MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery

MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery View ORCID Profile Qiaosi Tang , Daoyun Ding , View ORCID Profile Xiaoyong Yuan , View ORCID Profile Gustavo Seabra , Peter A Ramdhan , Chi-Yuan Liu , My T. Thai , Chenglong Li , View ORCID Profile Hendrik Luesch , View ORCID Profile Yanjun Li doi: https://doi.org/10.1101/2025.07.11.664424 Qiaosi Tang 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Qiaosi Tang Daoyun Ding 2 School of Computer, Data & Information Sciences, University of Wisconsin-Madison , Madison, WI 53706, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Xiaoyong Yuan 3 Holcombe Department of Electrical and Computer Engineering, Clemson University , Clemson, SC 29634, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Xiaoyong Yuan Gustavo Seabra 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Gustavo Seabra Peter A Ramdhan 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Chi-Yuan Liu 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site My T. Thai 4 Department of Computer & Information Science & Engineering, University of Florida , Gainesville, FL 32611, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Chenglong Li 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Hendrik Luesch 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA 5 Program in Cancer and Stem Cell Biology, Duke-NUS Medical School , Singapore 169857, Singapore Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Hendrik Luesch Yanjun Li 1 Department of Medicinal Chemistry, Center for Natural Products, Drug Discovery and Development, University of Florida , Gainesville, FL 32610, USA 4 Department of Computer & Information Science & Engineering, University of Florida , Gainesville, FL 32611, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Yanjun Li For correspondence: yanjun.li{at}ufl.edu Abstract Full Text Info/History Metrics Preview PDF Abstract Designing novel molecules with desired bioactivity remains a fundamental challenge in drug discovery. Most molecular design methods follow target-based drug discovery paradigms that rely on well-defined drug targets, thereby limiting their applicability to diseases lacking known targets or reference compounds. Here we introduce Morphology-Guided Molecule Generation (MGMG), a phenotypic drug discovery–oriented approach that integrates cellular morphological profiles from compound treatments with molecular textual descriptions without requiring molecular target information. Cell morphology offers the guidance on desired bioactivity-relevant phenotypic effects, while textual descriptions provide direct and interpretable cues about molecular structure. Leveraging complementary structural and bioactivity context, MGMG significantly enhances molecule generation performance, especially in scenarios where textual descriptions are under-informative or morphological signals are weak. MGMG can also be applied to genetic perturbations, enabling activator design from gene overexpression morphology without requiring knowledge of reference compound structure. In addition, in silico docking demonstrates that MGMG-generated molecules, despite lacking target information, exhibit binding affinities comparable to reference compounds, preserving key interactions while introducing structural diversity. Overall, MGMG jointly utilizes morphological and textual description inputs to guide molecule generation, enabling diverse, bioactivity-aware compound design in a target-agnostic fashion. Introduction Drug discovery is a challenging and time-consuming process; modern artificial intelligence-aided drug design (AIDD) methods aim to accelerate it by enabling the automated identification of drug candidates 1 . Most AI-based molecular design methods adhere to the principles of target-based drug discovery (TDD), relying on well-defined biological molecular targets and binding pockets for structure-based drug design 2 – 6 , or on known active molecule binders as reference compounds for ligand-based molecular generation 7 – 9 . Consequently, these methods have limited applicability for diseases lacking validated targets or existing reference compounds. In contrast to TDD, phenotypic drug discovery (PDD) does not depend on knowledge of specific drug targets or hypotheses about their roles in diseases. Instead, it relies on the chemical interrogation of disease-relevant biological systems in a molecular target-agnostic fashion and has successfully identified many of the first-in-class drugs approved by the US Food and Drug Administration (FDA) over the past decades 10 , 11 . High-content imaging (HCI), enabled by automated microscopy and image analysis, has paved the way for large-scale, high-throughput drug screening in PDD 12 . For example, the Cell Painting assay 13 , 14 , a prominent and low-cost HCI technique, provides a comprehensive view of the cellular state by staining major cellular components and producing morphological profiles as phenotypic descriptors. These morphological profiles have led to a comprehensive understanding of cellular behavior under chemical perturbations 13 . The target-agnostic nature and the ability to provide significant insights into biological responses make morphological profiling well-suitable for addressing the challenges faced by TDD-oriented de novo molecule design methods. While promising, only two previous studies have explored morphological profiling in the task of molecule design 15 , 16 , and both exhibit several limitations. First, because many different molecules can induce similar morphological effects, relying solely on cell morphology data for molecule generation often lacks specificity and control over the desired molecular structure or scaffold. Particularly in PDD, where drug target information is not available as a generation condition, the absence of additional guidance beyond the morphological data can lead to the omission of essential molecular properties, such as synthesizability and drug-likeness 15 , 16 , limiting the applicability of the generated molecules. Furthermore, generating molecules based on cell images from weak or inactive perturbations, which induce only subtle or minimal morphological changes, poses challenges in extracting clear, actionable signals for molecule generation 15 . In this work, we introduce MGMG ( M orphology- G uided M olecule G eneration), an AI-driven molecule design approach tailored for PDD, which addresses these challenges by enabling morphology-informed bioactivity awareness and structural and property control within a unified framework. To achieve this, we propose integrating cellular morphological data from compound treatments with molecular textual descriptions to jointly guide molecule design, without requiring molecular target information. Complementary to morphology, molecular textual descriptions offer a high degree of interpretability and have proven effective in conveying open-ended chemical information about molecular structures and properties across diverse drug discovery tasks 17 – 20 . Meanwhile, in medicinal chemistry, the availability of domain knowledge and heuristic “rules of thumb” are essential factors guiding real-world molecule discovery campaigns. Thus, driven by its unique advantages, we leverage molecular textual descriptions to embed domain knowledge of the desired structural and physicochemical attributes, complementing the guidance provided by morphological data. Through pretraining and contrastive learning paradigms, MGMG bridges the two modalities and their guidance signals by aligning them into a unified latent space and capturing their interactions. This combination enables molecule design with simultaneous bioactivity awareness, specificity, and structural control, overcoming the limitations of methods that rely solely on cell morphology. In parallel, unlike molecule captioning and text-based molecule generation approaches, our method provides explicit guidance on molecular biological effects and mechanisms of action, encoded through cellular morphological profiles. These aspects are challenging to capture precisely and thoroughly in textual descriptions, whereas cell morphology offers comprehensive, low-level phenotypic representations that are crucial for drug discovery and design. Through experiments, we demonstrate that the proposed approach effectively enhances specificity and control in molecule generation, achieving a higher recovery rate while maintaining a degree of structural diversity. We further show that the two input modalities compensate for each other, with morphological profiles improving molecule generation when textual descriptions are under-informative, and textual descriptions enhancing performance when morphological signals are weak. Our model can be extended from chemical to genetic perturbations, enabling activator design based on morphological profiles from gene overexpression, without requiring structural guidance from reference activators, even in cases of weak perturbations. We also explore whether the molecules generated in this target-agnostic manner can potentially bind to known protein targets. Our results show that the generated molecules can retain key interactions with target protein residues despite the input lacking explicit structural knowledge of the target. In some cases, structural variations introduced by MGMG enable the molecules to explore the binding pocket and access novel chemical space. In summary, we present MGMG as a versatile framework for the design of diverse and high-quality molecules with bioactivity context in a disease target-agnostic fashion. Results Overview of MGMG We introduce MGMG, a PDD-oriented small molecule generation framework that leverages both desired cellular morphological profiles and molecular textual descriptions to jointly guide the design of novel compounds ( Fig. 1a ). Cell morphology serves as a functional readout of chemical perturbations, capturing the biological impact of a designed compound at the phenotypic level. In contrast, textual descriptions convey structural and physicochemical information, reflecting the compound’s properties at the molecular level. MGMG integrates these complementary modalities to enable molecule generation that is both biologically informed and structurally guided. MGMG is built on an encoder-decoder Transformer architecture ( Fig. 1b ). The encoder comprises two branches: a cell morphology branch that processes the cellular morphological profile induced by the reference compound, and a textual description branch that extracts semantic representations from the reference compound’s textual description. The disentangled design allows MGMG to integrate the powerful existing models pretrained on a single modality. Specifically, we utilize the BioT5 model 21 pretrained on the large-scale text-molecule pairs as the base model for the text branch encoder to bridge the textual description associations and chemical structure. The morphology branch uses a multi-layer perceptron network to encode morphological profiles extracted by the bioimaging software CellProfiler 22 , which generates predefined biologically insightful features from raw Cell Painting images. These profiles have been successfully and widely used for characterizing compound mechanisms of action in phenotypic drug discovery 23 . To establish a unified representation space between these two modalities, MGMG employs the contrastive learning strategy 24 to align the morphology branch to the pre-trained text branch ( Fig. 1c ). After the alignment, the dual-modality inputs are integrated using an early fusion approach to generate a joint encoder representation, upon which a decoder autoregressively generates novel molecules in SELFIES format 25 . To train MGMG, we first curated a dataset named Mol-Instructions-BBBC036v1, a structure–text–morphology dataset comprising 25,070 compounds, by pairing morphological profiles from BBBC036v1 26 with molecular textual descriptions from PubChem 20 , 27 using the unique compound structures as the linking key. However, many of the original textual descriptions were under-informative with numerous distinct molecules sharing the same brief annotations, limiting their utility in molecule generation tasks. To enhance the quality and specificity of these descriptions, we employed the pre-trained molecular captioning model BioT5 21 to generate synthetic, detailed textual descriptions for each compound. The resulting dataset, MolCaptioned-BBBC036v1, was first used for multimodal alignment between the morphology and text branch encoders, followed by end-to-end encoder-decoder training for the entire MGMG molecule generation framework. Given the relatively limited dataset size, the Low-Rank Adaptation of Large Language Models (LoRA) was adopted to mitigate overfitting 28 . Download figure Open in new tab Fig. 1: Schematic of MGMG. a The overarching goal of MGMG is to design novel small molecules guided by cell morphology– informed bioactivity signals and molecular textual descriptions of a reference compound, without requiring prior knowledge of the molecular target. b Within an encoder-decoder Transformer framework, the complementary modalities are integrated using an early fusion approach and jointly guide molecule generation in the SELFIES format. c The contrastive learning strategy is utilized to align cell morphology profiles and textual descriptions to establish a unified latent representation space. MGMG generates molecules with desired chemical properties To our best knowledge, MGMG is the first approach to integrate both morphological and textual descriptions for de novo molecule design. Accordingly, we compared MGMG with the state-of-the-art unimodal baselines to assess its performance. Among the two existing morphology-based approaches, we compared MGMG to CPMolGAN, a Generative Adversarial Networks (GAN)-based model that generates molecules using morphological profiles as input 15 . We did not include the GFlowNet-based model in our comparison, as it requires training a separate model for each target compound, which does not scale to the size of our training dataset or align with our evaluation setup 16 . In addition to morphology-based models, we also compared MGMG with text-based molecule generation models, such as MolT5 29 and BioT5 21 . Notably, MGMG is a light model with only 14 million trainable parameters—just 5% of the model size of BioT5 or MolT5-base ( Supplementary Fig. 1 ). Because molecules can be represented as biosequence-like structures (e.g., SMILES or SELFIES), we evaluated the generated outputs using natural language processing metrics such as BLEU 30 and Levenshtein distance 31 to measure the similarity between generated and reference SELFIES or SMILES. Results in Table 1 demonstrate that MGMG outperforms all baseline methods on both metrics. In addition, both MGMG and BioT5 achieve 100% validity in the generated molecules, underscoring the robustness and error-tolerance of using SELFIES representations for molecular generation. Generating molecules with desired properties is crucial for drug discovery. Herein, we evaluated the molecules generated by MGMG and baseline methods based on a wide range of physicochemical, drug-likeness and synthesizability properties, even though these attributes were not monitored or optimized during the training process. Most molecules generated by MGMG adhere to the “Lipinski’s Rule of Five” 32 , with an average Molecular Weight (MW) <500 Da, Lipophilicity (LogP) <5, hydrogen bond acceptors (HBA) <10, and hydrogen bond donors (HBD) <5 ( Table 1 ). We observed a high correlation between the physicochemical properties of MGMG-generated molecules and their corresponding reference compounds ( Fig. 2a , Table 1 ), further demonstrating the model’s capability to accurately recover biologically relevant molecular structures. Supplementary Fig. 2 presents comparative examples of molecules generated by MGMG, BioT5, MolT5, and CPMolGAN. Overall, MGMG and BioT5 produced molecules structurally similar to the reference compound, with MGMG showing greater similarity. Molecules generated by MolT5 and CPMolGAN diverged more substantially from the reference compound. These observations were further supported by quantitative evaluations of drug-likeness and structural realism. MGMG generated molecules with the highest average drug-likeness (QED) scores among all models as well as the highest percentage of molecules exceeding a QED threshold of 0.5, indicating strong alignment with drug-likeness criteria. Moreover, the QED scores of MGMG-generated molecules showed the strongest correlation with those of the reference compounds, further highlighting its ability to preserve key drug-likeness attributes. In terms of structural realism, MGMG-generated molecules achieved the lowest Fréchet ChemNet Distance (FCD) scores 33 , indicating a closer resemblance to the distribution of known bioactive compounds. In addition to drug-likeness, we also examined the synthetic accessibility (SA) score, which estimates the ease of synthesis 34 . MGMG achieved the lowest average SA score and the highest percentage of molecules with an SA score below 4.5. It also exhibited the strongest correlation between the SA scores of generated molecules and their reference counterparts. These findings indicate that MGMG generates molecules with greater synthesizability than those produced by other models. Collectively, the strong performance across these criteria highlights MGMG’s ability to generate novel molecules with desired chemical properties. Detailed descriptions of these evaluation metrics are provided in the Methods section. Download figure Open in new tab Fig. 2: High Accuracy, Efficacy and Diversity in Molecule Generation by MGMG. a Properties of MGMG-generated molecules exhibit high correlation with their respective reference compounds. b An example demonstrating the diversity of MGMG-generated molecules: the top five molecules generated based on the textual description and morphological profile of a reference compound are shown, with the top-ranked molecule achieving the highest Tanimoto similarity to the reference compound. c Correlation between generated and reference compound properties across multiple chemical metrics. MGMG outperforms the unimodal baseline BioT5 on both datasets, with particularly pronounced improvements under limited-text conditions (Mol-Instructions-BBBC036v1). d t-SNE visualization of the molecules with weak morphological impact. e Correlation between generated and reference molecule properties with weak morphological impact. MGMG outperforms CPMolGAN across all evaluated chemical properties, with greater performance observed when detailed textual descriptions (MolCaptioned-BBBC036v1) were used. f Performance comparison on MolCaptioned-BBBC036v1 and Mol-Instructions-BBBC036v1 datasets with different ablation settings (y-axis). The x-axis shows the evaluation metrics: BLEU scores, Levenshtein scores, and FCD scores. MGMG achieves state-of-the-art (SOTA) performance in molecule generation across all metrics. All asterisks indicate statistical significance from student’s t-test: p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), p < 0.0001 (****) and not significant (ns). View this table: View inline View popup Download powerpoint Table 1. Properties of molecules generated by different models on MolCaptioned-BBBC036v1. MGMG generates molecules with diverse structures Analysis across all generated outputs revealed that MGMG produced a higher percentage of unique molecules and exhibited greater Murcko scaffold diversity compared to baseline models ( Table 1 ). These results suggest that MGMG generates structurally diverse molecules while preserving essential chemical properties. To further illustrate this capability, we analyzed the top five molecules generated from a single input consisting of a textual description and a morphological profile. The generated molecules were ranked based on their Morgan fingerprint similarity to the corresponding reference molecule ( Fig. 2b ). We observed that all five generated molecules shared a common structural core and functional groups with the reference molecule. Despite minor difference in stereochemistry, Sample-1 matched the reference molecule with a Tanimoto similarity score of 1.00, demonstrating MGMG’s ability to accurately recover reference molecular structures. Sample-2, 3, and 4 each achieved a Tanimoto similarity score of 0.83. These molecules, which are stereoisomers of one another, introduced a minor variation in the placement of the carbonyl group, while retaining the overall scaffold and functional motifs. Sample-5, with a lower Tanimoto similarity score of 0.63, incorporated more substantial structural changes, including the substitution of the pyridine ring with a phenyl ring and the introduction of a branched alkyl chain in contrast to the linear two-carbon chain in the reference, while still maintaining the general chemical framework observed in the higher-ranked samples. These findings indicate that MGMG maintains essential molecular scaffolds while incorporating structural diversity. Such variation is valuable for exploring novel analogs around a known active compound. Additional examples of top five generated molecules from various input conditions are provided in Supplementary Fig. 3 . Cellular morphological information improves generated molecule properties We next evaluated the contribution of incorporating morphological profiles to the properties of generated molecules. Molecules were generated under two input configurations: (1) combined molecular textual descriptions and morphological profiles, and (2) textual descriptions alone ( Supplementary Table 1 ). To challenge the model in conditions where limited chemical structural information is available, we used molecular descriptions from Mol-Instructions-BBBC036v1 20 , a dataset containing under-informative, non-unique textual descriptions sourced from PubChem. Compared to the detailed captions in the MolCaptioned-BBBC036v1 dataset, these under-informative descriptions provide substantially less structural detail ( Supplementary Fig. 4 ). This setup more accurately reflects real-world drug design scenarios, where prior knowledge of the specific structures and properties of the molecule to be designed is often limited. To assess the contribution of morphological information, we compared the outputs of MGMG with BioT5, a unimodal baseline model trained on the same molecular descriptions but without morphological input. We evaluated the correlation between generated molecules and their corresponding reference compounds across multiple chemical properties, including QED, SA, MW, LogP, HBA, and HBD. MGMG consistently outperforms BioT5 across both datasets, with particularly pronounced improvements in the Mol-Instructions-BBBC036v1 setting, where textual guidance is minimal. Notably, MGMG achieved a 282.039% increase in QED correlation and a 152.084% increase in SA correlation over BioT5 on Mol-Instructions-BBBC036v1, compared to more modest gains of 8.600% and 5.507%, respectively, on MolCaptioned-BBBC036v1 ( Fig. 2c ). This is expected, as the inclusion of rich phenotypic information becomes especially valuable when textual inputs lack detailed cues, whereas in the MolCaptioned-BBBC036v1 dataset, textual descriptions already provide informative guidance for molecular generation. Similar trends were observed for MW, LogP, HBA, and HBD ( Fig. 2c ), further highlighting the benefits of integrating morphological input. These findings indicate that morphological information enhances the drug-likeness and the synthesizability properties of generated molecules, especially when the available textual descriptions lack specificity. Textual information enhances molecule generation when morphological signals are weak While the addition of morphological information has been shown to improve molecule generation when textual descriptions are limited, it is equally important to assess whether incorporating textual descriptions can compensate in scenarios where morphological signals are weak. As previously discussed, weak perturbations often induce only subtle morphological changes, making it challenging to extract clear, actionable signals for molecule generation 15 . This limitation is especially relevant for models that rely exclusively on morphological input, such as CPMolGAN 15 . In such cases, their ability to generate meaningful compounds from weak morphological signals is limited 15 . In contrast, MGMG incorporates both morphological and textual descriptions, enabling it to compensate for weak phenotypic signals by leveraging textual information, thereby effectively enhancing the specificity and control in the molecule generation process. This dual-modality design allows MGMG to maintain generation performance even when morphological profiles are minimally informative. To test this hypothesis, we quantified the strength of morphological perturbation by computing the Euclidean distance between each treatment profile and the DMSO control—a commonly used phenotypically inert reference. Compounds were then ranked in ascending order by their distance to DMSO, and the bottom 5% (n = 130) with the weakest morphological signals were selected for analysis ( Fig. 2d ). On this subset, MGMG substantially outperformed CPMolGAN, achieving significantly higher correlations between generated molecules and their reference counterparts across key chemical properties, such as QED, SA, MW, LogP, HBA, and HBD ( Fig. 2e ). These performance gains were especially pronounced when detailed textual descriptions (MolCaptioned-BBBC036v1) were provided, as expected, compared to the under-informative textual descriptions from the Mol-Instructions-BBBC036v1 dataset ( Fig. 2e ). These findings underscore the importance of integrating textual descriptions, especially when phenotypic signals are weak. By leveraging both modalities, MGMG demonstrates greater robustness than morphology-only approaches, offering a more reliable strategy for molecule generation regardless of perturbation strengths. Ablation studies for MGMG To validate the design of MGMG, we conducted a series of ablation studies using both the MolCaptioned-BBBC036v1 and Mol-Instructions-BBBC036v1 datasets ( Fig. 2f , Supplementary Table 1 ). Our first set of experiments examined the effect of incorporating morphological profiles on molecule generation performance. Using identical input text, we compared two input configurations: (1) text-only generation and (2) multimodal generation with both morphological profiles and textual descriptions. Across both datasets, the inclusion of morphological input consistently improves reconstruction performance. On MolCaptioned-BBBC036v1, multimodal generation resulted in a 0.998% improvement in BLEU score and a 6.397% reduction in Levenshtein distance compared to text-only generation. On Mol-Instructions-BBBC036v1, the gains were more substantial, with BLEU increasing by 20.583% and Levenshtein distance decreasing by 9.352% ( Fig. 2f , Supplementary Table 1 ). We further evaluated correlations between generated and reference molecules across key chemical properties. Drug-likeness improved with multimodal generation, as shown by the reductions in FCD: 2.546 vs. 2.678 on MolCaptioned-BBBC036v1 (−4.932%) and 12.835 vs. 26.112 on Mol-Instructions-BBBC036v1 (−50.846%) ( Fig. 2f , Supplementary Table 1 ). MGMG-generated molecules also showed significantly higher QED correlations under the multimodal generation setting for both datasets. Notably, on Mol-Instructions-BBBC036v1, the incorporation of morphological profiles led to a 111.347% increase in QED correlation relative to the text-only baseline, while the gain on MolCaptioned-BBBC036v1 was more modest at 7.567% ( Supplementary Table 1 ). A similar pattern was observed for SA, with correlation increases of 96.805% on Mol-Instructions-BBBC036v1 and 5.560% on MolCaptioned-BBBC036v1 ( Supplementary Table 1 ). Consistent improvements were also observed across MW, LogP, HBA, and HBD, reinforcing the contribution of morphological profiles to molecule generation, particularly when textual descriptions are under-informative. In a second ablation study, we investigated the role of modality alignment in multimodal generation. Specifically, we evaluated whether aligning the morphology and text encoders via contrastive learning would improve performance. The model with aligned encoders outperformed the version without alignment across both datasets. For instance, on MolCaptioned-BBBC036v1, the MGMG model with modality alignment yielded a 0.515% improvement in BLEU score, a 3.231% reduction in Levenshtein distance, and a 2.875% reduction in FCD ( Fig. 2f ). These improvements were more pronounced on Mol-Instructions-BBBC036v1, where BLEU score increased by 23.960%, Levenshtein distance decreased by 9.905%, and FCD dropped by 32.796% ( Fig. 2f ). These results underscore the importance of projecting both modalities into a unified latent space to maximize the effectiveness of multimodal conditioning in molecule generation. MGMG enables activator design with genetic perturbations Given the success of MGMG in generating molecules guided by compound-perturbed morphological profiles, we next investigated whether the model could be applied to gene overexpression morphological profiles to design activators. We designed this case study not only to assess the model’s generalizability but also to challenge it to generate functionally relevant molecules without any prior knowledge of binding target or reference compound structure. Specifically, we applied MGMG to nine genes: NFKB1, BRCA1, HSPA5, TP53, CREBBP, STAT1, STAT3, HIF1A, and NFKBIA, using gene overexpression image profiles from the BBBC037v1 morphological profiling dataset 35 . According to prior studies, three of the nine genes—NFKB1, BRCA1, and HSPA5— induce strong morphological changes, while the remaining genes show weaker differences compared to empty vector controls 15 . For the textual descriptions, we used the GPT-4.0 36 to generate the functional descriptions of desired molecular effects for each gene, explicitly excluding any structural or scaffold-level hints, thereby mimicking a realistic early-stage drug design scenario where functional outcomes are defined but direct knowledge of reference compound is unavailable ( Fig. 3a,b , Supplementary Table 2 ). To evaluate the quality and relevance of the generated molecules, we compared their scaffolds to known activators retrieved from the ExCAPE dataset 37 , assuming that gene overexpression can phenocopy activator activity ( Fig. 3a ). All known activators were confirmed to be absent from the MGMG training set (BBBC036v1), ensuring that scaffold recovery was not due to data leakage or memorization. Among the nine genes tested, MGMG successfully recovered ExCAPE activator scaffolds for five genes, outperforming CPMolGAN, which recovered activators for two 15 ( Fig. 3c ). Notably, MGMG identified enriched scaffolds not only for genes with strong morphological signals such as NFKB1, BRCA1, and HSPA5 ( Fig. 3d ), but also for genes with weaker morphological impact, such as STAT3 and TP53 ( Fig. 3e ). For STAT3, two generated scaffolds matched potent activators in the ExCAPE dataset; for TP53, ten generated scaffolds matched known activators ( Supplementary Table 2 ). For example, the matched STAT3 activators (PubChem AID:1398; CID:483833 and CID:3797106) are reported to activate the STAT3 pathway in U3A cells with IC50 values of 2.06 µM and 6.19 µM, respectively—highlighting the potential functional relevance of the MGMG-generated compounds. These findings demonstrate that MGMG can be applied beyond chemical perturbations to genetic perturbation profiles and design molecules that resemble known activators even in the absence of their structural information. This capability underscores MGMG’s unique target-agnostic potential, enabling molecule generation based solely on phenotypic outcomes and functional goals. This target-agnostic approach opens new possibilities for molecule design, particularly in cases where structure-activity relationships are unknown or where conventional target-based approaches are limited. Download figure Open in new tab Fig. 3: Molecules can be generated with gene overexpression morphological profiles and exhibit structural features resembling known activators. a Schematic overview of the activator design pipeline using MGMG. b An example input pair showing TP53 gene overexpression image and textual description detailing the desired properties for the activators to be designed. c Number of enriched scaffolds as in ExCAPE activators and number of corresponding molecules generated per gene. Representative generated molecules from genes with d strong and e weak morphological impacts, respectively. Molecular docking of MGMG-generated molecules to potential protein targets Although MGMG is a target-agnostic molecule design approach, here we investigated whether the designed molecules could bind to potential targets when the reference compounds’ targets are available. For this case study, we docked the molecules generated by MGMG to the protein targets of the reference compounds, where the corresponding protein structures were fetched from the Protein Data Bank (PDB), and then we compared the docking results to those of the reference molecules. To ensure adequate structural diversity, the top five designed molecules were selected for docking. Docking scores were calculated by Glide SP 38 , based on the co-crystallized or docking-generated poses. In the first case ( Fig. 4a ), the reference molecule BRD-K85883481-001-02-6 was docked to its receptor, human liver 5β-reductase (AKR1D1) (PDB ID: 3CMF), yielding a docking score of -8.81 kcal/mol. The generated Sample-1 molecule retained the same steroid-like core structure as the reference molecule and only replaced the ketone on the 11-position by a hydroxyl. In contrast, the Sample-2 molecule demonstrated greater structural divergence, replacing the ketone on the 11-position with a hydroxyl group and removing the alpha hydroxyl group on the 17-position. Despite these differences, both the Sample-1 and Sample-2 molecules successfully retained the reference molecule’s key interactions with the target protein, including hydrogen bonds to SER169 and TYR58. These preserved interactions stabilized their binding within the protein pocket, resulting in docking scores of -8.46 kcal/mol for the Sample-1 molecule and -8.50 kcal/mol for the Sample-2 molecule ( Fig. 4a ). These results demonstrate MGMG’s capacity to generate structurally diverse molecules with similarly promising docking scores, highlighting its potential to explore novel chemical space. In another case ( Fig. 4b ), the reference molecule BRD-K55468218-001-04-8 was docked to the human dopamine D2 receptor (PDB ID: 7DFP), with a docking score of -10.00 kcal/mol. Compared to the reference molecule, Sample-1 and Sample-2 molecules both included an additional fluorobenzene ring, differing only in the fluorination position. This structural extension enabled the generated compounds to reach subpockets not accessed by the reference molecule. Despite these structural differences, both designs preserved the reference compound’s key interactions with the target protein, including the salt bridge and hydrogen bond between the positively charged nitrogen and ASP114, as well as the Pi-Pi stacking interactions with TRP386 and PHE390. These interactions resulted in docking scores of -9.88 kcal/mol and -10.51 kcal/mol, respectively. This case highlights MGMG’s potential for scaffold expansion to improve target engagement and explore new binding regions. Additional cases are presented in Supplementary Fig. 5 , further demonstrating the capability of MGMG to produce diverse molecules that not only maintain comparable binding affinities to the reference molecules but also introduce meaningful structural variations. Collectively, these findings show that molecules generated by MGMG, despite lacking explicit knowledge of the target protein, can retain the critical interactions with key target residues, thereby ensuring stable binding. At the same time, the structural variations introduced by MGMG allow the molecules to explore novel chemical space and previously untapped regions of the binding pocket, supporting the dual goals of structural novelty and optimization in ligand design. Download figure Open in new tab Fig. 4: Molecular docking of MGMG-generated molecules to potential protein targets. The top panel presents the input textual descriptions and cell morphology profiles of reference molecules targeting: a human liver 5β-reductase (PDB ID: 3CMF) and b human dopamine D2 receptor (PDB ID: 7DFP). The middle panel displays molecular structures and three-dimensional docking results of the ligands in their best-scored binding poses within the respective binding pockets. Proteins are shown in a grey cartoon representation, while ligands are shown as stick models. Carbon atoms are colored green for reference molecules and pink or light blue for MGMG-generated molecules. Oxygen and nitrogen atoms are colored red and blue, respectively. The target protein residues involved in interactions are highlighted in dark blue. Interaction types are represented as follows: hydrogen bonds (yellow dashed lines), salt bridges (magenta dashed lines), and Pi-Pi interactions (cyan dashed lines). The bottom panel quantifies molecular similarity and drug-like properties, comparing MGMG-generated molecules to the reference molecules based on Tanimoto similarity, docking scores, QED, and SA. Discussions Developing novel molecules with favorable chemical properties and bioactivity remains a longstanding challenge in drug discovery. Most existing in silico molecule design approaches follow the principles of target-based drug discovery, relying on well-defined disease targets and therefore offering limited applicability for diseases without validated molecular targets. In this study, we present MGMG, a phenotypic drug discovery-oriented framework that integrates morphological information from compound treatment with molecular textual descriptions to enable de novo molecule design without the need for target information. A comprehensive benchmark study demonstrates that our approach outperforms state-of-the-art models that rely solely on either morphological information or textual descriptions. Detailed ablation study further confirms that both the incorporation of morphology data and the multimodal alignment via contrastive learning contribute to the model’s improved performance. To illustrate the practical applicability of MGMG beyond benchmark evaluations, we present a case study in activator design, demonstrating how this framework not only can be applied from chemical to genetic perturbations but also enables target-agnostic molecule generation without requiring any structural information from reference compounds, relying solely on morphological and functional cues. In addition, we show that, despite the absence of explicit target structural knowledge, MGMG-generated molecules can preserve key interactions with target residues while introducing structural variations to explore novel chemical space. In this study, we propose the integration of cell morphology and molecular textual description modalities for target-agonistic molecule generation. These two distinct modalities address different aspects of the molecule design task and effectively complement each other. Cell morphology data, such as from Cell Painting assays, provide a comprehensive overview of cellular responses to compound treatment, effectively reflecting compound bioactivity. Large public datasets, such as the JUMP-CP consortium 39 , have been released to provide cellular morphological profiles from both chemical and genetic perturbations 12 . Textual descriptions provide a direct and often interpretable characterization of molecule structure. Recent datasets derived from PubChem and ChEMBL 20 , 40 have significantly expanded the availability of text-molecule pairs; however, these descriptions typically offer limited insight into biological functions of the molecules. One of the key findings of our study is the complementary role that morphological and textual contexts play in molecule generation, particularly under limited-information scenarios. In real-world settings, it is common for one source of inputs to be under-informative—either from non-specific textual descriptions due to limited prior knowledge of the reference compound, or from weak phenotypic perturbations that produce minimal morphological changes. Our experiments show that morphological profiles significantly enhance generation performance when textual description lacks structural details, and that textual descriptions can meaningfully compensate for low-information morphological profiles from phenotypically inert compounds. This mutual compensatory effect addresses another key practical challenge in molecule generation, where consistent and informative guidance is lacking from either context alone. By integrating both morphological and textual information, MGMG offers a robust, target-agonistic strategy for designing bioactive molecules. Although MGMG has demonstrated strong potential, it still faces several limitations that warrant further investigation. First, the current implementation relies on morphological profiles extracted using CellProfiler 22 , which requires additional software setup and manual parameter tuning. Moreover, these profiles are limited to a predefined set of hand-engineered features, which may not fully capture the rich and nuanced information embedded in the raw cell images. Leveraging end-to-end deep learning approaches, such as convolutional neural networks 41 or vision transformers 12 , 42 applied directly to Cell Painting images, could enable the model to learn more unbiased, expressive and biologically relevant representations, potentially enhancing its generalizability and performance. Another limitation of MGMG lies in its reliance on morphological profiles derived from the single cell line U2OS, which may limit the model’s ability to generalize across diverse cellular contexts. Incorporating morphological data from multiple cell types in future studies would allow the model to better capture context-dependent phenotypic responses to compounds, leading to a more comprehensive and biologically informed molecular design framework 39 . Conventional de novo molecular design methods based on TDD often fall short in complex, target-agnostic settings, particularly for diseases without validated targets or known reference compounds 10 . Through leveraging both morphological profiles and textual descriptions, our PDD-oriented framework addresses this gap by capturing holistic biological responses and enhancing molecule design in a target-independent manner, opening new directions for future development. With its demonstrated effectiveness, MGMG holds strong promise for driving target-agnostic molecule generation and expanding the reach of phenotypic-based approaches. Methods Dataset construction Mol-Instructions-BBBC036v1 and MolCaptioned-BBBC036v1 datasets In this study, we curated two tri-modal datasets, which aligned the molecular structure (represented as SELFIES), the corresponding textual description, and the induced cellular morphological profile under compound treatment. Wherein, the MolCaptioned-BBBC036v1 dataset was used for both model training and evaluation, while the Mol-Instructions-BBBC036v1 dataset was used exclusively for evaluation. We started from the BBBC036v1 dataset from the Broad Bioimage Benchmark Collection, which includes morphological profiles of 30,616 compounds treated on the U2OS cells in replicates and imaged after Cell Painting assay 26 . The morphological profiles were processed with the Pycytominer pipeline for normalization, feature selection (reduction) and replicated well consensus 43 . From this pipeline, we reduced feature dimension to 753. To acquire the textual descriptions for these compounds, their canonical SMILES representations were utilized to match corresponding entries in the Mol-Instructions 20 datasets. A total of 25,070 compounds were identified to have both the morphological profiles from BBBC036v1 and textual descriptions from PubChem, forming the Mol-Instructions-BBBC036v1 dataset 20 . To improve the description quality, we used pre-trained molecular captioning model BioT5 to generate molecule captions for these 25,070 compounds, resulting in the creation of MolCaptioned-BBC036v1 dataset. We partitioned this tri-modal dataset into training, validation and testing sets with an 8:1:1 ratio. MGMG model and training details MGMG encoder The encoder of MGMG follows the transformer-based architecture of BioT5 21 , which is built on the T5-v1.1-base 44 configuration. The encoder consists of 12 transformer layers, each comprising a multi-head self-attention mechanism and a feed-forward network. Each attention block includes 12 attention heads, with key and value projection dimensions of 64, and the model-wide hidden state dimension is set to 768. The feed-forward network within each block consists of a dense layer of size 3072 followed by a ReLU activation and a second dense layer projecting back to 768 dimensions. A dropout probability of 0.1 is applied throughout for regularization. The output of the encoder is passed to the decoder for autoregressive molecule generation. Morphological prefix and early fusion strategy To integrate phenotypic information, the cellular morphological profile extracted from CellProfiler 22 is projected into a 768-dimensional embedding space. The morphological embedding is then prepended to the input text sequence as the morphological prefix, forming a unified input to the MGMG encoder. This early fusion strategy allows the Transformer to jointly attend over both morphological and textual representations from the very first layer of processing. MGMG decoder The decoder of MGMG is structurally consistent with the encoder and also follows the T5-v1.1-base 44 configuration. It comprises 12 decoder layers, each with multi-head self-attention, cross-attention (attending to encoder outputs), and a feed-forward network. Dropout of 0.1 is applied uniformly. The decoder takes as input the embedded representation of SELFIES tokens and performs autoregressive decoding. The final decoder output is projected back to the SELFIES vocabulary space using a linear layer and softmax activation function, producing one token at a time during inference. This setup allows MGMG to generate chemically valid molecules conditioned on both textual description and morphological context. Modality alignment with contrastive learning In the first stage of training, MGMG learned to align the two input modalities—textual descriptions and cellular morphological profiles—by projecting them into a shared latent space. This was achieved using a contrastive learning approach by optimizing the InfoNCE 45 loss function. Given a batch of size N, we computed the pairwise similarity scores between text and morphology embeddings. Let S ii denote the similarity between a correctly aligned (positive) text– morphology pair, and S ij represent the similarity between mismatched (negative) pairs. A temperature hyperparameter τ controlled the sharpness of the similarity distribution. The contrastive loss was defined as: We used a batch size of 48 and the temperature hyperparameter τ was set as 0.1. The weights of the morphology encoder from this stage were subsequently used in the molecule generation stage, where both modalities were jointly processed in an autoregressive decoding framework. Training for molecule generation To train the MGMG model for molecule generation, we used the AdamW optimizer and applied a cosine annealing learning rate scheduler, starting at a base learning rate of 1 × 10 $% and decaying to a minimum of 1 × 10 $& . The batch size was 48. To reduce overfitting and improve parameter efficiency, Low-Rank Adaptation (LoRA) was applied during training with a rank of 128, alpha value of 128, and a dropout rate of 0.1. This reduced the number of trainable parameters to 14 million—just 5.3% of the full model’s 267 million parameters. Training was performed on one NVIDIA A100 (80G) GPU. The MGMG model was trained for ∼90 epochs, totaling ∼51 hours of wall clock time. Training and evaluation were repeated using three different random seeds to ensure reproducibility. Baseline models We benchmarked MGMG against a range of baseline models, including MolT5, BioT5, and CPMolGAN. Each model was trained on the MolCaptioned–BBBC036v1 dataset using its own default or published hyperparameter configurations. To ensure reproducibility, training and evaluation for each model were repeated across three independent random seeds. – CPMolGAN 15 is the first model to use cell image features that capture compound-treated cell morphology as the input for molecule generation. In this framework, morphological profiles derived from the Cell Painting assay are used to condition the GAN, guiding the generation of molecules toward desired phenotypic effects 15 . – MolT5 29 is built upon the T5 architecture for joint modeling of natural language and molecular representations. It is pretrained on the “Colossal Clean Crawled Corpus” (C4) 44 dataset for the text modality and the ZINC dataset for molecule modality, enabling tasks such as molecule captioning and text-based molecule generation. – BioT5 21 is a T5-based model that learns from both molecular structures and biomedical text. It is pretrained on 339K molecule–text pairs from PubChem, where each molecule is represented as a SELFIES string paired with its corresponding textual description 21 . This ensures chemically valid generation while enabling the model to capture meaningful associations between molecular structures and biological literature. Evaluation Metrics To evaluate the performance of MGMG in molecule generation, we used the following metrics: Reconstruction metrics: BLEU score. BLEU (Bilingual Evaluation Understudy) is a precision-based metric that quantifies the similarity between the generated and reference molecular sequences. It measures the overlap of n-grams while applying a brevity penalty to discourage excessively short outputs 30 . A higher BLEU score indicates that the generated sequences more accurately reconstruct the reference molecular sequences. Levenshtein distance. The Levenshtein distance measures the minimum number of single-character edits (insertions, deletions, and substitutions) required to transform the generated molecular sequence into the reference molecular sequence 31 . A lower Levenshtein distance indicates a closer sequential similarity to the reference molecule. Synthesizability: The Synthetic Accessibility (SA) score quantifies how easily a given molecule can be synthesized, with lower scores indicating greater synthetic feasibility 34 . Following the approach of Zapata et al. 15 , we applied a threshold of 4.5 to report the percentage of compounds considered synthetically accessible. Drug-likeness: Fréchet ChemNet Distance (FCD). FCD assesses the structural similarity between generated and real molecules by comparing their latent space representations obtained from a pretrained ChemNet model. It quantifies how closely the distribution of generated molecules resembles that of real, bioactive compounds. Specifically, for two sets of molecules—generated set G and reference set R —the FCD is computed as the Fréchet distance between their multivariate Gaussian embeddings: where μ G , μ R are the mean vectors and Σ G , Σ R are the covariance matrices of the latent activations for the generated and reference molecules, respectively. A lower FCD indicates that the generated molecules are more structurally and chemically similar to real compounds 33 . The PyTorch implementation provided by Insilico Medicine ( https://github.com/insilicomedicine/fcd_torch ) was used to compute this metric. Quantitative Estimate of Drug-likeness (QED): The QED score measures the drug-likeness of a molecule based on multiple physicochemical properties. Higher QED scores indicate stronger alignment with known drug-like characteristics 46 . In this work, we followed prior conventions 15 and applied a threshold of 0.5 to quantify the percentage of generated molecules with favorable drug-like properties. Physicochemical properties: We computed Molecular Weight (MW), Lipophilicity (LogP), Hydrogen Bond Acceptors (HBA) and Hydrogen Bond Donors (HBD) using RDKit 47 . Diversity: To assess the structural diversity of the generated molecules, we computed Murcko scaffolds 48 using RDKit, which extracts the core molecular structure by removing side chains while retaining ring systems and linkers. Diversity is defined as the proportion of unique scaffolds among all valid generated molecules, reflecting the model’s ability to explore diverse chemical backbones rather than repeatedly generating similar structures 49 . Validity: Validity is defined as the proportion of generated molecular strings, e.g., SMILES, that can be successfully parsed into chemically valid molecules using RDKit. A molecule is considered valid if it satisfies fundamental chemical rules, including atom valency and bond configurations, particularly within aromatic systems. This metric assesses the model’s ability to generate syntactically correct and chemically feasible structures 49 . Uniqueness: Uniqueness measures the proportion of unique molecules among all the valid generated molecules. This metric ensures that the model does not collapse into generating redundant or repetitive structures and reflects its capacity for diverse molecule generation 49 . Molecule generation for activator design task For the activator design study, we used gene overexpression morphological profiles from Zapata et al.’s work 15 . These profiles were originally derived from the BBBC037v1 dataset from the Broad Bioimage Benchmark Collection, which contains Cell Painting images of U2OS cells transfected to overexpress 220 genes 35 . We focused on nine genes with known matched activators in the ExCAPE activators 37 database. All profiles were processed using the normalization and aggregation strategy described in Zapata et al.’s work 15 . Regarding the textual descriptions, we used GPT-4.0 to generate simple functional annotations for the desired activators, avoiding explicit structural cues. These input textual descriptions are provided in Supplementary Table 2 . For each gene, 2,500 valid molecules were generated using consensus morphological profiles. The process followed three steps to ensure chemical and semantic diversity: (1) 50 molecules were generated based on the GPT-4.0-provided descriptions and the gene’s morphological profile; (2) these molecules were then captioned using BioT5; and (3) each caption was combined again with the same morphology profile to generate 50 additional molecules. All generated molecules were filtered using physicochemical criteria following Zapata et al. 15 , including: −2 < LogP < 7, HBA + HBD < 10, topological polar surface area (TPSA) < 150, rotatable bonds < 150, and no SureChEMBL alerts. Molecular Docking Glide docking 38 was performed to assess interactions between the target protein and MGMG-designed molecules as well as reference molecules. Protein structures were prepared using the Schrödinger Protein Preparation Workflow at pH 7.4. Protonation states of ionizable amino acids were predicted using PROPKA, and those of ligands were determined using Epik. Waters beyond 5Å from ligands were removed, and the structure was minimized with the OPLS4 force field, converging heavy atoms to a root-mean-square deviation (RMSD) of 0.3Å. The receptor grid box was generated based on the binding site of the reference molecule with default Glide Grid Generation settings. MGMG-designed molecules were prepared for docking using Schrödinger LigPrep with the OPLS4 force field. Preparation included converting 2D to 3D structures, adding hydrogens, computing partial charges, and optimizing the structures. The docking results were reported as GlideScores. All calculations were conducted using Maestro (version 14.2.118, Schrödinger). Data and Code Availability The code used to generate the results presented in this study will be available from the corresponding author upon request after the manuscript’s final publication. Author Contributions Y.L. conceived and supervised the research. Y.L. and Q.T. designed and implemented the methodology. Q.T. and D.D. implemented the baselines. X.Y., C.L., and H.L. contributed to discussions on experimental design. P.R. and C.L. refined the molecular textual descriptions. Y.L., Q.T. and G.S. wrote the paper. All authors reviewed and approved the manuscript. Declaration of Interests The authors do not have conflicts to declare. Supplementary Materials View this table: View inline View popup Download powerpoint Supplementary Table 1: Top 1 generated molecules by MGMG with or without morphology information on MolCaptioned-BBBC036v1 and Mol-Instructions-BBBC036v1 dataset. View this table: View inline View popup Download powerpoint Supplementary Table 2: Textual description inputs and matched scaffold hits in the activator design case study. Download figure Open in new tab Supplementary Fig. 1: Comparison of MGMG and unimodal baselines on trainable parameter counts and molecule reconstruction performance as measured by BLEU score. Download figure Open in new tab Supplementary Fig. 2: Example comparisons of the top 1 molecule generated by MGMG and unimodal baselines based on the same textual descriptions and morphological profiles of reference small molecules. Download figure Open in new tab Supplementary Fig. 3: Examples of the top 5 molecules generated by MGMG based on the textual description and morphological profile of a small molecule. Download figure Open in new tab Supplementary Fig. 4. An example of molecules generated by MGMG using the same morphological profile but paired with either rich-informative or under-informative textual descriptions from the MolCaptioned-BBBC036v1 and Mol-Instructions-BBBC036v1 datasets. Download figure Open in new tab Supplementary Fig. 5: Additional examples of molecular docking results of MGMG generated molecules. Acknowledgements This work was supported in part by the University of Florida (UF Startup Fund and UF Research AI Award to Y. L., UF Blue Future Medicine Initiative to H. L.), the Florida Department of Health (Florida Cancer Innovation Fund grant #MOABK), National Institutes of Health (RM1GM145426 to H. L.), the Debbie and Sylvia DeSantis Chair Professorship (to H. L.), the Bodor Professorship Fund (to C. L.), and the National Science Foundation (NSF South Carolina EPSCoR 25-GA04). We acknowledge UFIT Research Computing for providing computational resources. Figures 1 , 2, 3, 4 were created with BioRender.com. References ↵ Zhang , K. et al. Artificial intelligence in drug development . Nat. Med . 31 , 45 – 59 ( 2025 ). doi: 10.1038/s41591-024-03434-4 OpenUrl CrossRef PubMed ↵ Peng , X. et al. Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets . arXiv [cs.LG ] ( 2022 ). doi: 10.48550/arXiv.2205.07249 OpenUrl CrossRef Guan , J. et al. 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction . arXiv [q-bio.BM ] ( 2023 ). doi: 10.48550/arXiv.2303.03543 OpenUrl CrossRef Schneuing , A. et al. Structure-based drug design with equivariant diffusion models . Nat. Comput. Sci . 4 , 899 – 909 ( 2024 ). doi: 10.1038/s43588-024-00737-x OpenUrl CrossRef PubMed Jiang , Y. et al. PocketFlow is a data-and-knowledge-driven structure-based molecular generative model. Nat . Mach. Intell . 6 , 326 – 337 ( 2024 ). doi: 10.1038/s42256-024-00808-8 OpenUrl CrossRef ↵ Wang , Y. , Wang , Z. , Li , Y. , Yan , P. & Li , X . Improving Covalent and Noncovalent Molecule Generation via Reinforcement Learning with Functional Fragments . J. Chem. Inf. Model 65 , 5934 – 5944 ( 2025 ). doi: 10.1021/acs.jcim.5c00944 OpenUrl CrossRef PubMed ↵ Martinelli , D. D . Generative machine learning for de novo drug discovery: A systematic review . Comput. Biol. Med . 145 , 105403 ( 2022 ). doi: 10.1016/j.compbiomed.2022.105403 OpenUrl CrossRef PubMed Grisoni , F . Chemical language models for de novo drug design: Challenges and opportunities . Curr. Opin. Struct. Biol . 79 , 102527 ( 2023 ). doi: 10.1016/j.sbi.2023.102527 OpenUrl CrossRef ↵ Wang , Z. et al. AptaDiff: de novo design and optimization of aptamers based on diffusion models . Brief. Bioinform . 25 , bbae517 ( 2024 ). doi: 10.1093/bib/bbae517 OpenUrl CrossRef PubMed ↵ Vincent , F. et al. Phenotypic drug discovery: recent successes, lessons learned and new directions . Nat. Rev. Drug Discov . 21 , 899 – 914 ( 2022 ). doi: 10.1038/s41573-022-00472-w OpenUrl CrossRef PubMed ↵ Moffat , J. G. , Vincent , F. , Lee , J. A. , Eder , J. & Prunotto , M . Opportunities and challenges in phenotypic drug discovery: an industry perspective . Nat. Rev. Drug Discov . 16 , 531 – 543 ( 2017 ). doi: 10.1038/nrd.2017.111 OpenUrl CrossRef PubMed ↵ Tang , Q. et al. Morphological profiling for drug discovery in the era of deep learning . Brief. Bioinform . 25 ( 2024 ). doi: 10.1093/bib/bbae284 OpenUrl CrossRef ↵ Bray , M. A. et al. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes . Nat. Protoc . 11 , 1757 – 1774 ( 2016 ). doi: 10.1038/nprot.2016.105 OpenUrl CrossRef PubMed ↵ Seal , S. et al. Cell Painting: a decade of discovery and innovation in cellular imaging . Nat. Methods . 22 , 254 – 268 ( 2025 ). doi: 10.1038/s41592-024-02528-8 OpenUrl CrossRef ↵ Marin Zapata , P. A. , et al. Cell morphology-guided de novo hit design by conditioning GANs on phenotypic image features . Digit. Discov . 2 , 91 – 102 ( 2023 ). doi: 10.1039/d2dd00081d OpenUrl CrossRef ↵ Lu , S. Z. et al. Cell Morphology-Guided Small Molecule Generation with GFlowNets . arXiv [cs.LG ] ( 2024 ). doi: 10.48550/arXiv.2408.05196 OpenUrl CrossRef ↵ Pei , Q. et al. BioT5+: Towards Generalized Biological Understanding with IUPAC Integration and Multi-task Tuning . Findings of the Association for Computational Linguistics: ACL 2024 , 1216 – 1240 ( 2024 ). doi: 10.18653/v1/2024.findings-acl.71 OpenUrl CrossRef Edwards , C. , Zhai , C. & Ji , H . Text2Mol: Cross-Modal Molecule Retrieval with Natural Language Queries . Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , 595 – 607 ( 2021 ). doi: 10.18653/v1/2021.emnlp-main.47 OpenUrl CrossRef Liu , S. et al. Multi-modal molecule structure–text model for text-based retrieval and editing. Nat . Mach. Intell . 5 , 1447 – 1457 ( 2023 ). doi: 10.1038/s42256-023-00759-6 OpenUrl CrossRef ↵ Fang , Y. et al. Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models . The Twelfth International Conference on Learning Representations ( 2024 ). doi: 10.48550/arXiv.2306.08018 OpenUrl CrossRef ↵ Pei , Q. et al. BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations . Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , 1102 – 1123 ( 2023 ). doi: 10.18653/v1/2023.emnlp-main.70 ↵ Stirling , D. R. et al. CellProfiler 4: improvements in speed, utility and usability . BMC Bioinformatics 22 , 433 ( 2021 ). doi: 10.1186/s12859-021-04344-9 OpenUrl CrossRef PubMed ↵ Chandrasekaran , S. N. , Ceulemans , H. , Boyd , J. D. & Carpenter , A. E . Image-based profiling for drug discovery: due for a machine-learning upgrade? Nat. Rev. Drug Discov . 20 , 145 – 159 ( 2021 ). doi: 10.1038/s41573-020-00117-w OpenUrl CrossRef PubMed ↵ Radford , A. et al. Learning Transferable Visual Models From Natural Language Supervision . International Conference on Machine Learning 2021 ( 2021 ). doi: 10.48550/arXiv.2103.00020 OpenUrl CrossRef ↵ Krenn , M. , Häse , F. , Nigam , A. , Friederich , P. & Aspuru-Guzik , A . Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation . Machine Learning: Science and Technology 1 ( 2020 ). doi: 10.1088/2632-2153/aba947 OpenUrl CrossRef ↵ Bray , M. A. et al. A dataset of images and morphological profiles of 30 000 small-molecule treatments using the Cell Painting assay . Gigascience 6 , 1 – 5 ( 2017 ). doi: 10.1093/gigascience/giw014 OpenUrl CrossRef PubMed ↵ Davies , M. et al. ChEMBL web services: streamlining access to drug discovery data and utilities . Nucleic Acids Res . 43 , W612 – W620 ( 2015 ). doi: 10.1093/nar/gkv352 OpenUrl CrossRef PubMed ↵ Hu , E. J. et al. LoRA: Low-Rank Adaptation of Large Language Models . International Conference on Learning Representations ( 2022 ). doi: 10.48550/arXiv.2106.09685 OpenUrl CrossRef ↵ Edwards , C. et al. Translation between Molecules and Natural Language . Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , 375 – 413 ( 2022 ). doi: 10.18653/v1/2022.emnlp-main.26 OpenUrl CrossRef ↵ Papineni , K. , Roukos , S. , Ward , T. & Zhu , W.-J . BLEU: a method for automatic evaluation of machine translation . Proceedings of the 40th Annual Meeting on Association for Computational Linguistics , 311 – 318 ( 2002 ). doi: 10.3115/1073083.1073135 OpenUrl CrossRef ↵ Levenshtein , V. I . Binary codes capable of correcting deletions, insertions and reversals . Soviet Physics Doklady 10 , 707 -- 710 ( 1966 ). OpenUrl ↵ Lipinski , C. A. , Lombardo , F. , Dominy , B. W. & Feeney , P. J . Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings . Adv. Drug Deliv. Rev . 23 , 3 – 25 ( 1997 ). doi: 10.1016/S0169-409X(96)00423-1 OpenUrl CrossRef PubMed Web of Science ↵ Preuer , K. , Renz , P. , Unterthiner , T. , Hochreiter , S. & Klambauer , G . Fréchet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery . J. Chem. Inf. Model . 58 , 1736 – 1741 ( 2018 ). doi: 10.1021/acs.jcim.8b00234 OpenUrl CrossRef PubMed ↵ Ertl , P. & Schuffenhauer , A . Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions . J. Cheminform . 1 , 8 ( 2009 ). doi: 10.1186/1758-2946-1-8 OpenUrl CrossRef PubMed ↵ Rohban , M. H. et al. Systematic morphological profiling of human gene and allele function via Cell Painting . eLife 6 , e24060 ( 2017 ). doi: 10.7554/eLife.24060 OpenUrl CrossRef PubMed ↵ OpenAi et al. GPT-4 Technical Report . arXiv [cs.CL] ( 2024 ). doi: 10.48550/arXiv.2303.08774 OpenUrl CrossRef ↵ Sun , J. et al. ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics . J. Cheminform . 9 , 17 ( 2017 ). doi: 10.1186/s13321-017-0203-5 OpenUrl CrossRef PubMed ↵ Friesner , R. A. et al. Glide: A New Approach for Rapid, Accurate Docking and Scoring. 1. Method and Assessment of Docking Accuracy . J. Med. Chem. 47 , 1739 – 1749 ( 2004 ). doi: 10.1021/jm0306430 OpenUrl CrossRef PubMed Web of Science ↵ Chandrasekaran , S. N. et al. JUMP Cell Painting dataset: morphological impact of 136,000 chemical and genetic perturbations . bioRxiv ( 2023 ). doi: 10.1101/2023.03.23.534023 OpenUrl Abstract / FREE Full Text ↵ Liao , C. , Yu , Y. , Mei , Y. & Wei , Y . From Words to Molecules: A Survey of Large Language Models in Chemistry . arXiv [cs.LG ] ( 2024 ). doi: 10.48550/arXiv.2402.01439 OpenUrl CrossRef ↵ Moshkov , N. et al. Learning representations for image-based profiling of perturbations . Nat. Commun . 15 , 1594 ( 2024 ). doi: 10.1038/s41467-024-45999-1 OpenUrl CrossRef PubMed ↵ Kraus , O. et al. Masked Autoencoders for Microscopy are Scalable Learners of Cellular Biology . 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 11757 – 11768 ( 2024 ). doi: 10.1109/CVPR52733.2024.01117 ↵ Serrano , E. et al. Reproducible image-based profiling with Pycytominer . Nat. Methods . 22 , 677 – 680 ( 2025 ). doi: 10.1038/s41592-025-02611-8 OpenUrl CrossRef ↵ Raffel , C. et al. Exploring the limits of transfer learning with a unified text-to-text transformer . J. Mach. Learn. Res . 21 ( 2020 ). doi: 10.48550/arXiv.1910.10683 OpenUrl CrossRef PubMed ↵ van den Oord , A. , Li , Y. & Vinyals , O. Representation Learning with Contrastive Predictive Coding . arXiv [cs.LG] ( 2019 ). doi: 10.48550/arXiv.1807.03748 OpenUrl CrossRef ↵ Bickerton , G. R. , Paolini , G. V. , Besnard , J. , Muresan , S. & Hopkins , A. L . Quantifying the chemical beauty of drugs . Nat. Chem . 4 , 90 – 98 ( 2012 ). doi: 10.1038/nchem.1243 OpenUrl CrossRef PubMed ↵ RDKit : Open-source cheminformatics . https://www.rdkit.org doi: 10.5281/zenodo.591637 OpenUrl CrossRef ↵ Bemis , G. W. & Murcko , M. A . The Properties of Known Drugs. 1. Molecular Frameworks . J. Med. Chem . 39 , 2887 – 2893 ( 1996 ). doi: 10.1021/jm9602928 OpenUrl CrossRef PubMed Web of Science ↵ Polykovskiy , D. et al. Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models . Front. pharmacol . 11 , 565644 ( 2020 ). doi: 10.3389/fphar.2020.565644 OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted July 17, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery Qiaosi Tang , Daoyun Ding , Xiaoyong Yuan , Gustavo Seabra , Peter A Ramdhan , Chi-Yuan Liu , My T. Thai , Chenglong Li , Hendrik Luesch , Yanjun Li bioRxiv 2025.07.11.664424; doi: https://doi.org/10.1101/2025.07.11.664424 Share This Article: Copy Citation Tools MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery Qiaosi Tang , Daoyun Ding , Xiaoyong Yuan , Gustavo Seabra , Peter A Ramdhan , Chi-Yuan Liu , My T. Thai , Chenglong Li , Hendrik Luesch , Yanjun Li bioRxiv 2025.07.11.664424; doi: https://doi.org/10.1101/2025.07.11.664424 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioinformatics Subject Areas All Articles Animal Behavior and Cognition (7635) Biochemistry (17697) Bioengineering (13895) Bioinformatics (41953) Biophysics (21456) Cancer Biology (18595) Cell Biology (25521) Clinical Trials (138) Developmental Biology (13381) Ecology (19903) Epidemiology (2067) Evolutionary Biology (24323) Genetics (15612) Genomics (22511) Immunology (17738) Microbiology (40401) Molecular Biology (17184) Neuroscience (88623) Paleontology (667) Pathology (2833) Pharmacology and Toxicology (4825) Physiology (7644) Plant Biology (15158) Scientific Communication and Education (2046) Synthetic Biology (4296) Systems Biology (9825) Zoology (2271)