{"id":62,"date":"2024-03-18T17:59:19","date_gmt":"2024-03-18T17:59:19","guid":{"rendered":"http:\/\/localhost\/wordpress\/valancelabs\/?p=62"},"modified":"2025-05-20T12:29:00","modified_gmt":"2025-05-20T12:29:00","slug":"introducing-molgps-a-foundational-gnn-for-molecular-property-prediction","status":"publish","type":"post","link":"https:\/\/www.valencelabs.com\/fr\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/","title":{"rendered":"Pr\u00e9sentation de MolGPS \u2014 Un r\u00e9seau neuronal en graphe (GNN) fondamental pour la pr\u00e9diction des propri\u00e9t\u00e9s mol\u00e9culaires"},"content":{"rendered":"<figure class=\"wp-block-video\"><video height=\"990\" style=\"aspect-ratio: 1440 \/ 990;\" width=\"1440\" controls poster=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001444.png\" src=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Powered-by-Recursion_ver1.mp4\"><\/video><\/figure>\n\n\n\n<div class=\"fifty_fifty\">\n    <div class=\"container\">\n        <div class=\"left__col\">\n            <h2><span>Sur la scalabilit\u00e9<\/span> des GNN pour les graphes mol\u00e9culaires<\/h2>\n                    <\/div>\n        <div class=\"right__col\">\n            <div class=\"desc\">\n                <p>La mise \u00e0 l\u2019\u00e9chelle des mod\u00e8les d\u2019apprentissage profond a \u00e9t\u00e9 au c\u0153ur des r\u00e9centes r\u00e9volutions de mod\u00e9lisation du langage et de g\u00e9n\u00e9ration d\u2019images. Les praticiens ont observ\u00e9 une relation \u00e9troite entre la taille du mod\u00e8le, la taille de l\u2019ensemble de donn\u00e9es et les performances. Cependant, les architectures fond\u00e9es sur la structure, telles que les Graph Neural Networks (GNN), n\u2019ont pas encore d\u00e9montr\u00e9 les b\u00e9n\u00e9fices \u00e0 l\u2019\u00e9chelle, principalement en raison de l\u2019efficacit\u00e9 moindre des op\u00e9rations clairsem\u00e9es, des besoins importants en donn\u00e9es et du manque de clart\u00e9 quant \u00e0 l\u2019efficacit\u00e9 des diff\u00e9rentes architectures.<\/p>\n<p>Nous abordons cet inconv\u00e9nient des GNN en \u00e9tudiant leur comportement de mise \u00e0 l\u2019\u00e9chelle. Plus pr\u00e9cis\u00e9ment, nous analysons des r\u00e9seaux \u00e0 passage de messages, des transformateurs de graphes et des architectures hybrides sur la plus grande collection publique de graphes mol\u00e9culaires 2D. Pour la premi\u00e8re fois, nous observons que les GNN b\u00e9n\u00e9ficient consid\u00e9rablement de l\u2019augmentation de la largeur, du nombre de mol\u00e9cules, du nombre d\u2019\u00e9tiquettes et de la diversit\u00e9 des ensembles de donn\u00e9es de pr\u00e9\u2011entra\u00eenement.<\/p>\n<p>Nous sommes ravis de pr\u00e9senter MolGPS, un mod\u00e8le de 3B param\u00e8tres pour diverses t\u00e2ches de pr\u00e9diction de propri\u00e9t\u00e9s mol\u00e9culaires. Nous esp\u00e9rons que ce travail ouvrira la voie \u00e0 une \u00e8re o\u00f9 des GNN fondamentaux stimuleront la d\u00e9couverte de m\u00e9dicaments pharmaceutiques. Non seulement les performances du mod\u00e8le \u00e9voluent en fonction des param\u00e8tres, mais il b\u00e9n\u00e9ficie \u00e9galement \u00e9norm\u00e9ment de l'int\u00e9gration de donn\u00e9es ph\u00e9nomiques de haut niveau dans le tout.<\/p>            <\/div>\n        <\/div>\n    <\/div>\n<\/div>\n\n\n<div class=\"wp-block-columns blurred_container is-layout-flex wp-container-core-columns-is-layout-8f761849 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<div class=\"title__desc\">\n    <div class=\"container\">\n        <h2><span>D\u00e9tails du mod\u00e8le<\/span> et performances<\/h2>\n                    <div class=\"desc\">\n                <p>MolGPS a \u00e9t\u00e9 entra\u00een\u00e9 sur le m\u00e9lange d\u2019ensemble de donn\u00e9es <a href=\"https:\/\/arxiv.org\/abs\/2310.04292\" target=\"_blank\" rel=\"noopener\">LargeMix<\/a> , constitu\u00e9 de 5\u202fmillions de mol\u00e9cules regroup\u00e9es en 5 t\u00e2ches diff\u00e9rentes, chacune comportant plusieurs \u00e9tiquettes. LargeMix inclut notamment des ensembles de donn\u00e9es comme L1000_VCAP et L1000_MCF7 (transcriptomique), PCBA_1328 (bioessais), PCQM4M_G25 et PCGM4M_N4 (simulations DFT).<\/p>            <\/div>\n            <\/div>\n<\/div>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"920\" height=\"349\" src=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325822-1.png\" alt=\"\" class=\"wp-image-145\" srcset=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325822-1.png 920w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325822-1-300x114.png 300w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325822-1-768x291.png 768w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325822-1-18x7.png 18w\" sizes=\"auto, (max-width: 920px) 100vw, 920px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Nous avons \u00e9galement ajout\u00e9 un ensemble de donn\u00e9es de classification utilisant un sous\u2011ensemble des donn\u00e9es ph\u00e9nomiques de Recursion. Cet ensemble de donn\u00e9es a \u00e9t\u00e9 cr\u00e9\u00e9 \u00e0 l\u2019aide d\u2019un auto-encodeur masqu\u00e9 pr\u00e9\u2011entra\u00een\u00e9 regroupant les images ph\u00e9nomiques en 6\u202f000 classes diff\u00e9rentes, qui sont utilis\u00e9es ensuite pour des classifications binaires.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MolGPS a d\u2019abord \u00e9t\u00e9 pr\u00e9-entra\u00een\u00e9 \u00e0 l\u2019aide d'une strat\u00e9gie courante d'apprentissage multit\u00e2che supervis\u00e9e puis a \u00e9t\u00e9 finement ajust\u00e9 (ou sond\u00e9) pour diverses t\u00e2ches de pr\u00e9diction de propri\u00e9t\u00e9s mol\u00e9culaires afin d\u2019\u00e9valuer ses performances. Nous avons \u00e9valu\u00e9 les performances de MolGPS sur les bancs d'essai de Therapeutics Data Commons (TDC), MoleculeNet et Polaris.<\/p>\n\n\n\n<div class=\"title__desc\">\n    <div class=\"container\">\n        <h2>Therapeutics Data Commons (TDC) et <span>MoleculeNet<\/span><\/h2>\n            <\/div>\n<\/div>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"920\" height=\"623\" src=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001489.png\" alt=\"\" class=\"wp-image-148\" srcset=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001489.png 920w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001489-300x203.png 300w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001489-768x520.png 768w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001489-18x12.png 18w\" sizes=\"auto, (max-width: 920px) 100vw, 920px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Notre \u00e9tude se concentre sur 22 t\u00e2ches ADMET (absorption, distribution, m\u00e9tabolisme, excr\u00e9tion et toxicit\u00e9) disponibles dans TDC. Ce benchmark existe depuis des ann\u00e9es avec des soumissions continues de divers groupes, incluant \u00e0 la fois des mod\u00e8les d'apprentissage profond et d'apprentissage automatique traditionnel en t\u00eate du benchmark, avec un total de 8 mod\u00e8les se partageant les premi\u00e8res positions sur l\u2019ensemble des 22 t\u00e2ches. Simplement en mettant \u00e0 l'\u00e9chelle notre mod\u00e8le, nous avons constat\u00e9 que MolGPS surpasse SOTA dans 12\u202ft\u00e2ches sur\u202f22.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Nous \u00e9tudions \u00e9galement 4 jeux de MoleculeNet fr\u00e9quemment utilis\u00e9s dans des \u00e9tudes similaires\u202f: BACE (qui \u00e9value la liaison d\u2019un groupe d\u2019inhibiteurs ciblant la \u03b2\u2011s\u00e9cr\u00e9tase), BBBP (la p\u00e9n\u00e9tration de la barri\u00e8re h\u00e9mato\u2011enc\u00e9phalique, qui \u00e9value si une mol\u00e9cule peut p\u00e9n\u00e9trer dans le syst\u00e8me nerveux central), Clintox (qui est pertinent pour la toxicit\u00e9 des compos\u00e9s mol\u00e9culaires), et Sider (la ressource sur les effets secondaires, qui contient des informations sur les effets ind\u00e9sirables des m\u00e9dicaments dans une base de donn\u00e9es de m\u00e9dicaments commercialis\u00e9s). Nous avons constat\u00e9 que MolGPS surpasse SOTA (tous les mod\u00e8les auto-supervis\u00e9s ou les mod\u00e8les auto-supervis\u00e9s pr\u00e9-entra\u00een\u00e9s bas\u00e9s sur le quantum) sur les 4 t\u00e2ches.<\/p>\n\n\n\n<div class=\"title__desc\">\n    <div class=\"container\">\n        <h2><\/h2>\n                    <div class=\"desc\">\n                <h2><a href=\"https:\/\/polarishub.io\/\" target=\"_blank\" rel=\"noopener\">Polaris<\/a><\/h2>            <\/div>\n            <\/div>\n<\/div>\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"920\" height=\"444\" src=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001491.png\" alt=\"\" class=\"wp-image-152\" srcset=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001491.png 920w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001491-300x145.png 300w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001491-768x371.png 768w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-1000001491-18x9.png 18w\" sizes=\"auto, (max-width: 920px) 100vw, 920px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Bien que TDC et MoleculeNet soient des r\u00e9f\u00e9rences couramment utilis\u00e9es pour l'\u00e9valuation de la d\u00e9couverte de m\u00e9dicaments open-source, nous notons qu\u2019ils souffrent de biais de collecte et de traitement des donn\u00e9es sur des mol\u00e9cules diff\u00e9rentes. Ces limitations ont d\u00e9j\u00e0 \u00e9t\u00e9 d\u00e9crites lors de\u00a0<a href=\"https:\/\/practicalcheminformatics.blogspot.com\/2023\/08\/we-need-better-benchmarks-for-machine.html\">conversations<\/a>\u00a0au sein de la communaut\u00e9.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Polaris est une nouvelle collection de benchmarks et d\u2019ensemble de donn\u00e9es \u00e9labor\u00e9s au moyen d'un protocole d'\u00e9valuation normalis\u00e9, d\u00e9velopp\u00e9 par un\u00a0<a href=\"https:\/\/polarishub.io\/guidelines\/small-molecules\">consortium industriel<\/a>\u00a0de soci\u00e9t\u00e9s de biotechnologie et pharmaceutiques. Nous avons \u00e9tudi\u00e9 les performances de MolGPS sur 12 t\u00e2ches ADMET et de pr\u00e9diction de liaison et avons constat\u00e9 que MolGPS surpasse SOTA sur 11\/12 t\u00e2ches.<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"fifty_fifty\">\n    <div class=\"container\">\n        <div class=\"left__col\">\n            <h2><span>Exp\u00e9riences<\/span> de mise \u00e0 l'\u00e9chelle<\/h2>\n                            <div class=\"image_under_title\">\n                    <img loading=\"lazy\" decoding=\"async\" width=\"380\" height=\"344\" src=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Group-37666.png\" class=\"attachment-full size-full\" alt=\"\" srcset=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Group-37666.png 380w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Group-37666-300x272.png 300w, https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Group-37666-13x12.png 13w\" sizes=\"auto, (max-width: 380px) 100vw, 380px\" \/>                <\/div>\n                    <\/div>\n        <div class=\"right__col\">\n            <div class=\"desc\">\n                <p>Dans les exp\u00e9riences suivantes, nous examinons les performances de MolGPS en augmentant sa largeur de sorte que le nombre de param\u00e8tres passe de 1\u202fmillion \u00e0 3\u202fmilliards. Pour \u00e9valuer correctement les b\u00e9n\u00e9fices de l\u2019\u00e9chelle, nous mesurons les performances du mod\u00e8le sur 22 t\u00e2ches TDC en aval. Ici, la recherche d\u2019hyper-param\u00e8tres est effectu\u00e9e avec 10\u202fmillions de param\u00e8tres, et l\u2019augmentation de largeur est r\u00e9alis\u00e9e en zero\u2011shot \u00e0 l\u2019aide de la technique muTransfer.<\/p>\n<p><br \/>Nous observons \u00e9galement que les GNN profitent fortement de l\u2019augmentation de la largeur du mod\u00e8le, et que les gains sont coh\u00e9rents et lin\u00e9aires par rapport au logarithme du nombre de param\u00e8tres. Il n\u2019y a aucun ralentissement visible de la courbe de passage \u00e0 l\u2019\u00e9chelle, ce qui laisse penser que nous pouvons continuer \u00e0 am\u00e9liorer les performances avec des mod\u00e8les de plus en plus grands, \u00e0 l\u2019instar des LLM.<\/p>\n<p><br \/>Par ailleurs, MolGPS am\u00e9liore nettement les performances par rapport aux baselines TDC \u2014 il s\u2019agit des meilleures performances par t\u00e2che depuis l\u2019introduction de TDC en 2021. Sur l\u2019axe des ordonn\u00e9es, une valeur de 0 repr\u00e9sente la moyenne de toutes les soumissions au benchmark TDC. Compar\u00e9 au dernier SOTA sur TDC, notre ensemble de mod\u00e8les d\u00e9passe la courbe du meilleur mod\u00e8le par t\u00e2che, ce qui signifie qu\u2019il est g\u00e9n\u00e9ralement pr\u00e9f\u00e9rable d\u2019utiliser MolGPS plut\u00f4t que d\u2019essayer la trentaine de m\u00e9thodes du benchmark TDC.<\/p>\n<p><br \/>Nous notons aussi que le mod\u00e8le atteint une limite lorsqu\u2019il lorsqu'il est mis \u00e0 l'\u00e9chelle sur des donn\u00e9es publiques uniquement, mais que l\u2019ajout de donn\u00e9es ph\u00e9nomiques priv\u00e9es repousse consid\u00e9rablement les limites de l\u2019\u00e9chelle et des performances.<br \/>Nous rapportons ci\u2011dessus la \u00ab\u202fperformance normalis\u00e9e\u202f\u00bb, repr\u00e9sentant la moyenne du z\u2011score sur les 22 t\u00e2ches du TDC. Le z\u2011score sur la base des performances du mod\u00e8le par rapport au classement d'une t\u00e2che, ajust\u00e9 en fonction de la polarit\u00e9 de la m\u00e9trique de la t\u00e2che, c'est-\u00e0-dire multipli\u00e9 par -1 si \u00ab\u202fplus petit est meilleur\u202f\u00bb.<\/p>            <\/div>\n        <\/div>\n    <\/div>\n<\/div>\n\n\n<div class=\"fifty_fifty\">\n    <div class=\"container\">\n        <div class=\"left__col\">\n            <h2>Mol\u00e9cules et <span>langage<\/span><\/h2>\n                    <\/div>\n        <div class=\"right__col\">\n            <div class=\"desc\">\n                <p>Dans le contexte des LLM, nos r\u00e9sultats de mise \u00e0 l\u2019\u00e9chelle peuvent sembler surprenants puisque nos mod\u00e8les ne sont entra\u00een\u00e9s que sur quelques millions de points de donn\u00e9es (mol\u00e9cules), tandis que les LLM s\u2019entra\u00eenent g\u00e9n\u00e9ralement sur des ensemble contenant des billions de jetons (tokens). Pour mieux comprendre la hausse de performance et l\u2019\u00e9cart de taille des ensembles de donn\u00e9es, il est utile d\u2019\u00e9tablir quelques analogies entre mol\u00e9cules et langage.<\/p>\n<p><br \/>Dans notre contexte, les mol\u00e9cules sont analogues aux phrases dans le traitement du langage, tandis que les atomes et liaisons sont analogues aux jetons. De plus, la t\u00e2che est supervis\u00e9e et certaines mol\u00e9cules poss\u00e8dent des milliers d\u2019\u00e9tiquettes associ\u00e9es issues de donn\u00e9es exp\u00e9rimentales. Cela permet aux int\u00e9grations mol\u00e9culaires apprises d'\u00eatre beaucoup plus riches que celles d\u00e9riv\u00e9es de la simple r\u00e9cup\u00e9ration d'un jeton manquant.<\/p>            <\/div>\n        <\/div>\n    <\/div>\n<\/div>\n\n<div class=\"blog_cta\">\n    <div class=\"container\">\n        <div class=\"bg\">\n            <img decoding=\"async\" src=\"https:\/\/www.valencelabs.com\/wp-content\/themes\/valencelabs\/src\/images\/blog_cta_bg.png\" alt=\"Backgroung image\">\n        <\/div>\n\n        <div class=\"content\">\n            <h2>Vous souhaitez en savoir plus sur MolGPS ?<\/h2>\n                            <div class=\"desc\">\n                    Contactez notre \u00e9quipe.                <\/div>\n                                        <a href=\"#\" class=\"read_more\">\n                    Nous contacter                    <svg width=\"24\" height=\"24\" viewbox=\"0 0 24 24\" fill=\"none\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\">\n                    <path d=\"M5 12H19M19 12L13 6M19 12L13 18\" stroke=\"white\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"\/>\n                    <\/svg>\n                <\/a>\n                    <\/div>\n        \n    <\/div>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>Scaling deep learning models has been at the heart of recent revolutions in language modeling and image generation. Practitioners have observed a strong relationship&#8230;<\/p>","protected":false},"author":3,"featured_media":133,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[3],"class_list":["post-62","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized","tag-blog"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.6 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.valencelabs.com\/fr\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs\" \/>\n<meta property=\"og:description\" content=\"Scaling deep learning models has been at the heart of recent revolutions in language modeling and image generation. Practitioners have observed a strong relationship...\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.valencelabs.com\/fr\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/\" \/>\n<meta property=\"og:site_name\" content=\"Valence Labs\" \/>\n<meta property=\"article:published_time\" content=\"2024-03-18T17:59:19+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-05-20T12:29:00+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png\" \/>\n\t<meta property=\"og:image:width\" content=\"680\" \/>\n\t<meta property=\"og:image:height\" content=\"791\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Valence Labs\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Valence Labs\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"2 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/\"},\"author\":{\"name\":\"Valence Labs\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/#\\\/schema\\\/person\\\/1b232c7b905afbb857b4eccc92cc87d0\"},\"headline\":\"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction\",\"datePublished\":\"2024-03-18T17:59:19+00:00\",\"dateModified\":\"2025-05-20T12:29:00+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/\"},\"wordCount\":326,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.valencelabs.com\\\/wp-content\\\/uploads\\\/2025\\\/04\\\/Frame-2087325827.png\",\"keywords\":[\"Blog\"],\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/\",\"url\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/\",\"name\":\"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.valencelabs.com\\\/wp-content\\\/uploads\\\/2025\\\/04\\\/Frame-2087325827.png\",\"datePublished\":\"2024-03-18T17:59:19+00:00\",\"dateModified\":\"2025-05-20T12:29:00+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/#\\\/schema\\\/person\\\/1b232c7b905afbb857b4eccc92cc87d0\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#breadcrumb\"},\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.valencelabs.com\\\/wp-content\\\/uploads\\\/2025\\\/04\\\/Frame-2087325827.png\",\"contentUrl\":\"https:\\\/\\\/www.valencelabs.com\\\/wp-content\\\/uploads\\\/2025\\\/04\\\/Frame-2087325827.png\",\"width\":680,\"height\":791},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.valencelabs.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/#website\",\"url\":\"https:\\\/\\\/www.valencelabs.com\\\/\",\"name\":\"Valence Labs\",\"description\":\"Recursion\u2019s AI Research Engine\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.valencelabs.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.valencelabs.com\\\/#\\\/schema\\\/person\\\/1b232c7b905afbb857b4eccc92cc87d0\",\"name\":\"Valence Labs\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g\",\"caption\":\"Valence Labs\"},\"url\":\"https:\\\/\\\/www.valencelabs.com\\\/fr\\\/author\\\/valencelabs\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.valencelabs.com\/fr\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/","og_locale":"fr_FR","og_type":"article","og_title":"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs","og_description":"Scaling deep learning models has been at the heart of recent revolutions in language modeling and image generation. Practitioners have observed a strong relationship...","og_url":"https:\/\/www.valencelabs.com\/fr\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/","og_site_name":"Valence Labs","article_published_time":"2024-03-18T17:59:19+00:00","article_modified_time":"2025-05-20T12:29:00+00:00","og_image":[{"width":680,"height":791,"url":"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png","type":"image\/png"}],"author":"Valence Labs","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Valence Labs","Est. reading time":"2 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#article","isPartOf":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/"},"author":{"name":"Valence Labs","@id":"https:\/\/www.valencelabs.com\/#\/schema\/person\/1b232c7b905afbb857b4eccc92cc87d0"},"headline":"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction","datePublished":"2024-03-18T17:59:19+00:00","dateModified":"2025-05-20T12:29:00+00:00","mainEntityOfPage":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/"},"wordCount":326,"commentCount":0,"image":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#primaryimage"},"thumbnailUrl":"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png","keywords":["Blog"],"inLanguage":"fr-FR","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/","url":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/","name":"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction - Valence Labs","isPartOf":{"@id":"https:\/\/www.valencelabs.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#primaryimage"},"image":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#primaryimage"},"thumbnailUrl":"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png","datePublished":"2024-03-18T17:59:19+00:00","dateModified":"2025-05-20T12:29:00+00:00","author":{"@id":"https:\/\/www.valencelabs.com\/#\/schema\/person\/1b232c7b905afbb857b4eccc92cc87d0"},"breadcrumb":{"@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#primaryimage","url":"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png","contentUrl":"https:\/\/www.valencelabs.com\/wp-content\/uploads\/2025\/04\/Frame-2087325827.png","width":680,"height":791},{"@type":"BreadcrumbList","@id":"https:\/\/www.valencelabs.com\/introducing-molgps-a-foundational-gnn-for-molecular-property-prediction\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.valencelabs.com\/"},{"@type":"ListItem","position":2,"name":"Introducing MolGPS \u2014 A Foundational GNN for Molecular Property Prediction"}]},{"@type":"WebSite","@id":"https:\/\/www.valencelabs.com\/#website","url":"https:\/\/www.valencelabs.com\/","name":"Valence Labs","description":"Recursion\u2019s AI Research Engine","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.valencelabs.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":"Person","@id":"https:\/\/www.valencelabs.com\/#\/schema\/person\/1b232c7b905afbb857b4eccc92cc87d0","name":"Valence Labs","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/secure.gravatar.com\/avatar\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/417eefa5271ddcb42c8459f4bfb948e58ed4385707bd51ef5dda05a7da5d394f?s=96&d=mm&r=g","caption":"Valence Labs"},"url":"https:\/\/www.valencelabs.com\/fr\/author\/valencelabs\/"}]}},"_links":{"self":[{"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/posts\/62","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/comments?post=62"}],"version-history":[{"count":25,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/posts\/62\/revisions"}],"predecessor-version":[{"id":273,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/posts\/62\/revisions\/273"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/media\/133"}],"wp:attachment":[{"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/media?parent=62"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/categories?post=62"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.valencelabs.com\/fr\/wp-json\/wp\/v2\/tags?post=62"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}