Moshkov et al. (Nature Communications, 2023) benchmarked three data modalities — chemical structure (CS), Cell Painting morphological profiles (MO), and L1000 gene-expression profiles (GE) — for predicting the outcomes of 270 historical drug-discovery assays, and found that existing data-fusion strategies fail to exploit the modalities' complementarity: late fusion of all three predicted 28 assays at AUROC (area under the receiver-operating-characteristic curve) > 0.9, no better than morphology alone (28) or CS+MO (31), and far short of the 52 assays attainable in retrospect by ideally selecting the best single modality per assay. The open problem is to design fusion methodology whose blind per-assay performance approaches that retrospective ideal-selection ceiling (on benchmarks controlled for a since-discovered cell-count confound), to identify and validate the determinants of assay predictability (target identity, assay biology, disease context) that the original study could not analyze, and, as deployment extensions, to establish transferability beyond the single U2OS cell line and historical compound library and to develop active-learning strategies for choosing which compounds to physically test for new assays. Fusion methods published through 2026 (refs 2, 6, 9–13) do not report recovering the ideal-selection complementarity, and no validated account of assay predictability determinants exists; the core remains open as of August 2026.

Background

In early drug discovery, a compound library can be characterized once by several profiling modalities and the resulting profiles then used to predict, without further experimentation, whether each compound will be active ('hit') in other biological assays — a strategy called 'profile once, predict many'. The three modalities studied here are: (i) chemical structure (CS), encoded as molecular fingerprints or as learned graph representations (the cited study used Chemprop, a directed-message-passing graph-convolutional neural network operating on the molecular graph); (ii) morphological profiling by Cell Painting, a high-content microscopy assay that stains eight cellular components with fluorescent dyes and extracts hundreds to thousands of quantitative image features per treatment (in the cited study, hand-engineered features from the open-source CellProfiler software); and (iii) transcriptomic profiling by the L1000 assay, which measures messenger-RNA levels of 978 'landmark' genes capturing most transcriptional variance. Predictive performance per assay is measured with AUROC (area under the receiver-operating-characteristic curve; 0.5 is random, 1.0 perfect), with AUROC > 0.9 conventionally called 'well predicted'. Two integration families were compared: early fusion (concatenating feature vectors before training one model) and late fusion (training one model per modality and aggregating their prediction scores). The key reference point is the retrospective ideal selection: an oracle that, for each assay, picks the single best-performing modality after looking at hold-out performance. It is not a deployable method — it quantifies the complementarity ceiling that real fusion should aim for.

Moshkov et al. (2023) assembled ~30,000 compounds profiled in U2OS cells (a human osteosarcoma cell line) by both Cell Painting and L1000, together with readouts of 270 non-redundant assays from historical Broad Institute projects, and trained per-assay predictors from each modality and their combinations. Each modality carried independent information, but fusion failed to harvest it: CS alone gave 16 assays with AUROC > 0.9, morphology 28, late fusion of CS+MO 31, yet fusing all three modalities gave 28 — no better than morphology alone — while the retrospective ideal selection could reach 52 assays (57 counting all predictor combinations). The authors attributed this to a noise-to-signal problem: for most assays at least one modality carries no signal, and a fusion method that cannot detect which modalities are informative dilutes the informative ones. At the more practical AUROC > 0.7 bar, 51–64% of assays were predictable with phenotypic data versus 37% from structure alone, so the stakes are practical: reliable fusion would let screening teams virtualize many future assays instead of physically screening full libraries. The authors explicitly listed what remained open: improved fusion strategies; how much learned representations (convolutional-network image features, gene-expression latent spaces, alternative chemical embeddings) would improve over the baseline features used; what determines whether an assay is predictable (target and assay type were unavailable as metadata); and strategies for choosing the sparse subset of compounds to physically test for a new assay so as to maximize predictive performance. The study covered a single cell line (U2OS) and one historical compound library, so transfer beyond them is untested. Two later results frame any answer: prospective validations of single phenotypic modalities have succeeded (an industrial Cell Painting study across 140 assays with experimental follow-up, and a dose-dependent transcriptomics model prospectively validated in a malaria assay), while a 2026 reanalysis showed that the number of cells in Cell Painting images alone — a proxy for cytotoxicity — predicts a substantial fraction of benchmark assays, so claims of fusion gains must now be evaluated against a cell-count baseline on assays not confounded by cell health.

Problem Statement

The problem, as posed by Moshkov et al. (2023) and unresolved as of August 2026, is to make the empirically demonstrated complementarity of chemical structure, Cell Painting morphology, and L1000 gene-expression profiles actually usable for compound activity prediction. Its core is a fusion objective: design a data-fusion strategy whose blind, per-assay predictive performance, assessed on the cell-count-controlled benchmark (the 270-assay benchmark with cell-count-confounded assays filtered out and a cell-count baseline added, per the 2026 reanalysis), approaches the retrospective ideal-selection ceiling while clearly exceeding the best single modality and the best pairwise late fusion, with the single-modality, pairwise-fusion and ideal-selection references likewise established on that controlled benchmark. That complementarity gap is quantified by the source's counts on the original, un-filtered benchmark — CS alone 16, morphology alone 28, best pairwise late fusion (CS+MO) 31, all-three late fusion 28, and 52 assays predictable at AUROC > 0.9 by ideally choosing the best modality per assay (57 counting all predictor combinations). Alongside this core, and stated by the source as separately open: identify and validate the determinants of assay predictability — target identity, assay type and assay biology, disease context, and the correlation between each modality and the bioactivity of interest — explaining when fusion helps and which modalities carry signal for which assays; and, as deployment extensions of the same 'profile once, predict many' objective, establish whether the three-modality conclusions transfer to other cell types, larger compound spaces, and prospective screening, and develop active-learning strategies that choose which sparse subset of compounds to physically test for a new assay so as to maximize predictive performance. These are parallel open questions rather than a single conjunctive completion rule: the source poses the fusion objective and the determinants question separately, and progress on the determinants, transferability, or active-learning strands does not by itself resolve the fusion objective, which in turn does not presuppose a completed determinants account.

References

  1. Moshkov, N.; Becker, T.; Yang, K.; Horvath, P.; Dancik, V.; Wagner, B. K.; Clemons, P. A.; Singh, S.; Carpenter, A. E.; Caicedo, J. C. Predicting compound activity from phenotypic profiles and chemical structures. Nature Communications 14, 1967 (2023). https://doi.org/10.1038/s41467-023-37570-1
  2. Seal, S.; Yang, H.; Trapotsi, M.-A.; Singh, S.; Carreras-Puigvert, J.; Spjuth, O.; Bender, A. Merging bioactivity predictions from cell morphology and chemical fingerprint models using similarity to training data. Journal of Cheminformatics 15, 56 (2023). https://doi.org/10.1186/s13321-023-00723-x
  3. Haslum, J. F.; Lardeau, C.-H.; Karlsson, J.; Turkki, R.; Leuchowius, K.-J.; Smith, K.; Müllers, E. Cell Painting-based bioactivity prediction boosts high-throughput screening hit-rates and compound diversity. Nature Communications 15, 3470 (2024). https://doi.org/10.1038/s41467-024-47171-1
  4. Seal, S.; Dee, W.; Shah, A.; Cerisier, N.; Zhang, A.; Miglietta, E.; Titterton, K.; Cabrera, Á. A.; Boiko, D.; Beatson, A.; Slabaugh, G.; Taboureau, O.; Carreras Puigvert, J.; Singh, S.; Spjuth, O.; Bender, A.; Carpenter, A. E. Counting cells can accurately predict small-molecule bioactivity benchmarks. Nature Communications 17, 2436 (2026). https://doi.org/10.1038/s41467-026-68725-5
  5. Herman, D.; Kańduła, M. M.; Freitas, L. G. A.; van Dongen, C.; Le Van, T.; Mesens, N.; Jaensch, S.; Gustin, E.; Micholt, L.; Lardeau, C.-H.; Varsakelis, C.; Reumers, J.; Zoffmann, S.; Will, Y.; Peeters, P. J.; Ceulemans, H. Leveraging Cell Painting Images to Expand the Applicability Domain and Actively Improve Deep Learning Quantitative Structure–Activity Relationship Models. Chemical Research in Toxicology 36, 1028–1036 (2023). https://doi.org/10.1021/acs.chemrestox.2c00404
  6. Sun, Y.; Zhang, X.; Zheng, Q.; Li, H.; Zhang, J.; Hong, L.; Wang, Y.; Zhang, Y.; Xie, W. Phenotypic Bioactivity Prediction as Open-set Biological Assay Querying. bioRxiv preprint (2026). https://doi.org/10.64898/2026.02.28.708711
  7. Serrano, E.; Peters, J.; Wagner, J.; Graham, R. E.; Chen, Z.; Feng, B. Y.; Miranda, G.; Kalinin, A. A.; Vulliard, L.; Tomkinson, J.; Mattson, C.; Lippincott, M. J.; Kang, Z.; Sitani, D.; Bunten, D.; Seal, S.; Carragher, N. O.; Carpenter, A. E.; Singh, S.; Marin Zapata, P. A.; Caicedo, J. C.; Way, G. P. Progress and new challenges in image-based profiling. Molecular Systems Biology 22, 624–658 (2026). https://doi.org/10.1038/s44320-026-00197-7
  8. Moshkov, N.; Bornholdt, M.; Benoit, S.; Smith, M.; McQuin, C.; Goodman, A.; Senft, R. A.; Han, Y.; Babadi, M.; Horvath, P.; Cimini, B. A.; Carpenter, A. E.; Singh, S.; Caicedo, J. C. Learning representations for image-based profiling of perturbations. Nature Communications 15, 1594 (2024). https://doi.org/10.1038/s41467-024-45999-1
  9. Liu, G.; Seal, S.; Arevalo, J.; Liang, Z.; Carpenter, A. E.; Jiang, M.; Singh, S. Learning Molecular Representation in a Cell. ICLR 2025; arXiv:2406.12056 (2024). https://arxiv.org/abs/2406.12056
  10. Li, M.; Zang, Z.; Xing, W.; Chen, J.; Zhang, R.; Luo, J.; Li, S. Z. Learning Cell-Aware Hierarchical Multi-Modal Representations for Robust Molecular Modeling. AAAI 2026; arXiv:2511.21120 (2025). https://arxiv.org/abs/2511.21120
  11. Rao, J.; Lin, H.; Chen, L.; Xie, J.; Zheng, S.; Yang, Y. Multi-modal Contrastive Learning with Negative Sampling calibration for Phenotypic Drug Discovery. Proceedings of CVPR 2025, 30752–30762 (2025). https://doi.org/10.1109/CVPR52734.2025.02864
  12. Wang, S.; Han, Q.; Qin, W.; Wang, L.; Yuan, J.; Cai, F.; Zhao, Y.; Ren, P.; Zhang, Y.; Tang, Y.; Li, R.; Li, Z.; Zhang, W.; Gao, S.; Bai, F. PhenoModel: A multimodal phenotypic drug design foundation model for discovering novel potential inhibitors of multiple cancer cells. Acta Pharmaceutica Sinica B 16, 1219–1232 (2026). https://doi.org/10.1016/j.apsb.2025.09.036
  13. Lian, X.; Ma, P.; Ma, T.; Ren, Z.; Cai, X.; Cheng, Z.; Song, B.; Wang, H.; Pan, X.; Chen, Y.; Yuan, S.; Lin, C. Empowering chemical structures with biological insights for scalable phenotypic virtual screening. Bioinformatics 42, btag517 (2026). https://doi.org/10.1093/bioinformatics/btag517
  14. Chandrasekaran, S. N.; Ackerman, J.; Alix, E.; Ando, D. M.; Arevalo, J.; Bennion, M.; Boisseau, N.; Borowa, A.; Boyd, J. D.; Brino, L.; et al. (JUMP Cell Painting Consortium). JUMP Cell Painting dataset: morphological impact of 136,000 chemical and genetic perturbations. bioRxiv preprint (2023). https://doi.org/10.1101/2023.03.23.534023
  15. Xiang, L.; Wang, Y.; Shao, W.; Li, Q.; Yu, X.; Wei, M.; Gui, Y.; Li, S.; Qin, P.; Hu, C.; Zhang, G.; Zhang, X.; Wang, J.; Li, Y.; An, J.; Luo, Y.; Liao, Y.; Deng, J.; Tai, X.; Xu, R. Y.; Huang, L.; Guo, D.; Zhang, G.; Zhang, Z.; Xie, Z.; Deng, Y.; Xu, J.; Wang, D. High-throughput profiling of chemical-induced gene expression across 93,644 perturbations. Nature Methods 22, 1954–1963 (2025). https://doi.org/10.1038/s41592-025-02781-5
  16. Pahl, A.; Liu, J.; Patil, S.; Rezaei Adariani, S.; Schölermann, B.; Warmers, J.; Bonowski, J.; Koska, S.; Akbulut, Y.; Seitz, C.; Sievers, S.; Ziegler, S.; Waldmann, H. Illuminating Dark Chemical Matter Using the Cell Painting Assay. Journal of Medicinal Chemistry 67, 8862–8876 (2024). https://doi.org/10.1021/acs.jmedchem.4c00160
  17. Godinez, W. J.; Trifonov, V.; Fang, B.; Kuzu, G.; Pei, L.; Guiguemde, W. A.; Martin, E. J.; King, F. J.; Jenkins, J. L.; Skewes-Cox, P. Compound activity prediction with dose-dependent transcriptomic profiles and deep learning. bioRxiv preprint (2023). https://doi.org/10.1101/2023.08.03.551883
  18. Abe, R.; Miyao, T.; Bajorath, J. Accurate prediction of activity cliff compounds based on bioactivity profiles depends on assay nearest neighbor relationships. Journal of Cheminformatics 18, 72 (2026). https://doi.org/10.1186/s13321-026-01210-9
  19. Bushiri Pwesombo, D.; Beese, C.; Schmied, C.; Sun, H. Semisupervised Contrastive Learning for Bioactivity Prediction Using Cell Painting Image Data. Journal of Chemical Information and Modeling 65, 528–543 (2025). https://doi.org/10.1021/acs.jcim.4c00835