Species-Anchored Cell Painting for Dual-Action Bone Drugs

Osteoporosis drugs that only stop bone breakdown also slow bone building, because the two processes are chemically coupled. The goal of "dual-action" therapy is to suppress bone-resorbing cells (osteoclasts) while stimulating bone-forming cells (osteoblasts). Candidates are almost always judged the same way. Someone runs a pair of stains on mouse or rat cells, gives the compound to ovariectomized mice or steroid-treated rats, and scans their bones with micro-computed tomography, an X-ray method that reconstructs bone in 3D at micrometer resolution. What this pipeline never asks is whether the cellular effect that made the drug look good in a rodent is the same effect it has on human or large-animal bone cells.

We propose to answer that question at the cell level, where it is cheap. We would run Cell Painting, a six-dye imaging assay that turns each cell's shape and organelles into a vector of thousands of numbers, on primary osteoblast and osteoclast precursors from mouse, rat, pig, and human, all exposed to the same compounds. We would put the four species in a shared coordinate system using a panel of reference compounds whose biology is well understood. Then we would test two things. First, whether an early (48-hour) profile predicts the bone outcome that micro-CT measures in rodents weeks later. Second, whether agreement between species in the dish predicts agreement between species in the animal. The deliverable is a cheap screen that flags compounds whose rodent efficacy rests on species-specific cell behavior before any animal work is done.

Core Claim

A compound's effect on bone-cell morphology splits into a component shared across species and a component specific to one species. Only the shared component should carry over from rodents to humans. We can estimate the shared component by aligning each species' profiles on a common set of reference perturbations. Once aligned, we claim three things. A 48-hour profile should rank compounds by their eventual micro-CT effect in mice better than the standard two-number readout (osteoblast mineralization plus osteoclast count). A classifier trained on rodent cells should label human-cell dual-action status better than one trained on unaligned profiles. And disagreement between mouse and rat in the dish should predict which compounds give discordant micro-CT results in mouse and rat in vivo. We treat that last prediction as the decisive one, because it is the only cross-species claim we can check against living bone within budget.

Why this is not already done

Three rodent studies show the current design. One tests a Dolichos lablab extract in cell culture and in senescence-prone mice with micro-CT morphometry (Dolichos Lablab Linné…). One tests a You-Gui-Yin herbal fraction on osteoclast signaling and on bone in ovariectomized mice (N-Butanol Extract of Modified You-Gui-Yin…). One tests alendronate plus Plastrum testudinis in steroid-treated rat spine (Promotion effect of extracts from plastrum testudinis…). All three use single-marker cell assays, thresholded micro-CT, and one species. They motivate the proposal by showing the pipeline we would replace.

The profiling literature supplies the method but has never touched bone lineages or other species. That work includes Cell Painting itself (Cell Painting, an image-based assay), the prediction of 70 cell-health readouts from profiles (Predicting cell health phenotypes), and the adaptation of frozen vision models for mechanism classification (Bridging Self-Supervision and Mechanism of Action Discovery). All of it lives in transformed human lines such as U2OS and A549. The file itself notes that models trained on such lines often fail to generalize to primary cells or other species.

Two constraints shape the design. The matched chemical–genetic resource CPJUMP1 reports that "only 5–25% of compounds correctly matched their 'sister' compounds" (Three million images…), and MOTIVE notes that some real pathway effects are morphologically silent (MOTIVE). We therefore do not assume profiles capture mechanism. We test directly whether they predict bone outcome.

On the imaging side, Local Binary Patterns have been applied to 2D slices of 24 human osteoarthritic bone cores. These are texture codes read straight from grayscale voxels without first separating bone from marrow, and their entropy correlated with histological grade at $R^2 = 0.67$ (Local Binary Patterns…). We extend these codes to 3D and use them as a secondary in vivo readout, without making them the centerpiece.

Research Program

Cells. We would isolate two primary cell types from each of four species: marrow stromal cells, differentiated toward osteoblasts, and marrow monocytes, differentiated toward osteoclasts with RANKL (the signal that makes precursors become osteoclasts). Human cells would come from surgical waste bone and pig cells from abattoir femurs, which keeps large-animal cost low.

Compounds. The library has about 300 compounds. It includes antiresorptives such as alendronate, putative dual-action agents from the file such as strontium salts, gallium nitrate, and morroniside, and single phytochemicals identified in the herbal mixtures (rutin, glycyrrhizin, trigonelline). The remainder is chosen to overlap with JUMP compounds so that existing U2OS profiles serve as a free baseline.

Imaging and labels. We image a modified Cell Painting panel at 48 hours. The modification adds a TRAP counterstain channel for osteoclasts, since TRAP is the standard osteoclast marker enzyme. Profiles are built from CellProfiler features and frozen DINO embeddings. Late ground truth is collected on the same plates: day-21 Alizarin Red mineralization for osteoblasts, and day-7 multinucleated TRAP-positive counts plus pit resorption for osteoclasts. These define a per-species in vitro dual-action label.

Species alignment. Every plate carries vehicle controls and a fixed anchor panel. The panel includes BMP-2, dexamethasone, graded RANKL, alendronate, a Wnt activator, and a cytotoxic control. For species $s$, let $x_s(c)$ be the control-normalized profile of compound $c$. We learn a linear map $W_s$ into a shared space by minimizing

$$\sum_{a \in A} \sum_{s} \lVert W_s x_s(a) - \bar z(a) \rVert^2,$$

where $A$ is the anchor set and $\bar z(a)$ is the cross-species mean of the mapped anchor responses. $W_s$ is constrained to be orthogonal so that species cannot collapse onto each other. The species-concordance score of a test compound is the cosine similarity of its mapped profiles across species. Anchors are never test compounds.

In vivo arm. About 16 compounds, spread across the concordance range, go into 12-week ovariectomized mice. Eight of them also go into dexamethasone-treated rats. Micro-CT reports standard thresholded morphometry (BV/TV, Tb.Th, Tb.Sp, Tb.N, BMD) and 3D grayscale texture. Serum CTX and PINP, blood markers of resorption and formation, are collected as a systemic comparator.

Baselines

The naive reference point is the two-scalar score bone labs already use: the osteoblast mineralization change plus the osteoclast count change, both measured late and in the same species. It costs almost nothing beyond the plates we already run. Any claim for profiling must beat it at the same measurement budget. "Beating it" means a higher Spearman correlation with in vivo BV/TV change across held-out compounds, under leave-one-compound-out evaluation. A 48-hour profile that merely ties a day-21 scalar would still count as a speed win. We would report that as a speed result, not an information result.

The strongest existing profiling method is frozen DINO embeddings with a TRex-style focal-loss multilayer perceptron (Bridging Self-Supervision and Mechanism of Action Discovery). On the JUMP-CP pilot, TRex reports mechanism-of-action mean average precision (mAP) rising from 0.15 to 0.32, and from 0.24 to 0.28 on unseen compounds. We would train it on the same bone-cell images and label set, with the same GPU budget and the same hyperparameter search size as our model, but without species alignment. I found no reported result on Lacuna for any profiling method applied to primary osteoblasts or osteoclasts, or to cross-species transfer, so these numbers anchor capability only, not the target.

The key-ingredient ablation pools the four species with ordinary per-plate normalization and no anchor-based maps. If this ablation matches the aligned model on rodent-to-human transfer AUROC, the alignment contributes nothing and we would say so. A second ablation drops the osteoclast arm, to test whether the dual-lineage design matters. A third compares 3D texture against thresholded morphometry as the in vivo target, to check that our conclusions do not depend on the choice of bone summary.

Why not just…?

Why not just use the ALP/Alizarin and TRAP readouts? This is the real threat. "Dual action" is by definition a two-number claim, and those two numbers may carry all the signal that predicts micro-CT. The in vivo arm settles it directly. If the two-scalar score correlates with mouse BV/TV as well as the aligned profile, the profile's remaining value is reduced to early readout and toxicity flagging, and we would narrow the claim to that. Way et al. showing that profiles predict apoptosis and ROS (Predicting cell health phenotypes) suggests profiles catch cytotoxic false positives that the scalars miss. That is a narrower benefit than predicting bone outcome.

Why not just reuse existing U2OS profiles from JUMP? U2OS cells are bone-derived, and the data already exist in the Cell Painting Gallery. Screening on them would cost nothing. They are, however, a transformed tumor line with no osteoclast counterpart and no species axis. Because our library overlaps JUMP, we run exactly this baseline: U2OS profiles as predictors of mouse micro-CT outcome and of human primary-cell labels. If U2OS matches primary cells, the primary-cell work is unnecessary.

Why not just measure serum CTX and PINP in the animals? These markers capture systemic turnover, as the alendronate-plus-extract rat study used them (Promotion effect of extracts from plastrum testudinis…). However, they still require an animal per compound, and they are still rodent-only. We collect them anyway and report how well they agree with micro-CT. They do not answer the cross-species question.

Why not just test in pigs in vivo? That would be the most direct translational check. It is also the cost and ethics barrier the file identifies (Bridging the Cross Species Clinical Translation Gap), and it cannot be run on 300 compounds. Pig cells in the dish are our cheap stand-in. Our design tells us whether dish-level concordance is informative enough to decide which few compounds deserve a pig study.

Decisive Checks

The first check is the mouse–rat test. Among the eight compounds run in both rodents, we predict that low in vitro mouse–rat concordance will mark the compounds whose in vivo BV/TV effects diverge between the two species. We test this with a permutation test on the rank correlation between concordance and in vivo discordance. With eight compounds this is underpowered for a small effect, so we pre-register the direction. A clear failure would kill the claim that dish concordance predicts anything about animal concordance.

The second check is rodent-to-human transfer AUROC for the in vitro dual-action label on held-out compounds, comparing aligned profiles, pooled profiles, and U2OS profiles.

The third check is qualitative but pointed. The reference antiresorptive should come out low-concordance in its formation-lineage profile if the model is capturing the coupling of formation to resorption. If it comes out high-concordance, we would suspect the model is reading only resorption suppression.