{"id":98919,"date":"2026-10-18T19:33:06","date_gmt":"2026-10-18T16:33:06","guid":{"rendered":"https:\/\/twodots.gr\/?p=98919"},"modified":"2026-10-18T19:33:06","modified_gmt":"2026-10-18T16:33:06","slug":"molemb-multimodal-ai-moriaki-analysis","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/molemb-multimodal-ai-moriaki-anazitisi\/","title":{"rendered":"MolEmb: When Multimodal AI Models Become Molecular Search Engines"},"content":{"rendered":"<p>Large multimodal artificial intelligence models have been presented primarily as systems that \u00absee,\u00bb read, and generate responses. The study <strong>MolEmb<\/strong> examines a different role: whether the same type of model can be converted into a framework for representing molecules, generating vectors that are reused in property prediction, search, and retrieval of scientific information.<\/p>\n<p>The change is not merely technical. A classic molecular encoder typically takes a representation of the molecule\u2014such as a graph, a SMILES string, or its geometry\u2014and returns a fixed vector. MolEmb proposes an interface where the molecule is presented as a two-dimensional visualization and canonical SMILES, while a natural-language prompt defines the scientific context. Thus, the same molecular identity can take on different useful representations when the query concerns, for example, toxicity, solubility, or crossing of the blood-brain barrier.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">The short answer<\/p>\n<p>MolEmb shows that a multimodal AI model can become a useful molecular embedding model only when it is aligned with data that covers real-world search tasks. The key finding is not merely high performance, but that context sensitivity arises primarily from the supervision scheme and must be controlled separately from molecule recognition.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#stathero-vector-oikogeneia-anaparastaseon\">From a fixed vector to a family of representations<\/a><\/li>\n<li><a href=\"#pos-molemb-prosarmozei-mllm\">How MolEmb adapts an MLLM<\/a><\/li>\n<li><a href=\"#dedomena-efthygrammisis-poikilia\">Alignment Data and Why Their Variety Matters<\/a><\/li>\n<li><a href=\"#provlepsi-moriakon-idiotiton\">Prediction of molecular properties in eight benchmarks<\/a><\/li>\n<li><a href=\"#molecule-text-retrieval\">The Crucial Test: Molecule-Text Retrieval<\/a><\/li>\n<li><a href=\"#molcar-polla-contexts\">MolCAR: When a Single Molecule Has Multiple Correct Answers<\/a><\/li>\n<li><a href=\"#geniki-efthygrammisi-lysi\">The Failure of General Alignment and the Solution<\/a><\/li>\n<li><a href=\"#data-lesson-ai-products\">The Essential Data Lesson for AI Products<\/a><\/li>\n<li><a href=\"#oria-erevnas\">What the research does not prove<\/a><\/li>\n<li><a href=\"#praktiki-stratigiki-retrieval\">A Practical Strategy for Teams Building Retrieval Systems<\/a><\/li>\n<li><a href=\"#ti-kratame-molemb\">What We Take Away from MolEmb<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"stathero-vector-oikogeneia-anaparastaseon\">From a fixed vector to a family of representations<\/h2>\n<p>The authors define the <em>general molecular embedding model<\/em> as a function that takes two arguments: a molecular profile containing the molecule\u2019s available faces and a semantic context. The output remains a fixed-length vector, but it is not necessarily the same for every use. For a given molecule, the set of embeddings resulting from different contexts forms a context-aware family.<\/p>\n<p>This distinguishes three concepts that are often confused. The molecular identity identifies the molecule. The profile describes what observations are available\u2014in this particular study, a 2D image and SMILES. Context specifies the scientific lens through which the representation should be organized. The goal is not for a vector to \u00abknow everything,\u00bb but to activate the right information for the right task.<\/p>\n<p>For companies investing in AI infrastructure, this concept reflects a broader pattern: value lies not only in content production, but also in the creation of stable representations that power search, recommendations, matching, and retrieval-augmented workflows. The same distinction appears in <a href=\"https:\/\/twodots.gr\/multi-vector-embeddings-domain-finetuning-etairiki-anazitisi\/\">multi-vector embeddings for corporate search<\/a>, where the correct representation depends on the domain and the query. In chemistry, however, the cost of an incorrect generalization is high, which is why evaluation must clearly distinguish between prediction, retrieval, and context sensitivity.<\/p>\n<div class=\"td-comparison td-comparison-cards--horizontal\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">Three Levels of Information<\/p>\n<p class=\"td-chart-title\">The embedding must distinguish between identity, available views, and purpose<\/p>\n<p class=\"td-chart-intro\">If the three layers are merged into a single opaque vector, the evaluation cannot show whether the model understood the object or simply recognized a familiar pattern.<\/p>\n<\/div>\n<div class=\"td-comparison-grid\">\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Id. at<\/p>\n<p>What is the molecule, regardless of the question asked?.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Company<\/p>\n<p>Which views are available? Here are the 2D representation and the canonical SMILES.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Context<\/p>\n<p>Which qualification or scientific paper should guide the retrieval?.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"pos-molemb-prosarmozei-mllm\">How MolEmb adapts an MLLM<\/h2>\n<p>MolEmb retains the pretrained backbone of the multimodal large language model and adds small, trainable LoRA adapters. The input combines the two-dimensional molecular representation, the canonical SMILES, and, where necessary, a natural language prompt. The hidden state at the end-of-sequence position is used as the final embedding, without additional mean or CLS pooling.<\/p>\n<p>Training aligns pairs of molecules and descriptions in the same space. The molecular side and the textual side pass through the same MLLM backbone in separate forward passes. A bidirectional contrastive objective of the InfoNCE type brings the correct pair closer together and pushes the incorrect pairs further apart. The temperature is set to 0.07, and the vectors are normalized before calculating the cosine similarity.<\/p>\n<p>The researchers are testing three backbones: Intern-S1-mini 8B, Qwen3-VL-8B, and Qwen3.5-0.8B. This selection allows them to examine both a scientifically pre-trained model and, more generally, multimodal models of different sizes. Training was performed on eight NVIDIA A100 GPUs, with LoRA applied to both the language and visual components.<\/p>\n<h2 id=\"dedomena-efthygrammisis-poikilia\">Alignment Data and Why Their Variety Matters<\/h2>\n<p>The basic molecule-text alignment uses three sources: MolTextNet, KnowMol-100k, and ChEBI-20-MM. The MolTextNet schema is based on a corpus of descriptions from a single primary source. The Mixed condition combines the three sources to add names, functional groups, ontology-type definitions, and physicochemical properties.<\/p>\n<p>After the global canonical SMILES check, the Mixed dataset includes 411,709 training pairs, 4,149 validation pairs, and 4,165 test pairs. The team removed 5,041 duplicate pairs. These numbers are important not to showcase scale as an end in itself, but because the study concludes that covering scientific contexts is more important than simply adding more similar descriptions.<\/p>\n<p>In property tests, Mixed alignment improves on certain endpoints, particularly ESOL and Tox21, but it does not outperform other methods across the board. MolTextNet remains stronger on other tasks. This precludes an oversimplified interpretation along the lines of \u00abmore datasets always mean a better model.\u00bb The composition and semantic diversity of the training data determine what knowledge is transferred.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">Why variety isn't just about volume<\/p>\n<p>A large corpus with repetitive descriptions can improve recognition without teaching context routing. The right question for an operational retrieval system is whether the examples cover actual usage intentions, not just whether their total number is large.<\/p>\n<\/div>\n<h2 id=\"provlepsi-moriakon-idiotiton\">Prediction of molecular properties in eight benchmarks<\/h2>\n<p>The evaluation covers three regression benchmarks\u2014ESOL, Lipophilicity, and FreeSolv\u2014and five classification benchmarks: BACE, BBBP, ClinTox, Tox21, and SIDER. For regression, RMSE is used, where a lower value is better. For classification, ROC-AUC is used, with macro averaging applied to the multi-label datasets Tox21 and SIDER.<\/p>\n<p>The direct adaptation of MLLMs, without prior molecule-text alignment, already matches the performance of several pretrained graph methods, especially in regression tasks. For example, Intern-S1-mini Direct achieves an RMSE of 0.827 on ESOL and 1.839 on FreeSolv, while on ClinTox it achieves a ROC-AUC of 99.51%. These are results specific to this protocol and do not prove that an MLLM generally outperforms other methods in every setting.<\/p>\n<p>The most cautious conclusion is that the multimodal interface can be optimized for molecular prediction. Removing the 2D image leads to small but consistent losses on most tasks, so the visual aspect adds information beyond SMILES. Nevertheless, prediction using a task-specific head is not sufficient to prove that we have a reusable embedding model.<\/p>\n<h2 id=\"molecule-text-retrieval\">The Crucial Test: Molecule-Text Retrieval<\/h2>\n<p>To verify whether the space is truly shared between molecules and text, the authors perform molecule-to-text and text-to-molecule retrieval on 2,970 held-out MolTextNet pairs. The random R@1 is approximately 0.034%. An off-the-shelf Qwen3-VL-Embedding-8B achieves approximately 3% R@1 in both directions, indicating that a general multimodal embedding does not, on its own, possess a reliable chemical space.<\/p>\n<p>After contrastive alignment, all MolEmb backbones exceed 73% R@1 in both directions. Intern-S1-mini achieves 85.39% for molecule-to-text and 84.88% for text-to-molecule, while Qwen3.5-0.8B achieves 83,97% and 84,58%, respectively. The scientifically pre-trained Intern-S1-mini achieves the best R@1 performance among the three backbones.<\/p>\n<p>The comparison shows why it is not enough to simply purchase a \u00abgeneral-purpose embedding API\u00bb and assume that it will work in a specialized scientific field. The backbone provides capabilities and a common interface, but a reliable domain space results from targeted alignment and validation using data that represents real-world usage. The same reasoning applies when designing <a href=\"https:\/\/twodots.gr\/hybrid-ai-search-papers-with-code\/\">Hybrid AI search using keywords, embeddings, and fallbacks<\/a>.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">Four metrics that explain MolEmb's surge<\/p>\n<p class=\"td-chart-subtitle\">The values are taken from the study protocol and do not represent general accuracy for drug discovery.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">2.970<\/span><span class=\"td-metric-label\">held-out pairs<\/span><span class=\"td-metric-note\">MolTextNet Retrieval Test<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">85,39%<\/span><span class=\"td-metric-label\">molecule-to-text R@1<\/span><span class=\"td-metric-note\">Intern-S1-mini after alignment<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">84,88%<\/span><span class=\"td-metric-label\">text-to-molecule R@1<\/span><span class=\"td-metric-note\">Intern-S1-mini after alignment<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">99,8%<\/span><span class=\"td-metric-label\">Context R@1<\/span><span class=\"td-metric-note\">MolCAR-Structured after continued alignment<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"molcar-polla-contexts\">MolCAR: When a Single Molecule Has Multiple Correct Answers<\/h2>\n<p>The MolCAR benchmark isolates a more challenging question: Can the model select the correct information for the same molecule when only the instruction changes? Each target document links a real experimental outcome from one of eight task families to the same molecular profile. Thus, the model cannot succeed simply by recognizing the molecule. It must discern the scientific context.<\/p>\n<p>MolCAR-Structured deterministically generates cards from real labels and molecular descriptors. The evaluation set consists of 4,092 molecule-context pairs across 1,653 held-out molecules. MolCAR-Natural replaces the structured cards with scientific-style notes generated by Intern-S1 235B, but reconnects the verified observed results and the molecular profile to preserve the ground truth.<\/p>\n<p>The eight metrics include BACE-1 inhibition, blood-brain barrier permeability, water solubility, hydration free energy, lipophilicity, Tox21 toxicity outcomes, clinical toxicity, and side-effect profile. The main metric, Context R@1, examines whether the correct context ranks first among the documents for the same molecule, largely eliminating the advantage of simple identity recognition.<\/p>\n<h2 id=\"geniki-efthygrammisi-lysi\">The Failure of General Alignment and the Solution<\/h2>\n<p>Prior to continued alignment, task instructions have little effect on context ranking. In MolCAR-Structured, Intern-S1-mini achieves 44.4% Context R@1 with instructions and 40.4% without instructions. Qwen3.5-0.8B scores 41.11 TP3T compared to 40.41 TP3T. In other words, the general molecule-text alignment model learned to link molecules and descriptions, but it did not learn to reliably distinguish between multiple valid descriptions of the same molecule.<\/p>\n<p>The authors attribute this breakdown to the structure of the data: typical descriptions are relatively narrow and templated, and rarely present the same molecule through different task lenses. MolCAR-Train adds precisely this supervision. It includes 10,844 pairs for continued alignment, after evaluation molecules based on canonical SMILES have been removed from the training set.<\/p>\n<p>After continued alignment, the Intern-S1-mini achieves 99.81 TP3T Context R@1 on the structured benchmark, compared to 40.41 TP3T without task instructions. The Qwen3.5-0.8B achieves 69.41 TP3T. In the natural variant, the corresponding values are 99.21 TP3T and 72.01 TP3T. This change does not mean that the open scientific question has been resolved. It demonstrates, in a controlled benchmark of known task families, that context sensitivity can be learned when the right supervision is provided.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">Critical threshold for interpretation<\/p>\n<p class=\"td-decision-title\">Do not confuse object recognition with understanding intent<\/p>\n<p>A retrieval system may find the correct product, document, or molecule while simultaneously returning incorrect information in response to the user\u2019s query. The evaluation must keep the identity constant and change only the context in order to reveal this failure mode.<\/p>\n<\/div>\n<h2 id=\"data-lesson-ai-products\">The Essential Data Lesson for AI Products<\/h2>\n<p>MolEmb\u2019s most important finding is that context-aware embedding is primarily a property of the training data. A larger or more general backbone is not sufficient if the training corpus does not demonstrate how the relevant information changes depending on the task. Embeddings do not derive their operational significance from the model\u2019s name, but from the contrasts they observed during training.<\/p>\n<p>This principle extends well beyond chemistry. In an e-commerce retrieval system, the same product may be relevant for different purposes: compatibility, profit margin, sustainability, or fast delivery. In a B2B knowledge base, the same document may be evaluated differently for sales enablement, compliance, or technical support. This is a design analogy, not a result measured by the study. The practical lesson is to collect task-conditioned examples that correspond to actual product decisions.<\/p>\n<p>For AI and marketing technology teams, this also changes how we measure performance. A single retrieval score may mask the fact that the system correctly identifies the object but ignores the intent. Tests are needed that keep the identity constant and change only the context, exactly as MolCAR does. Similar attention is needed in <a href=\"https:\/\/twodots.gr\/multimodal-anazitisi-eikonon-embeddings-frontier-llms\/\">Multimodal image search using embeddings<\/a>, where the most powerful general model is not always the best solution for that particular retrieval task.<\/p>\n<h2 id=\"oria-erevnas\">What the research does not prove<\/h2>\n<p>MolEmb is considered an adaptation route rather than a fully scaled molecular foundation model. The molecular profile is limited to 2D depictions and canonical SMILES. It does not include conformers, reactions, spectra, or assay metadata, although these could be integrated into the same interface in the future.<\/p>\n<p>MolCAR covers eight property-related task families and does not test transfer to completely unknown task families or open-ended scientific search. Furthermore, the Mixed corpus remains small and less diverse than industrial-scale embedding pretraining. Therefore, the 99.81 TP3T score on a controlled context benchmark should not be presented as a general measure of accuracy in drug discovery.<\/p>\n<p>There is still the usual gap between benchmarks and production use: data quality, drift, scientific validation, computational cost, and governance. This paper documents a feasible architectural approach and a diagnostic tool, not a ready-to-use clinical or pharmaceutical decision-making system. In a production environment, documented risk controls and human oversight are also required, in accordance with the NIST AI RMF framework.<\/p>\n<h2 id=\"praktiki-stratigiki-retrieval\">A Practical Strategy for Teams Building Retrieval Systems<\/h2>\n<p>The first step is to define the object\u2019s views and usage contexts separately. In this study, the views are images and SMILES, while the contexts are scientific papers. In another domain, the views might be text, images, attributes, and history, while the contexts must be derived from actual business decisions.<\/p>\n<p>Next, a benchmark is needed that tests difficult negative examples. The correct object with the wrong context must be considered incorrect; otherwise, the metric will reward simple identification. The evaluation must also distinguish between object retrieval, document retrieval, and within-object context ranking, so that it is clear where the system is failing.<\/p>\n<p>Finally, the team needs a data loop. When context routing fails, the solution is not automatically a larger model. MolEmb shows that outcome-grounded, task-diverse supervision can substantially alter the geometry of the embedding space through parameter-efficient adaptation. The priority, therefore, is to improve the structure of the examples and repeat the evaluation without leakage.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">Designing Context-Aware Retrieval in Seven Steps<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Use Cases<\/span><strong>Map out their true intentions<\/strong>\n<p>List the decisions that the retrieval system will support, such as compatibility, technical support, compliance, or scientific validity.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Views<\/span><strong>Identify the faces of the object<\/strong>\n<p>Determine which text, images, attributes, and historical data make up the profile before deciding how to combine them.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Cons<\/span><strong>Create challenging negative examples<\/strong>\n<p>Keep the same object and change the context so that the model doesn't rely solely on identity recognition.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Metrics<\/span><strong>Measure different types of retrieval<\/strong>\n<p>Distinguish between object retrieval, document retrieval, and context ranking, rather than combining all errors into a single score.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Baseline<\/span><strong>Compare with a simple domain baseline<\/strong>\n<p>Check whether the complexity of MLLM adds value beyond that of a specialized encoder or a hybrid search architecture.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Data loop<\/span><strong>Turn Failures into Supervision<\/strong>\n<p>Collect outcome-grounded examples from contexts where the system confused the correct object with the wrong preposition.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Guardrails<\/span><strong>Set usage limits and review them<\/strong>\n<p>Monitor drift, cost, scientific validity, and human review before embeddings influence critical decisions.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"ti-kratame-molemb\">What We Take Away from MolEmb<\/h2>\n<p>The study shows that multimodal language models can serve as a basis for reusable molecular embeddings, not just as generators or chemistry assistants. Simple direct adaptation performs competitively on several property benchmarks, molecule-text contrastive alignment creates a robust shared retrieval space, and targeted continued alignment adds task sensitivity.<\/p>\n<p>The most well-founded conclusion is not that specialist molecular encoders have become obsolete. It is that a second path is emerging: a common multimodal interface, natural-language conditioning, and embeddings that can serve more than one downstream workflow. The success of this path will depend on datasets that capture not only what a molecule is, but also why the user is examining it.<\/p>\n<p>For a business, the practical value lies in the rigor of the evaluation. The model must demonstrate that it retrieves the correct information for the intended purpose, using sufficient data, clear baselines, and operational limits. Simply choosing a larger model does not replace this work.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-eyebrow\">Context-aware AI with measurable value<\/p>\n<p class=\"td-service-cta-title\">Design a retrieval system that understands intent, not just the object<\/p>\n<p>TWO DOTS maps data, contexts, baselines, and guardrails so that an AI search or automation workflow can be evaluated based on your company\u2019s actual decisions.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Check out Business Automation &amp; AI<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is MolEmb?;<\/summary>\n<div class=\"td-faq-content\">\n<p>It is a lightweight framework for fine-tuning multimodal large language models to generate reusable molecular embeddings. It aligns molecular profiles and textual descriptions in a shared space using bidirectional contrastive training.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Which faces of a molecule does it use?;<\/summary>\n<div class=\"td-faq-content\">\n<p>This implementation uses two-dimensional visualization and canonical SMILES. The authors cite conformers, reactions, spectra, and assay metadata as potential future views.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Why isn't a general multimodal embedding model enough?;<\/summary>\n<div class=\"td-faq-content\">\n<p>In the held-out MolTextNet retrieval task, the general Qwen3-VL-Embedding-8B achieved approximately 3% R@1, while the MolEmb models, after domain alignment, exceeded 73% in both directions.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What does MolCAR measure?;<\/summary>\n<div class=\"td-faq-content\">\n<p>It measures whether a model selects the correct task context for the same molecule. It keeps the molecular identity constant and changes the scientific lens in order to distinguish between molecule recognition and understanding the instruction.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is the main finding from the data?;<\/summary>\n<div class=\"td-faq-content\">\n<p>Context-aware behavior emerged primarily when the supervision included task-diverse, outcome-grounded examples. General molecule-description alignment alone did not yield reliable context routing.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Can MolEmb be used directly in drug discovery?;<\/summary>\n<div class=\"td-faq-content\">\n<p>This paper presents and evaluates a research approach, not a ready-made system for making pharmaceutical or clinical decisions. Further validation, more comprehensive molecular analyses, and productive controls are required.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What do high Context R@1 scores mean?;<\/summary>\n<div class=\"td-faq-content\">\n<p>This means that, in the validated MolCAR benchmark, the correct context ranked first among the contexts for the same molecule. It is not a general accuracy rate across all of chemistry.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is the practical course for other fields?;<\/summary>\n<div class=\"td-faq-content\">\n<p>As a design analogy, a retrieval system must be trained and evaluated based on actual usage intentions, not just on the identity of the objects. This analogy is an interpretation applicable to AI products and is not an experiment in this study.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.23646\" target=\"_blank\" rel=\"noopener\">Zhao et al.: MolEmb \u2013 Multimodal Large Language Models Can Be Effective Molecular Embedding Models<\/a><\/li>\n<li><a href=\"https:\/\/openreview.net\/pdf?id=dD95A7ATXn\" target=\"_blank\" rel=\"noopener\">OpenReview: MolEmb's full paper and supplementary tables<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/InternLM\/Intern-S1\" target=\"_blank\" rel=\"noopener\">InternLM: the official repository for Intern-S1<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/publications\/artificial-intelligence-risk-management-framework-ai-rmf-10\" target=\"_blank\" rel=\"noopener\">NIST: Artificial Intelligence Risk Management Framework 1.0<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>The MolEmb study shows how multimodal AI models can generate context-aware molecular embeddings and why supervision determines the value of retrieval.<\/p>","protected":false},"author":1,"featured_media":111641,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","footnotes":""},"categories":[199],"tags":[20639,21280,21279,21278,18467],"class_list":["post-98919","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-contrastive-learning","tag-drug-discovery","tag-molecular-embeddings","tag-molemb","tag-multimodal-ai"],"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/98919","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=98919"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/98919\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/111641"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=98919"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=98919"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=98919"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}