-throughput virtual screening regarding DDR2 in osteoarthritis disease models
- Machine Learning Based Drug Indication Prediction Using Linked Open Data
- Large-scale AI-Ready Data for Anti-Cancer Drug Response Modeling
- Deep learning for drug repurposing: Methods, databases, and applications
- SEA CDM: Study-Experiment-Assay Common Data Model and Databases for Cross-Domain Data Integration and Analysis
- Naringenin regulates production of matrix metalloproteinases in the knee-joint and primary cultured articular chondrocytes and alleviates pain in rat osteoarthritis model
- N6-methyladenosine (m6A) methyltransferase WTAP-mediated miR-92b-5p accelerates osteoarthritis progression
- FTO-mediated m6A demethylation of pri-miR-3591 alleviates osteoarthritis progression