The emerging synergy of experimental and computational approaches for therapeutic modulation of biomolecular condensates Authors: Priyesh Mohanty, Shiv Rekhi, Aahil Khambhawala, Qizan Chen, Jeetain Mittal Published: 2026 ABSTRACT Biomolecular condensates (BCs) are membraneless organelles which play roles in key biological functions such as RNA metabolism, signal transduction and DNA repair, reflecting their importance in cellular organization and function. The dysregulation of condensate self-assembly and its internal material properties due to aberrant phase separation has been linked to neurodegeneration, cancers, viral infections, and cardiac diseases. Consequently, there is growing interest in the discovery and development of therapeutic molecules, referred to as condensate modifiers (c-mods), that specifically target BCs and/or their components which are associated with disease. In this perspective, we first provide readers with a brief overview of the possible modes of action of c-mods and the strategies underlying their design for effective targeting of BCs. Next, we highlight the role of traditional computer-aided drug discovery (CADD) in synergy with modern AI/ML methods in targeting BCs as illustrated in recent studies. Finally, we discuss the physicohemical features of the condensate microenvironment and c-mods that enable the favorable partitioning of the latter, thereby opening new avenues for targeting “undruggable” proteins within the condensate microenvironment. We conclude by providing an overview of the challenges that remain to successfully integrate experiment and computation, and discuss potential strategies to overcome them. Introduction The intracellular environment is organized into compartments or organelles, usually by means of a lipid membrane, which separates its constituents from the surrounding cytoplasmic milieu. These organelles control their interior environment via selective transport of macromolecules, enabling them to carry out specific functions. Over the past decade, membraneless organelles often referred to as biomolecular condensates (BCs),[1,2] have been shown to play roles in key biological functions (e.g., gene regulation, stress response), reflecting their importance in cellular organization. From a mechanistic standpoint, BCs are generally believed to form via the phase separation of proteins, nucleic acids, and small molecules, leading to the formation of a macromolecule-rich phase with unique composition and material properties compared to the bulk cytoplasm[3–6]. Dysregulation of the assembly and material properties of BCs due to aberrant phase separation has been linked to numerous pathological conditions, such as neurodegenerative diseases, cancers, viral infections, and cardiac diseases[7–10]. It has become apparent that several key proteins implicated in these diseases are also enriched within BCs, prompting significant interest in the discovery and development of therapeutic molecules (small molecules, RNA, and peptides) referred to as condensate modifiers (c-mods), which modulate the self-assembly, composition, and material properties of BCs[11]. Aside from directly modulating the properties of BCs, another emerging possibility is to leverage them as hubs to target “undruggable” biomolecules, such as intrinsically disordered proteins (IDPs) or multi-domain proteins which contain a modular organization comprising of one or more folded domains and intrinsically disordered regions (IDRs). Due to their inherent conformational heterogeneity and lack of stable binding pockets,[12, 13] IDPs/IDRs generally exhibit weaker binding (Kd > 1 μM) compared to folded domains. They are also particularly conducive for the formation of dynamic higher-order assemblies such as BCs due to their compositional biases and sequence-encoded multivalent interactions [14,15]. The enrichment of these IDPs/IDRs within BCs as compared to the surrounding cellular milieu could allow for improved target engagement and thereby compensate for their weaker binding affinities. In this perspective, we first provide the readers with a brief overview of the possible modes of c-mod action, and current progress in development of strategies for modulation of BC properties. Next, we highlight the role of conventional computer-aided drug discovery (CADD) approaches and artificial intelligence/machine learning (AI/ML) methods in targeting BCs. We further provide the readers with an in-depth discussion on the physicochemical features of small molecules that dictate their propensity to localize within BCs (i.e., partitioning), opening avenues for targeting biomolecules within the condensate environment. In conclusion, we discuss the challenges associated with the successful integration of AI/ML and physics-based simulations with high-throughput screening assays for the rational design of c-mods and highlight potential strategies to overcome them. Mechanisms and strategies for therapeutic modulation of BCs The capacity of small molecules to modulate condensates has been reported in several studies. Hyman and colleagues provided evidence that ATP can inhibit or promote the formation of FUS condensates in a concentration dependent manner, pointing to a possible mechanism by which cells regulate condensates through endogenous metabolites[16]. Alkanediols such as 1,6 hexanediol have been used to assess the stability of condensates formed by some proteins[17–20]. A key insight into the potential therapeutic relevance of small molecule drugs for modulating condensate behavior was provided by Klein et al. through their investigation of the partitioning of cancer therapeutics into nuclear condensates[21]. They discovered that the molecules were preferentially localized into condensates formed by their targets, suggesting the action of these drugs may be, in part, linked to their ability to partition within condensates. Murcko and colleagues coined the term, c-mod,[11, 22] to describe molecules (e.g., small molecules, peptides, and RNA) that can modify the physicochemical properties and/or function of BCs to reverse the disease state. The objective underlying the use of a c-mod for disease modulation may be one of the following: Condensatopathy repair: A condensatopathy refers to the dysfunction of a BC that drives a specific disease phenotype. Therefore, repair of the disease phenotype involves disrupting the behavior of a disease-associated condensate without explicit consideration of a target biomolecule. Condensatopathy repair: A condensatopathy refers to the dysfunction of a BC that drives a specific disease phenotype. Therefore, repair of the disease phenotype involves disrupting the behavior of a disease-associated condensate without explicit consideration of a target biomolecule. Target inactivation: The goal here is to disrupt the abnormal function of a specific target biomolecule (usually a protein or nucleic acid species) within the native condensate or disfavor its partitioning within the condensate. Target inactivation: The goal here is to disrupt the abnormal function of a specific target biomolecule (usually a protein or nucleic acid species) within the native condensate or disfavor its partitioning within the condensate. C-mod classification based on their phenotypic outcomes A classification scheme for c-mods has been proposed based on their modes of action, i.e., the condensate phenotypes observed through microscopy as described previously[22]. Here we provide a broad overview of the possible modes of action for c-mods (Fig. 1) with representative examples: Dissolvers: Numerous studies observed an association between aberrant condensation and neurodegenerative diseases[7,23]. Mutations, post-translational modifications, and environmental factors can impair the dissolution of condensates formed by aggregation-prone proteins such as TDP-43, FUS, tau, and α-synuclein[22,24]. The persistence of these condensates beyond the duration of their physiological requirement can promote a transition from liquid-like droplets to solid-like aggregates. Therefore, small molecules that disrupt the intermolecular interactions stabilizing these condensates provide an opportunity to dissolve them before they mature into pathological aggregates. For example, planar compounds such as mitoxantrone and pyrvinium were shown to dissolve persistent TDP43 condensates localized to the cytoplasm of iPSc motor neurons carrying ALS-associated mutations in TDP-43 or FUS[25]. More recently, the small molecule, lipoamide, was also shown to promote the dissolution of stress granules (SGs) in a redox-state sensitive manner, leading to the resolution of ALS-associated pathological hallmarks in motor neurons and fly models[26]. Dissolvers: Numerous studies observed an association between aberrant condensation and neurodegenerative diseases[7,23]. Mutations, post-translational modifications, and environmental factors can impair the dissolution of condensates formed by aggregation-prone proteins such as TDP-43, FUS, tau, and α-synuclein[22,24]. The persistence of these condensates beyond the duration of their physiological requirement can promote a transition from liquid-like droplets to solid-like aggregates. Therefore, small molecules that disrupt the intermolecular interactions stabilizing these condensates provide an opportunity to dissolve them before they mature into pathological aggregates. For example, planar compounds such as mitoxantrone and pyrvinium were shown to dissolve persistent TDP43 condensates localized to the cytoplasm of iPSc motor neurons carrying ALS-associated mutations in TDP-43 or FUS[25]. More recently, the small molecule, lipoamide, was also shown to promote the dissolution of stress granules (SGs) in a redox-state sensitive manner, leading to the resolution of ALS-associated pathological hallmarks in motor neurons and fly models[26]. Inducers: In some instances, inducing condensate formation can mitigate disease-related effects. In spinal muscular atrophy, forming nuclear condensates via an antisense oligonucleotide (Nusinersen) can enhance mRNA processing, rescuing the expression of functional survival motor neuron protein, SMN2[27]. Promoting condensation was shown to be a viable strategy for the treatment of B-cell malignancies: a small molecule (BI-3082) induced oligomerization of the oncogenic transcriptional repressor BCL6, accelerating its intracondensate ubiquitination and subsequent proteasomal degradation [28]. In context of hepatocellular carcinoma, the small molecule RQ was demonstrated to force β-catenin into cytoplasmic condensates in liver cancer cells which prevented nuclear localization of the former and subsequent activation of cancer-promoting genes, thereby resulting in the suppression β-catenin-driven tumor growth[29]. In contrast, sulforaphane induces the formation of nuclear β-catenin condensates, and reduces the concentration of the free protein which is implicated in the progression of colorectal cancer[30]. Inducers: In some instances, inducing condensate formation can mitigate disease-related effects. In spinal muscular atrophy, forming nuclear condensates via an antisense oligonucleotide (Nusinersen) can enhance mRNA processing, rescuing the expression of functional survival motor neuron protein, SMN2[27]. Promoting condensation was shown to be a viable strategy for the treatment of B-cell malignancies: a small molecule (BI-3082) induced oligomerization of the oncogenic transcriptional repressor BCL6, accelerating its intracondensate ubiquitination and subsequent proteasomal degradation [28]. In context of hepatocellular carcinoma, the small molecule RQ was demonstrated to force β-catenin into cytoplasmic condensates in liver cancer cells which prevented nuclear localization of the former and subsequent activation of cancer-promoting genes, thereby resulting in the suppression β-catenin-driven tumor growth[29]. In contrast, sulforaphane induces the formation of nuclear β-catenin condensates, and reduces the concentration of the free protein which is implicated in the progression of colorectal cancer[30]. Localizers: The functional roles of BCs depend on their internal environments, which they regulate their composition by selectively concentrating specific biomacromolecules[31–33]. Aberrant changes in condensate composition can ultimately lead to dysfunction and disease. Therefore, c-mods that modify protein-protein and protein-RNA interaction patterns, and prevent their recruitment to or promote expulsion from condensates, offer a means for compositional modulation without complete dissolution. For example, a frameshift mutation in the nuclear export signal (NES) of the nucleolar protein, NPM1, causes its mislocalization from the nucleolus to the cytoplasm by exportin-1, resulting in acute myeloid leukemia (AML). The small molecule, Avrainvillamide,[34] effectively blocked this process by interacting with both NPM1 and exportin-1, and thereby promoted NPM1 retention within the nucleolus[35]. Localizers: The functional roles of BCs depend on their internal environments, which they regulate their composition by selectively concentrating specific biomacromolecules[31–33]. Aberrant changes in condensate composition can ultimately lead to dysfunction and disease. Therefore, c-mods that modify protein-protein and protein-RNA interaction patterns, and prevent their recruitment to or promote expulsion from condensates, offer a means for compositional modulation without complete dissolution. For example, a frameshift mutation in the nuclear export signal (NES) of the nucleolar protein, NPM1, causes its mislocalization from the nucleolus to the cytoplasm by exportin-1, resulting in acute myeloid leukemia (AML). The small molecule, Avrainvillamide,[34] effectively blocked this process by interacting with both NPM1 and exportin-1, and thereby promoted NPM1 retention within the nucleolus[35]. Morphers: The intermolecular interactions in condensates give rise to distinct material states, ranging from viscous liquids to viscoelastic gels, which in turn govern their function[36,37]. The liquid-like nature of transcriptional condensates enables the rapid diffusion of transcription factors between the condensate and its surroundings, providing a mechanism for cells to regulate transcription[38]. The nuclear pore complex (NPC) forms a selectively permeable barrier to regulate the transport of biomacromolecules into the nucleus due to its dense cross-linked gel-like structure[39]. Changes to the material state of the NPC can impair its selectivity [40] potentially resulting in the mislocalization of key nuclear proteins such as TDP-43, FUS, and NPM1[41]. Beyond determining the material properties, intermolecular interactions also dictate the spatial organization of components within condensates to achieve compartmentalization. For instance, the nucleolus is organized into three distinct subcompartments, each with unique functions[42]. Environmental cues, such as oxidative stress, have also been shown to promote a morphological transition of SGs, whereby the proteins G3BP1 and TDP-43 demix, resulting in a TDP-43-enriched phase which can then form pathological aggregates[43]. A notable example of a c-mod which affects the condensate material properties was reported by Altmeyer and colleagues[44] who discovered that the steroidal alkaloid, cyclopamine, and its analogue A3E, inhibited respiratory synctial virus (RSV) condensates by promoting its “hardening” which was associated with a reduction in the intracondensate dynamics of the transcription factor, M2–1, as inferred from FRAP experiments. Similarly, Zhang et al. devised an aromatic, non-natural micropeptide referred to as the “killswitch”, which was able to arrest the internal dynamics of a variety of endogenous (e.g., nucleolus) and disease-associated (e.g., fusion oncoprotein and adenoviral) nuclear condensates, by directly altering the local microenvironment upon its recruitment[45]. Morphers: The intermolecular interactions in condensates give rise to distinct material states, ranging from viscous liquids to viscoelastic gels, which in turn govern their function[36,37]. The liquid-like nature of transcriptional condensates enables the rapid diffusion of transcription factors between the condensate and its surroundings, providing a mechanism for cells to regulate transcription[38]. The nuclear pore complex (NPC) forms a selectively permeable barrier to regulate the transport of biomacromolecules into the nucleus due to its dense cross-linked gel-like structure[39]. Changes to the material state of the NPC can impair its selectivity [40] potentially resulting in the mislocalization of key nuclear proteins such as TDP-43, FUS, and NPM1[41]. Beyond determining the material properties, intermolecular interactions also dictate the spatial organization of components within condensates to achieve compartmentalization. For instance, the nucleolus is organized into three distinct subcompartments, each with unique functions[42]. Environmental cues, such as oxidative stress, have also been shown to promote a morphological transition of SGs, whereby the proteins G3BP1 and TDP-43 demix, resulting in a TDP-43-enriched phase which can then form pathological aggregates[43]. A notable example of a c-mod which affects the condensate material properties was reported by Altmeyer and colleagues[44] who discovered that the steroidal alkaloid, cyclopamine, and its analogue A3E, inhibited respiratory synctial virus (RSV) condensates by promoting its “hardening” which was associated with a reduction in the intracondensate dynamics of the transcription factor, M2–1, as inferred from FRAP experiments. Similarly, Zhang et al. devised an aromatic, non-natural micropeptide referred to as the “killswitch”, which was able to arrest the internal dynamics of a variety of endogenous (e.g., nucleolus) and disease-associated (e.g., fusion oncoprotein and adenoviral) nuclear condensates, by directly altering the local microenvironment upon its recruitment[45]. Molecular-level strategies for targeting disease-associated condensates and their components The ultimate goal of c-mod development is to design therapeutic molecules with well-defined modes of action at the molecular level in order to achieve a more precise control over condensate assembly, morphology, and dynamics. Below, we discuss examples of studies where rich insights into c-mod action were garnered at the molecular-level, and/or their application was demonstrated to achieve the rational design of c-mods with predictable outcomes: Modulation of scaffold interaction network: The stability, composition, and material properties of BCs is critically dependent on the strength and valency of interactions between (co)-scaffolds that recruit “client” molecules. C-mods which perturb the multivalent interaction network through either weak (nonspecific) or high affinity (site-specific) interactions can exert significant effects on the general properties of condensates. Yeo and colleagues reported the discovery that small molecules with extended planar aromatic moieties such as Mitoxantrone (a topoisomerase inhibitor) which were found to block the recruitment of ALS-associated RNA-binding proteins (clients) such as TDP-43, FUS, and HNRNPA2B1 to SGs in iPSc neurons, possibly through chelation with nucleic acids,IDRs, and the disruption of protein-RNA interactions[25]. Taylor and colleagues[46] designed two peptide-inspired small molecules which bound to a specific pocket in the NTF2L domain of scaffold proteins - G3BP1/2 with sub-micromolar affinity. These peptidomimetics were shown to disrupt the co-condensation of RNA, G3BP1, and caprin 1 in vitro, prevent SG formation prior to or during stress, and dissolve pre-existing SGs. Modulation of scaffold interaction network: The stability, composition, and material properties of BCs is critically dependent on the strength and valency of interactions between (co)-scaffolds that recruit “client” molecules. C-mods which perturb the multivalent interaction network through either weak (nonspecific) or high affinity (site-specific) interactions can exert significant effects on the general properties of condensates. Yeo and colleagues reported the discovery that small molecules with extended planar aromatic moieties such as Mitoxantrone (a topoisomerase inhibitor) which were found to block the recruitment of ALS-associated RNA-binding proteins (clients) such as TDP-43, FUS, and HNRNPA2B1 to SGs in iPSc neurons, possibly through chelation with nucleic acids,IDRs, and the disruption of protein-RNA interactions[25]. Taylor and colleagues[46] designed two peptide-inspired small molecules which bound to a specific pocket in the NTF2L domain of scaffold proteins - G3BP1/2 with sub-micromolar affinity. These peptidomimetics were shown to disrupt the co-condensation of RNA, G3BP1, and caprin 1 in vitro, prevent SG formation prior to or during stress, and dissolve pre-existing SGs. Modulating the conformational and interaction landscape of “undruggable” targets: BCs are enriched in multi-domain proteins which comprise of both folded and IDRs. IDRs exhibit high conformational heterogeneity and do not possess stable binding pockets and, making the development of high potency binders extremely challenging. Further, self- or co-IDR interactions can also promote nucleation within condensates to give rise to amyloid fibrils with varying morphologies[47–49]. Vendruscolo and colleagues[50] reported that the aminosterol, claramine, stabilized α-synuclein condensates and inhibited its liquid-to-amyloid transition both in vitro and in a C. elegans model of Parkinson’s disease. Based on chemical kinetics experiments, they established that the primary mode of action for claramine was to inhibit primary amyloid nucleation within the condensates. Wheeler and colleagues[26] discovered that lipoamide could interact with and stabilize the methionine-oxidized form of the SG protein, SFPQ (splicing factor proline- and glutamine-rich), to promote SG dissolution and thereby counteract the detrimental effects of prolonged SG accumulation. Modulating the conformational and interaction landscape of “undruggable” targets: BCs are enriched in multi-domain proteins which comprise of both folded and IDRs. IDRs exhibit high conformational heterogeneity and do not possess stable binding pockets and, making the development of high potency binders extremely challenging. Further, self- or co-IDR interactions can also promote nucleation within condensates to give rise to amyloid fibrils with varying morphologies[47–49]. Vendruscolo and colleagues[50] reported that the aminosterol, claramine, stabilized α-synuclein condensates and inhibited its liquid-to-amyloid transition both in vitro and in a C. elegans model of Parkinson’s disease. Based on chemical kinetics experiments, they established that the primary mode of action for claramine was to inhibit primary amyloid nucleation within the condensates. Wheeler and colleagues[26] discovered that lipoamide could interact with and stabilize the methionine-oxidized form of the SG protein, SFPQ (splicing factor proline- and glutamine-rich), to promote SG dissolution and thereby counteract the detrimental effects of prolonged SG accumulation. Targeting condensation to alter regulatory processes: Dys-regulation of enzymatic activity associated with BCs such as DNA repair and transcription can lead to aberrant gene expression which is associated with numerous types of cancers. Several fusion proteins composed of IDRs and DNA-binding domains (DBDs), are strongly associated with aberrant gene expression [51,52]. Through phenotypic screening, Wang et al. discovered a small molecule (LY2835219), which could effectively dissolve condensates in Ewing sarcoma fusion-expressing cell lines and partially reverse aberrant gene expression[53]. Using a similar approach, Xie et al. discovered a small molecule - ET516 which selectively disrupted androgen receptor condensation and suppressed its transcriptional activity, thereby inhibiting the proliferation and tumor growth of prostate cancer cells expressing AR-resistant mutants[54]. Targeting condensation to alter regulatory processes: Dys-regulation of enzymatic activity associated with BCs such as DNA repair and transcription can lead to aberrant gene expression which is associated with numerous types of cancers. Several fusion proteins composed of IDRs and DNA-binding domains (DBDs), are strongly associated with aberrant gene expression [51,52]. Through phenotypic screening, Wang et al. discovered a small molecule (LY2835219), which could effectively dissolve condensates in Ewing sarcoma fusion-expressing cell lines and partially reverse aberrant gene expression[53]. Using a similar approach, Xie et al. discovered a small molecule - ET516 which selectively disrupted androgen receptor condensation and suppressed its transcriptional activity, thereby inhibiting the proliferation and tumor growth of prostate cancer cells expressing AR-resistant mutants[54]. Co-opting regulatory processes to modulate condensation: Post-translational (chemical) modification of a condensate component (e.g., phosphorylation, acetylation, and ubiquitination)[55] by certain cellular enzymes, can be exploited by cells to control the assembly and activity of BCs. In principle, therapeutic molecules which bind and stimulate the activity of enzymes critical to the formation and function of BCs can be exploited to achieve regulatory control in a dose-dependent manner. For example, it is possible to develop heterobifunctional small molecules known as PROTACS (Proteolysis Targeting Chimeras),[56] which may selectively bind a target protein associated with a disease condensate and promote its ubiquitination by additionally recruiting a specific E3 ligase in its proximity[57,58]. Consequently, ubiquitination of the target can regulate its cellular concentration and availability for condensate formation via proteasomal degradation as shown in the case of a BRD4-targeting PROTAC which regulates formation of the super-enhancer condensate[59]. Co-opting regulatory processes to modulate condensation: Post-translational (chemical) modification of a condensate component (e.g., phosphorylation, acetylation, and ubiquitination)[55] by certain cellular enzymes, can be exploited by cells to control the assembly and activity of BCs. In principle, therapeutic molecules which bind and stimulate the activity of enzymes critical to the formation and function of BCs can be exploited to achieve regulatory control in a dose-dependent manner. For example, it is possible to develop heterobifunctional small molecules known as PROTACS (Proteolysis Targeting Chimeras),[56] which may selectively bind a target protein associated with a disease condensate and promote its ubiquitination by additionally recruiting a specific E3 ligase in its proximity[57,58]. Consequently, ubiquitination of the target can regulate its cellular concentration and availability for condensate formation via proteasomal degradation as shown in the case of a BRD4-targeting PROTAC which regulates formation of the super-enhancer condensate[59]. Role of computer-aided drug discovery (CADD) in combination with AI/ML for targeting BCs CADD approaches have seen rapid adoption over the last decade across the entire drug discovery pipeline due to the generation of enormous data on target protein structures and their interactions with ligands, the rapid expansion of make-on-demand chemical libraries (>1 billion compounds), breakthroughs in AI/ML, and abundant computation capabilities[60–62]. For early-stage hit discovery against a target protein, virtual screening[63] (VS) is a well-established and cost-effective alternative to expensive high-throughput chemical screens and falls primarily into two distinct categories: (i) ligand-based: where the target structure is not required and/or is based on previously identified active compounds which bind to and modulate target activity and (ii) structure-based: wherein a reliable three-dimensional structure of the target protein is used to predict the site-specific binding of ligand molecules. Successful applications of both approaches[64–66] relevant to the therapeutic targeting of BCs are described in the following sub-sections. Ligand-based virtual screening Ligand-based VS relies on the knowledge of active compounds against the target protein based on in vitro or cell-based assays (Fig. 2 A). Here, the underlying assumption is that molecules which share similar structural and chemical features with known active compounds could possess similar activity towards the target. Common approaches include the use of a molecular similarity method (e.g., Tanimoto coefficient) computed based on 1D-fingerprints, pharmacophore mapping, and 3D-molecular shape matching[67,68]. The successful application of ML models for hit identification relies on high quality data generated from high-throughput screening experiments which typically aim to quantify the effect of small molecules on a cellular phenotype based on fluorescence imaging. High-throughput screens can be carried out either in vitro or in cell-based systems, and the latter approach is commonly referred to as high-content screening (HCS)[69]. Narlikar and colleagues developed an in vitro screening platform based on chromatin condensation to recreate physiological or disease-relevant chromatin states and identify small molecule modulators of the disease-state[70]. Further, they confirmed that the small molecule hits identified in vitro against a heterochromatin state mediated by the protein HP1α, also impacted chromatin structure within cells. An automated HCS platform[71] (Corelet™) was developed by Nereid therapeutics to identify small molecule modulators of BCs (upto hundreds of thousands of compounds) based on a blue light-inducible condensate formation system that does not require chemical stimuli or substantial overexpression of fluorescently-tagged proteins. These attributes ultimately minimize cellular perturbation, resulting in greater accuracy for the identification of BC modulators. Given the limited resolution of fluorescence imaging used in HCS for the identification of phenotypic changes associated with micron-scale condensates, Silver and colleagues developed two alternative cell-based assays based on high-throughput single molecule tracking (htSMT) and proximity-based condensate biosensors[72]. The authors carefully demonstrated that both approaches could reliably report on the condensation of SARS-CoV-2 nucleocapsid (N) protein and be employed to rapidly screen small molecule libraries, thereby providing an accurate readout of c-mod activity independent of HCS. By coupling the large amount of data generated from screening experiments with ML-based approaches, it is possible to construct quantitative structural activity relationships (QSARs)[73] and predict c-mods from substantially larger and diverse compound libraries for experimental validation. Such an approach was demonstrated in a recent study by Yu and co-workers[74] who performed an imaging-based HCS of ~5000 FDA-approved compounds for their activity towards TDP-43-containing anisosomes in cells[75]. To predict c-mods from a larger compound library (>170,000 compounds), they employed a graph convolutional network (GCN), which is based on molecular graphs, to build a QSAR model which predicted ~1100 candidates (0.65%) with potential biological activity against TDP-43 anisosomes. Out of the top five compounds chosen for experimental validation, two were found to reduce the number and size of TDP-43 anisosomes. Structure-based virtual screening Structure-based VS typically relies on molecular docking algorithms [76] which enable an efficient search of possible ligand binding sites on the surface of the target protein structure and ranks protein-ligand complexes based on an empirical scoring function (Fig. 2 B). Khanna and co-workers[77] applied structure-based VS to modulate the toxicity of TDP-43, a nuclear multi-domain protein which localizes to cytoplasmic SGs under cellular stress and forms pathological aggregates linked to neurodegenerative diseases such as amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD)[43]. Here, the authors performed targeted docking of a small library (50K compounds) against the RNA-recognition motif-1 (RRM1) domain of TDP-43, and determined that five out of eight top-scoring compounds were true binders based on saturation transfer difference (STD) NMR experiments. Interestingly, the highest affinity binder based on MST experiments, rTRD01 (Kd ~ 89 μM), was found to reduce the affinity of TDP-43 RRMs towards the hexanucleotide RNA-repeat of the disease-linked c9orf72 gene which is found in SGs and improve larval locomotion in an ALS fly model. Although the study did not explicitly investigate the ability of rTRD01 to affect the extent of TDP-43 localization to SGs in stressed neurons, it demonstrates that structure-based VS can identify small molecules that directly bind TDP-43 and modulate its RNA interactions, providing a potential route to influence disease-relevant biochemical activity. Active ML can enable efficient prediction of hits through iterative docking, model training and inference steps without the need to dock all molecules from ultra-large chemical libraries[76]. The adoption of such efficient, high-throughput strategies paves the way for the rapid discovery of diverse compounds across a wide range of binding affinities and specificities. The inherent conformational plasticity of IDR/IDPs, which are important for condensate formation, poses a challenge for conventional CADD methods. This has motivated the recent development of ensemble docking,[78] wherein molecular dynamics (MD) simulations of IDRs are conducted and small molecules are docked to its ensemble. Docked ensembles of three small molecules to α-synuclein generated following this approach were found to agree well with experiments. This suggests that common docking scoring functions could also be used for IDRs, provided the conformational ensemble of the protein is sampled adequately. Features of the condensate microenvironment and small molecules which govern partitioning Given that BCs are typically enriched in a variety of macromolecules (such as multidomain proteins, DNA, and RNA), it is vital to establish a comprehensive physicochemical framework for understanding the basis of small molecule partitioning into condensates and ultimately develop c-mods with high efficacy towards target condensates and their components. Here, we provide a detailed discussion of the emerging insights from recent studies which investigated small molecule partitioning into condensates and constructed QSAR models based on AI/ML and HCS experiments. In addition, we discuss computational studies which utilized physics-based simulations (atomistic and coarse-grained) to better understand the physicochemical basis of small molecule partitioning and provide complementary insights into the significance of the condensate microenvironment. Insights from QSAR studies utilizing HCS in combination with AI/ML Kilgore et al. were the first to comprehensively investigate the distinct chemical solvation environments presented by different condensates by quantifying the partitioning of small molecules into homotypic condensates (containing scaffold protein only) reconstituted in vitro[79]. To understand the physicochemical rules which govern the selective enrichment of small molecules into condensates with distinct chemical environments, the authors studied partitioning of >1500 intrinsically fluorescent small molecules (FDA-approved) spanning different chemical classes. By measuring the fluorescence intensity inside versus outside condensates for each molecule using microscopy, they reported partition coefficients (PCs) for the molecules in three different condensate systems. The authors observed an order of magnitude variation of PC for the molecules in their dataset and weak correlation between different condensate systems. Based on the Tanimoto similarity between Morgan fingerprints of the small molecules tested, they inferred that chemical probes with similar chemical features exhibited similar degree of selective enrichment within a given condensate. Subsequently, they constructed a QSAR model based on the chemical features of small molecules and their corresponding partitioning ratios using deep learning message-passing neural networks (MPNNs), to predict new molecules which exhibit selective partitioning into scaffold-only condensates in vitro. Interestingly, the QSAR model was also partially capable of predicting new drug molecules which exhibited selective partitioning into nuclear condensates in vivo (nucleoli, heterochromatin), with chemical complementarity, supporting their proposed hypothesis that distinct chemical environments presented by BCs influence the partitioning efficiency of small molecules belonging to different chemical classes. An investigation into factors which govern the small molecule partitioning of BCs was carried out by Thody et al.[80] wherein the authors quantified the extent of partitioning of ~1700 small molecules (metabolites and FDA-approved compounds) for four two-component condensates, each with a distinct macromolecular composition. In contrast to fluorescence-based approaches, partitioning was quantified in pooled sub-libraries (~300 compounds) via mass spectrometry and subsequently validated for a subset of 30 intrinsically fluorescent molecules using confocal microscopy. For a given condensate system, the authors observed a nearly six-orders of magnitude variation in the extent of partitioning across the small molecules evaluated, and despite differences in their macromolecular composition, the partitioning behavior was correlated across all condensates. Finally, a ML model constructed based on the physicochemical features of small molecules determined that partitioning of small molecules was largely dictated by two physical features: the aqueous solubility and hydrophobicity, as opposed to site-specific interactions with macromolecules. Overall, the works of Kilgore et al. and Thody et al. provide complementary perspectives on small-molecule partitioning into BCs by leveraging distinct experimental modalities and chemical libraries. Taken together, their findings suggest that the same molecules partition either strongly or weakly across different condensates over a wide range of partition coefficients, while condensate-specific effects emerge within a narrower range of PC values (~0–1). The above findings imply that a deeper understanding of the influence of condensate microenvironment on partitioning efficiency is critical to achieving target specificity. Toward this goal, Ouyang et al. assessed the influence of micropolarity within the condensate environment on drug-targeting efficiency both in vitro and in cellulo, and observed that higher enrichment of non-polar (hydrophobic) residues within condensates was generally associated with favorable partitioning of more hydrophobic drug molecules[81]. These observations were further corroborated in the case of estrogen receptor 1 (ESR1) whose IDR is enriched in non-polar residues, wherein inhibitors with both high target affinity and hydrophobicity showed improved partitioning efficiency into its condensates versus those with high affinity and lower hydrophobicity. Finally, Random Forest models constructed to predict inhibitor targeting efficiency from molecular properties, identified hydrophobicity as a key determinant of drug-partitioning efficiency. Overall, these observations are consistent with the findings of Thody et al. and establish that the hydrophobicity of the condensate microenvironment is an important feature which must be taken into consideration for the development of c-mods with high target efficiency. Insights from multiscale molecular dynamics (MD) simulations The insights from experiments point to the microenvironment within condensates as a dominant contributor to the selective inclusion of small molecules. However, detailed investigations of the interior environment in the dense phase and the role of protein sequence in determining the environment are challenging using experiments alone at present. As a result, molecular simulations of condensates at atomistic resolution have emerged as a valuable tool to bridge this gap and uncover the physicochemical features of the condensate microenvironment[82–84]. In the context of the driving forces for the partitioning of small molecules, a key feature is the solvation environment within the condensate [79–81]. The primary determinant of the solvation environment is the relative enrichment of protein and water in the dense phase. Previous studies of condensates at atomistic resolution, as well as NMR experiments, have demonstrated that the interior environment can approach ~60% water by weight, indicating a relatively hydrated environment [82,85,86]. Atomistic simulations of minimal peptide fragments suggest that the extent of hydration is strongly sequence-dependent with more polar-rich peptides leading to more hydrated environments[87,88]. Investigations of the partitioning of amino acid sidechains between the dilute and dense phase of model protein condensates reveal that due to hydration, the transfer free energies deviate from predictions using measures of solvation free energy[89,90]. A molecular view of the interaction modes that facilitate the incorporation of small molecules within the condensate can be achieved using simulations at atomistic resolution. For instance, Wang et al. computed interaction fingerprints from simulations of a set of small molecules in three protein condensates and proposed that the molecules form distinct interaction networks within the condensate[91]. This suggests that small molecules selectively bind to certain regions of the protein within the crowded dense phase. While conventional CADD approaches are useful to identify small molecules that bind to IDRs in the dilute phase,[78,92] two main challenges arise within the condensate: (i) accounting for possible conformational changes of the IDR within the dense phase[93] and (ii) ensuring the molecule partitions into the condensate. The long time- and length-scales required to investigate phenomena associated with BCs have led to the adoption of multiscale simulations to elucidate the partitioning and molecular interactions of small molecules. Kelley et al. used minimal coarse-grained (CG) models to demonstrate that, while the partitioning of molecules into the condensate reduces with increasing size, sufficiently favorable interactions between the particles and the condensate-forming protein can lead to the incorporation of arbitrarily large molecules in the dense phase[94]. To investigate the partitioning of small molecules, Emelianova et al. used atomistic simulations of minimal peptide fragments and small molecules in the slab geometry to directly calculate the partition coefficient of the molecules between phases[88]. To facilitate the modeling of biologically-relevant proteins and small molecules, they developed a method for the parametrization of small molecules[88] in the framework of the Mpipi CG model[95] for protein phase separation. An alternate method for the estimation of partition coefficients is through the direct measure of the transfer free energies of the molecules from the dilute to the dense phase of protein condensates. While such approaches have been developed to calculate the transfer free energies of amino acids,[89,90] they have not yet been extended to larger (drug-like) molecules. Together, the approaches above provide a path for understanding the relationship between the chemical features of molecules and their localization within condensates. This is particularly relevant in the case of multiphasic condensates wherein each subphase has a distinct chemical environment[42,96]. Recent work also suggests that even in some single component systems, the condensate interior can have clusters of higher hydrophobicity than the surrounding regions[97]. The implications for the design of c-mods are that the chemical features of the molecules can be tuned to achieve optimal localization to a particular subphase or region within the condensate. Through computational modeling, substantial progress has been made in understanding the interplay between factors such as molecular dimensions, intermolecular interactions, and composition of the condensate in determining the partitioning and localization of small molecules into the dense phase. Along with experiments and AI/ML methods, computer simulations have been instrumental in highlighting the importance of the condensate micro-environment in determining the thermodynamics of small molecule partitioning (Fig. 3). Current challenges associated with the application of machine learning/artificial intelligence-based approaches for identifying c-mods with optimal partitioning properties As highlighted in sections 3.1 and 4, predictive models based on modern AI/ML methods which are trained on high-quality data obtained from established phenotypic screening assays are a highly promising tool to predict which of the physicochemical features of a c-mod, such as aqueous solubility, lipophilicity, hydrogen-bond donors/acceptors, and presence of specific chemical groups may be most critical for determining its selective partitioning into a disease-relevant condensate. However, several challenges persist at present with regard to the construction of AI/ML-based models for predicting the partitioning behavior of potential c-mods, as described in the sub-sections below. Imbalanced data sets and experimental noise Recent work suggests that a key feature in determining the partitioning of small molecules into condensates is the octanol-water partition coefficient (logP)[80]. However, a major obstacle in using logP to predict small molecule partitioning favorability is the imbalanced nature of available datasets. Existing datasets are predominantly comprised of molecules that have extreme partitioning values, while true positive examples of molecules that enrich within condensates moderately are comparatively scarce. This makes it difficult to learn the chemical features that truly drive favorable partitioning into a specific condensate. Therefore, the models tend to perform worse in the moderate to high partitioning regime that is most relevant for drug discovery, since high logP values are underrepresented. For generative approaches, the absence of prior examples of strongly partitioning molecules (i.e., reference points in training data) may result in a high failure rate in proposing novel molecules with high enrichment potential or produce candidates that appear promising in silico but pose considerable challenges for chemical synthesis[98]. A second challenge relates to the ambiguity near the partitioning boundary, as a large fraction of logP values are tightly clustered around zero. Therefore, small variations in experimental conditions, or even minor measurement noise, can shift these molecules to either side of the threshold, leading to complications in both model training and evaluation. Such near-boundary cases introduce instability in ML models, classifiers may flip predictions based on slight perturbations, while regression models may report low numerical error for small absolute differences in logP. As a result, they may show deceptively high overall accuracy while failing to capture the sparse positive cases that are of the greatest biological and therapeutic interest[99,100]. Addressing this issue requires models that can explicitly incorporate uncertainty and probabilistic boundaries, rather than treating partitioning as a sharp binary outcome[101,102]. Closely linked to this issue is the lack of a clear understanding of the underlying molecular rules that govern small molecule partitioning. While empirical measurements such as lipophilicity, hydrophobicity, and aqueous solubility are generally recognized as important factors in condensate partitioning, there are many cases where molecules with very similar values for these properties display drastically different partitioning behavior. Likewise, structurally similar compounds can show opposite outcomes, with one enriching strongly in condensates and the other failing to partition at all. From a modeling perspective, this gap forces algorithms to infer rules from incomplete signals, often leading to overfitting or reliance on correlations that do not hold outside the training set. The result is a lack of confidence in AI-based predictions and a persistent difficulty in generalizing to novel chemical scaffolds. Another important issue is that molecules which exhibit very strong partitioning often do so because they possess extreme physicochemical properties, such as very high logP that make them unsuitable as drug-like compounds due to poor water or blood solubility[103]. Designing compounds that combine both adequate drug-like properties and strong partitioning therefore represents a major challenge. Modeling and evaluation In addition to the above-highlighted issues with the current data sets, the design of model training and evaluation strategies introduces unique challenges. One critical issue is that condition-aware data splits are rarely enforced. Many molecules are tested under multiple experimental conditions, such as different salt concentrations or crowding environments. If molecules measured under one set of conditions are included in the training set, while the same molecules under slightly different conditions are included in the test set, this creates information leakage. The model may appear to generalize well, but it is memorizing molecular identities rather than learning condition-specific rules. Such leakage inflates reported performance and give a misleading picture of true predictive capacity. Developing splitting strategies that explicitly control for condition overlap is essential for generating realistic benchmarks of model capability[104–106]. A further obstacle is the absence of standardized benchmarks. Unlike more established areas of drug discovery, where curated datasets and agreed-upon splits, such as those used for kinase inhibitors,[107,108] or protein-ligand binding,[109] enable rigorous comparison, condensate prediction lacks any common reference point. Each group develops its own assays, curates its own datasets, and applies its own evaluation metrics. As a result, reported model performances are not directly comparable across studies, and progress is difficult to measure. Without shared benchmarks, it is impossible to distinguish genuine methodological advances from differences in dataset quality or task framing. Establishing community datasets with standardized evaluation protocols will be essential for accelerating progress and avoiding fragmentation of the field. Conclusions and future outlook for c-mod discovery and development The development of drug modalities that specifically target the microenvironment of BCs is challenging. The desired action of these molecules based on phenotypic outcomes can range from dissolving the condensate, recruiting or secluding certain components, altering its material properties or, in some cases facilitating condensate formation itself. Recent studies[79–81] have begun to elucidate the links between small molecule chemistries and their partitioning into different in vitro BCs, pointing to physical features such as hydrophobicity and chemical features such as aromaticity being of importance for condensate partitioning. However, given the relatively similar compositional biases of low complexity IDRs present within multi-domain proteins that form the “scaffold” of numerous BCs, questions remain as to whether these condensates can selectively sequester small molecules with a particular chemistry to minimize off-target activity in vivo. Collectively, this highlights that there is a need for a precise understanding of how small molecule chemistries along with protein sequence, encode the partitioning of small molecules and their subsequent modes of action. With regard to condensate microenvironment, an additional consequence of hydration within the condensate, is an increased dielectric constant as compared to the hydrophobic core of proteins[90,110]. The value of the dielectric constant, and hence the micropolarity of the medium, can also be sequence-dependent[87]. Changes in the dielectric constant can lead to different electrostatic environments within the condensate, which can have a marked effect on the predicted binding of molecules to protein scaffolds. From these studies, a question that emerges is whether the condensate microenvironment can be modeled as a hydrophobic medium such as octanol or cyclohexane[89,90]. Therefore, to what extent, or in which sequence contexts, features such as logP are applicable as a valid metric to screen small molecules for condensate partitioning remains unclear. An emerging area of interest in the context of the condensate microenvironment are the pH gradients and interphase electric potentials generated by condensates[96,111, 112]. The implications for the partitioning of small molecules are changes in charge state of the molecule within the condensate compared to the dilute phase, as well as an interfacial transport barrier due to interphase potentials. While the modeling of these effects in physics-based molecular simulations is currently in its infancy, recent simulation studies have begun to elucidate the nature of the interphase potentials, as well as their effect on the partitioning of charged molecules[82,89,113,114]. We anticipate that advances in the investigations of the condensate microenvironment and its sequence-dependence will provide a robust physics-based framework for understanding the thermodynamic driving forces underlying the partitioning of small molecules. A major limitation regarding the effective adoption of ML-based QSAR models for predicting c-mod efficacy within condensates is the lack of activity data beyond partitioning itself. For most datasets, the only available label is whether a molecule partitions, with little or no annotation of downstream biochemical or cellular effects. Although enrichment within condensates is necessary for function, it does not ensure productive engagement with the intended target or a beneficial cellular outcome. A molecule may enrich strongly yet fail to modulate the relevant process or may perturb material properties in harmful ways. This disconnect slows rational design and prevents models trained only on partitioning from supporting drug discovery pipelines where functional modulation is the endpoint. Without paired datasets that link enrichment to activity, the field risks treating enrichment as a proxy for efficacy. The scarcity of c-mod activity data also shapes objective design. A clear limitation is the lack of an explicit task to predict the biological effect of a molecule once it has partitioned, together with safeguards against detrimental changes in condensate properties. To avoid collapse to intracondensate enrichment alone, models should produce multidimensional outputs that capture enrichment strength, functional potential, and drug-likeness. Additionally, conflicts with solubility and pharmacokinetics further constrain the design space, which underscores the need to balance efficacy, safety, and developability in the learning objectives. For example, molecules with high lipophilicity may display strong enrichment yet suffer from poor aqueous solubility, limited absorption, suboptimal exposure at the target site, and unfavorable clearance profiles which collectively undermine therapeutic viability despite attractive partitioning. To effectively address the above highlighted challenges, a careful integration of simulation, AI/ML, and experiment in an iterative loop (Fig. 4) is required. We envision that within the loop, atomistic simulations of condensates will provide an initial characterization of the condensate microenvironment which will guide the feature selection and interpretation of feature importance in AI/ML models trained to predict partitioning. Through screening of large databases using these models, practitioners can identify candidate molecules for validation in high content screening experiments which can then be used to filter the chemical space for c-mods belonging to various therapeutic modalities. Through the iterative discovery, testing and validation of candidate molecules, we anticipate the development of heuristics for the design and development of c-mods to target BCs with high specificity and efficacy coupled with a favorable drug profile.