<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by 310 AI on Medium]]></title>
        <description><![CDATA[Stories by 310 AI on Medium]]></description>
        <link>https://medium.com/@310AI?source=rss-6c9b84f0871a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*4I-gvPXCrGEn0tcbEZZpbw.jpeg</url>
            <title>Stories by 310 AI on Medium</title>
            <link>https://medium.com/@310AI?source=rss-6c9b84f0871a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 16:44:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@310AI/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Text2Enzyme: MP4 AI generates proteins across full EC universe]]></title>
            <link>https://medium.com/310-ai/text2enzyme-mp4-ai-generates-proteins-across-full-ec-universe-587fd212e6ea?source=rss-6c9b84f0871a------2</link>
            <guid isPermaLink="false">https://medium.com/p/587fd212e6ea</guid>
            <category><![CDATA[enzyme-engineering]]></category>
            <category><![CDATA[protein-design]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[310 AI]]></dc:creator>
            <pubDate>Tue, 01 Oct 2024 04:25:18 GMT</pubDate>
            <atom:updated>2024-10-01T16:58:55.945Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Written by </em><a href="https://medium.com/u/0f5f1e168625"><em>Oleg Matusovsky</em></a><em> and </em><a href="https://medium.com/u/066a35762034"><em>Kathy Y. Wei, Ph.D.</em></a><em> on Oct. 1 2024</em></p><ul><li>MP4 is a versatile <strong>AI molecule programming</strong> model that can also generate specific enzymes from text input.</li><li>The model has been used to build the EC repo, which showcases <strong>6,590 AI-generated protein</strong> sequences covering all seven enzyme classes, and includes 725 novel, potentially active enzymes. Explore <a href="https://310.ai/mp/ecrepo">here</a>.</li><li>Discover over 5,300 AI-generated <strong>enzymes binding</strong> with cofactors and substrates <a href="https://310.ai/docs/design/mp">here</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P0slEbAIGuciZa3syPz9AA.gif" /><figcaption>MP4 discovered enzymes of all classes</figcaption></figure><h3>From Your Medicine Cabinet to Your Cheese Plate: The Surprising Role of Enzymes in Everyday Life</h3><p>Enzymes are nature’s molecular machines, efficiently catalyzing a wide range of biological processes — even in minuscule amounts. These essential proteins work tirelessly behind the scenes, quietly orchestrating vital bodily functions from digesting our food to defending us against disease.</p><p>The Enzyme Commission (EC) system classifies all known enzymatic reactions. It consists of four levels of organization, from general (e.g., oxidoreductases) to specific (e.g., alcohol dehydrogenase). The system includes 7 first-level, 79 second-level, 320 third-level, and 8,269 fourth-level categories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fu3mswAGc0awTbQ5rO0e0g.png" /><figcaption>Enzymes of all classes are designed and produced to make consumer, commercial, and medical products</figcaption></figure><h3>Text2Enzyme: Transforming Enzyme Design with an AI Molecule Programming Model</h3><p>Molecule Programming model version 4 (MP4) revolutionizes protein design by using simple natural language prompts to generate sequences with specific functions. MP4 tackles one of the three major challenges in protein science: programmability. While significant progress has been made in solving the protein folding problem (predicting a protein’s structure from its sequence) and the inverse folding problem (designing sequences to achieve a specific structure) with solutions like AlphaFold3 and ProteinMPNN, respectively, the challenge of programmability — designing proteins to perform specific functions — remains largely unsolved. MP4 is trained using approximately 3,800 AMD-Instinct GPU-days.</p><p>MP4 has been used at scale to generate a collection of &gt;6,500 enzymes, encompassing 100% of top-level, 87% of second-level, and 73% of third-level Enzyme Commission (EC) space. This comprehensive collection is freely accessible for exploration <a href="https://310.ai/mp/ecrepo">here</a>. EC repo features predicted structures for all the enzymes generated by the text2enzyme model, including 5,306 structures docked with suitable ligands (cofactors and substrates) and 781 structures with predicted binding sites.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qEkn947nqTAyiEduSMTFbw.jpeg" /><figcaption>MP4 generates hundreds of novel enzymes across the enzyme space</figcaption></figure><p>The AI-generated enzymes in the EC repo show high sequence and structural quality as well as good functionality indicators. The repository includes novel proteins with low sequence similarity to known proteins and predicted to belong to the EC class that the input prompt specified. Predicted structural models of the text2enzyme sequences show a blend of helices and sheets, indicating that the MP4 model is capable of designing varying secondary structures. More than 80% of the enzymes have been docked with appropriate ligands (including coenzymes and substrates), and the ligand binding site of an additional ~12% have been predicted.</p><h3>EC1 Oxidoreductase: Electrical experts</h3><p>Oxidoreductases are the electrical powerhouses of the cell, facilitating the transfer of electrons between molecules. Imagine them as conduits that pass energy down a chain. By driving these essential reactions, oxidoreductases are crucial for cellular respiration and metabolism, ensuring that our energy levels stay charged and operational.</p><blockquote><em>Glucose Oxidase (EC 1.1.3.4), produced by Novozyme under the brand name Gluzyme® Fortis, is extensively utilized across various industries. It is especially beneficial in food production, where it enhances dough stability and improves dough-handling characteristics.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DvJbHU-76ZVfGAsaKWn8JA.gif" /><figcaption>A novel AI-generated oxidoreductase [yellow] docked with NADH compared to a natural enzyme [gray]</figcaption></figure><p>An EC1 example from EC repo is <a href="https://310.ai/mp/ecrepo/d3bd62ae-4680-4f0e-915d-c7f737808c52">M1Y7K</a>, an alcohol dehydrogenase (yellow cartoon) which is a fairly novel protein sequence with a <em>seqdif </em>score of 54. The <em>seqdif</em>, a metric for evaluating sequence novelty, is a derivative of the percentage identical matches (pident) to known protein sequences. Values of <em>seqdif</em> (defined as 100 — pident) above 50 indicate sequence novelty. The protein also has a very realistic structure with a very low <em>structdif </em>score of 4.6, which translates to a <em>tmscore </em>of 0.954 (<em>structdif </em>is defined as 100 — <em>tmscore </em>* 100). The <em>nlmsim </em>score, a measure of text-to-text similarity between the input prompt and ProtNLM function prediction, is also high, indicating that the predicted function highly correlates to the input prompt. NADH docked to the AI-generated enzyme binds to the same pocket as in the experimental NADH-bound dehydrogenase structure (gray cartoon), indicating the reliability of the enzyme generated by MP4.</p><h3>EC2 Transferase: Swap specialists</h3><p>Transferases are like molecular couriers, swapping functional groups between molecules. This transfer process is essential for many biological functions, including the construction of new proteins and the detoxification of harmful substances.</p><blockquote><em>Alanine Aminotransferase (EC 2.6.1.2), produced by Roche Diagnostics under the brand name ALT, is used in clinical diagnostics to measure alanine transaminase levels in human blood. It is widely employed in hospitals for diagnosing liver diseases such as hepatitis and cirrhosis. In 2023, Roche reported global sales of diagnostic tests, including ALT, totaling approximately $17B.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9q3BtGDM37Z0E5Gp8TJolg.gif" /><figcaption>A novel text2enzyme kinase [green] docked with its substrate and compared with a known enzyme [gray]</figcaption></figure><p>An interesting example is <a href="https://310.ai/mp/ecrepo/97aef535-ba4f-4fb3-bf78-741e08ba19e4">MVEK7</a>, a novel EC 2.7.4.2 transferase (green cartoon) with a <em>seqdif </em>score of 52.1 is shown. The enzyme is a phosphomevalonate kinase which has its substrate (PMV) docked in the same pocket as a crystallographic PMV-bound PMV kinase (gray cartoon). While the sequence is novel, the <em>structdif </em>score shows that the structure model is very similar to the naturally-occurring kinase, highlighting the potential of the AI-generated enzyme for the target function.</p><h3>EC3 Hydrolase: Water-driven breakup agents</h3><p>Hydrolases act as the cellular cleanup team, breaking down large molecules by adding water. This process is essential for digestion, as hydrolases help decompose proteins, fats, and carbohydrates into smaller, absorbable components.</p><blockquote><em>Lipase (EC 3.1.1.3), produced by Novozymes under the brand name Lipozyme® TL IM, breaks down fats into fatty acids and glycerol. This enzyme is primarily used in biodiesel production and food processing. In 2023, it generated annual revenue of approximately $1B and was utilized in 25% of biodiesel production.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5j09P8arg9RGbFYW_mc14Q.gif" /><figcaption>Example of a text2enzyme result [violet] docked with GTP and compared with a natural enzyme [gray]</figcaption></figure><p>One example for the hydrolase class in the EC repo is <a href="https://310.ai/mp/ecrepo/b7d0384c-4c0c-48fa-8cf8-040d7750bb9e">MFT5G</a>, a high quality EC 3.6.5.n1 AI-generated elongation factor 4/LepA (violet cartoon). The AI enzyme was docked with GTP and compared with a natural, structurally similar (<em>structdif</em> 5.9) GDP-bound LepA (gray cartoon). As shown, the nucleotide binds to the same pocket as the natural enzyme, highlighting the potential of the enzyme generated by the molecule programming model.</p><h3>EC4 Lyase: Waterless bond-breaking experts</h3><p>Lyases are employed in the production of food, textiles, paper, and pharmaceuticals. These enzymes break down molecules without the use of water, often resulting in the formation of simpler molecules or new structures.</p><blockquote><em>Recombinant L-Phenylalanine Ammonia-Lyase (PAL) (EC 4.3.1.24), produced by BioMarin Pharmaceutical Inc. under the brand name Palynziq, is used to break down excess phenylalanine in the treatment of phenylketonuria (PKU). This product generated approximately $304M in revenue in 2023.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8TWkXXboFJTeIP_x12EvmQ.gif" /><figcaption>A novel AI-generated lyase [magenta] docked with its product and compared to a close natural lyase [gray]</figcaption></figure><p>Example <a href="https://310.ai/mp/ecrepo/b0a5bc69-6323-4f5d-8934-6e5458a4ed2d">MC0CE</a> is an EC 4.2.99.20 lyase (magenta cartoon) from the EC repo, and is a novel text2enzyme SHCHC lyase. It belongs to the synthase subclass of lyases. The enzyme’s product docked in the same pocket as with the crystallographic SHCHC-complex structure (gray cartoon).</p><h3>EC5 Isomerase: Shape-shifters</h3><p>Isomerases rearrange the structure of molecules, transforming them into different forms, or isomers, without adding or removing components. This ability to rearrange atoms is crucial for converting molecules into forms that cells can more effectively utilize.</p><blockquote><em>Glucose Isomerase (EC 5.3.1.5), produced by Novozymes under the brand name Sweetzyme®, catalyzes the conversion of glucose to fructose. This process is crucial for producing high-fructose corn syrup, a widely used ingredient in many foods and beverages.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qhTLTRjUJptDyRaPyglXmg.gif" /><figcaption>A novel AI-generated isomerase [dark blue] docked with substrate and compared with a close natural enzyme</figcaption></figure><p>For the EC5 class, a novel text2enzyme EC 5.1.1.13 isomerase, <a href="https://310.ai/mp/ecrepo/cfbbd47e-38e3-41a5-ac5e-a9d1ee2e7b67">MJZMP</a>, designed to catalyze the D-to-L-aspartate interconversion is featured (purple cartoon). The AI-generated aspartate racemase was docked with aspartate and compared with a close natural glutamate-bound aspartate/glutamate racemase (gray cartoon). The functional score <em>nlmsim </em>90 and the high structural similarity (<em>structdif</em> 9.4) to the known racemase with the same function indicate the potential of the novel enzyme for the target function.</p><h3>EC6 Ligase: Bonding pros</h3><p>Ligases are the master builders of the cell, responsible for joining two molecules together. They act like molecular glue, facilitating the bonding of everything from DNA strands to proteins. Ligases are also essential for DNA replication and repair, as they join DNA molecules by forming bonds between their ends.</p><blockquote><em>DNA Ligase (EC 6.5.1.1), produced by QIAGEN under the brand name T4 DNA Ligase, is essential in various biotech applications such as next-generation sequencing, molecular cloning, and DNA amplification. These processes are crucial for genetic testing related to ancestry and health. In 2023, this product generated approximately $1.2B in revenue.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ul2e3BV2laIS82Ki1YmJ1Q.gif" /><figcaption>An example AI-generated ligase [pink] docked with ATP and compared to a similar natural ligase [gray]</figcaption></figure><p>An example from the MP4-generated ligases is the featured <a href="https://310.ai/mp/ecrepo/60f65939-2b81-44a3-bc3d-79b3375dc8fb">MY7V3</a>, which is an EC 6.3.2.4 ligase that shows decent sequence divergence (<em>seqdif </em>46.6) from the closest natural protein sequence. The ATP docked to the AI-generated D-ala — D-ala ligase (pink cartoon) binds to the same pocket and in a similar conformation as compared to an experimental structure of an ATP-bound natural D-ala — D-ala ligase (gray cartoon).</p><h3>EC7 Translocase: Traffic controllers</h3><p>Translocases are essential for moving molecules, ions, amino acids, and proteins across cell membranes, ensuring the crucial separation between the inside and outside of cells. Without them, this separation would be disrupted.</p><blockquote><em>Fe3+-Transporting ATPase (EC 7.2.2.7), produced by Novartis under the brand name Jadenu®, is used in chelation therapy to remove excess iron from the body by suppressing the activity of Fe3+-transporting ATPases. This product generated approximately $870M in revenue in 2023.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iPLlIJ2sqtDZ7Epv480W5g.gif" /><figcaption>Example text2enzyme ATP-docked translocase [brown] compared to a known similar translocase [gray]</figcaption></figure><p>The example AI-generated translocase <a href="https://310.ai/mp/ecrepo/0249c3be-4b65-4b13-aeff-9c5ed4ad0c01">MM131</a> (orange cartoon) is an ABC-type transporter with a <em>seqdif</em> score of 46.7, showing high structural similarity (<em>structdif</em> 15.6) to a naturally-occurring ABC-type transporter. The potential of the AI-generated enzyme is also supported by the high protNLM-based function prediction metric (<em>nlmsim </em>90). Compared to an ATP-bound crystal structure of a known, close natural transporter (gray cartoon), ATP docks in the same pocket and similar conformation.</p><h3>Beyond enzymes</h3><p>While enzymes are central, there is much more to what proteins can do for us. And while natural enzymes have been the starting point for all of our useful products so far, with the the rise of AI, a new era of machine made solutions is on the horizon.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GVP-dNceA6U43fg92Qs_yQ.jpeg" /></figure><h3>References</h3><ol><li><a href="https://310.ai/docs/">310 Copilot documentation</a></li><li><a href="https://310.ai/docs/design/mp">310 MP4 documentation</a></li><li>Merck Full-Year 2023 Financial Results</li><li>The Novozymes Report 2023</li><li>Johnson, M.E. (2017). “A 100-Year Review: Cheese production and quality”. Journal of Dairy Science. 100 (12): 9952–9965. doi:10.3168/jds.2017–12979</li><li>BioMarin Pharmaceutical Inc. 2023 Financial Report</li><li>QIAGEN Annual Report 2023</li><li>Novartis Annual Reports</li><li>Roche Financial Report 2023</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=587fd212e6ea" width="1" height="1" alt=""><hr><p><a href="https://medium.com/310-ai/text2enzyme-mp4-ai-generates-proteins-across-full-ec-universe-587fd212e6ea">Text2Enzyme: MP4 AI generates proteins across full EC universe</a> was originally published in <a href="https://medium.com/310-ai">310 AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MP4: AI Text2Protein Breakthrough Tackles the Molecule Programming Challenge]]></title>
            <link>https://medium.com/310-ai/mpm4-ai-text2protein-breakthrough-tackles-the-molecule-programming-challenge-870045a8c1ad?source=rss-6c9b84f0871a------2</link>
            <guid isPermaLink="false">https://medium.com/p/870045a8c1ad</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[molecule]]></category>
            <category><![CDATA[protein]]></category>
            <category><![CDATA[language-model]]></category>
            <dc:creator><![CDATA[310 AI]]></dc:creator>
            <pubDate>Tue, 16 Jul 2024 02:23:04 GMT</pubDate>
            <atom:updated>2024-09-30T22:56:26.567Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Written by </em><a href="https://medium.com/u/066a35762034"><em>Kathy Y. Wei, Ph.D.</em></a><em> and </em><a href="https://medium.com/u/c566b30abf98"><em>Koosh</em></a><em> on 16 July 2024</em></p><ul><li><a href="http://310.ai/">310</a> AI’s MP4 model compresses billions of years of life into a powerful AI that transforms text prompts into novel and useful protein sequences.</li><li>Originally developed in 2021, this breakthrough could revolutionize the application of biology in medicine, industry, and beyond.</li><li>Explore molecule programming with the <a href="https://310.ai/mpm/repo">MP4 repo</a>, featuring over 1,000+ AI-generated, scientist-approved proteins.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3hABUN3YVsc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3hABUN3YVsc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3hABUN3YVsc%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2b213517e474c07f54fcd478981e854d/href">https://medium.com/media/2b213517e474c07f54fcd478981e854d/href</a></iframe><p>Protein design and engineering focus on manipulating proteins to achieve specific functions — either by creating entirely new proteins from scratch (design) or by modifying existing ones (engineering). Understanding how to make protein functions controllable and designable has intrigued scientists ever since we first gained the ability to analyze their shapes.</p><p>In a major step towards a new chapter in protein sciences, 310 AI’s<strong> m</strong>olecule<strong> p</strong>rogramming<strong> </strong>model<strong> </strong>version 4 <strong>(MP4) </strong>is able to generate protein sequences from text prompts describing the desired function. The earliest version of MP was built in 2021, featured in <a href="https://www.youtube.com/watch?t=329&amp;v=DiGB5uAYKAg&amp;feature=youtu.be">2023</a>, and has grown significantly in capabilities. With just a simple sentence, it’s now possible to invent de novo sequences that not only match the desired function but also deliver high quality, excellent foldability, and uniqueness.</p><p>Many of the world’s greatest opportunities are fundamentally linked to protein functions. A fast, accurate way to create proteins with desired functions will revolutionize:</p><ul><li><em>Medicine</em>: Design drugs for currently undruggable disease targets.</li><li><em>Industry</em>: Develop high-value chemicals and materials with non-toxic, sustainable, environmentally-friendly processes.</li><li><em>Environment</em>: Deploy enzymes or whole microorganisms for environmental cleanup or nutrients recycling.</li><li><em>Research</em>: Discover tools to tackle currently unanswerable questions.</li></ul><p>In recent years, there has been a fervent focus on leveraging AI in protein sciences. The extraordinary success of AlphaFold has set the stage for a wave of breakthroughs in this dynamic field.</p><h3>The Programmability Challenge</h3><p>Protein engineering has long been closely linked to healthcare. Since the introduction of Humulin, a <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3714061/">recombinant insulin</a> for diabetes treatment in 1982, scientists have been eager to explore how modified proteins can lead to even better disease treatment. Fast forward to 2021, when <a href="https://www.sciencedirect.com/science/article/pii/S000649712106465X">NL-201</a>, the first protein designed from scratch on a computer, entered clinical trials.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o7XccqnhMGixsMQw79XTXQ.png" /></figure><p><a href="https://www.nature.com/articles/s41586-021-03819-2">AlphaFold</a> and <a href="https://www.nature.com/articles/s41586-024-07487-w">AlphaFold3</a> represent significant advances in the folding problem (sequence &gt; structure), while others, including <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9997061/">ProteinMPNN</a> and <a href="https://www.biorxiv.org/content/10.1101/2024.07.01.600583v1">ESM3</a> address the inverse folding problem (structure &gt; sequence). However, none have comprehensively tackled the programming problem (function &gt; sequence) as MP4 has.</p><p>Despite these advancements, most protein design technologies — both experimental and computational — have relied on searching increasingly larger portions of sequence space. This approach is inherently unsustainable; for a typical protein of 300 amino acids with 20 possible amino acids, the number of potential proteins is 20³⁰⁰, vastly exceeding the number of atoms in the universe. This is where modern AI techniques offer the potential to overcome the limitations of traditional search-based methods. While 20³⁰⁰ is an enormous search space, it’s roughly equivalent to the number of options in a 98-word English paragraph. Modern natural language AI models can easily generate such text with high coherence and readability.</p><h3>A Repository of Text2Protein AI Results</h3><p>At 310 AI, we can take a text prompt such as “Ribosomal protein L33 contains a conserved site and is located in the chloroplast” and generate a full protein sequence such as “MAK GKD ARV TVI LEC TSC ERN GVN KKS TGI SRY ITQ KNR HNT PGR LEL RKF CPY CRK HTI HGE IKK”, using our proprietary molecule programming system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Nj0y-tDmMS1z3sup.gif" /><figcaption>Examples of text2protein generation</figcaption></figure><p>In particular, freeform human-readable specifications of a protein are processed into a proprietary protein language with a vocabulary that describes functions, properties, processes, families, domains, active sites, motifs, taxonomy, and more. While only a fraction of the vast freeform text protein programming space is understood by MP4, it is already successfully generating exciting examples from diverse and complex prompts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vEPahjTNz6gh0sdg6dc1Nw.jpeg" /><figcaption>Overview of repo functions</figcaption></figure><p>We are releasing our <a href="https://310.ai/mpm/repo">MP4 repo</a>, featuring an initial 1000+ AI-generated sequences. More will be added soon!</p><h3>Creating Quality Molecules</h3><p>One of the major challenges in protein design and engineering is the lack of simple metrics to assess the “goodness” of protein sequence, structure, or function. The most well-developed metrics focus on protein structure, applicable to both experimental and computational results. Here, we use three main metrics to evaluate sequence, structure, and function, respectively.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uG9lxbazuggSAZ9Iv1Rh5w.png" /></figure><ul><li><a href="https://www.sciencedirect.com/topics/immunology-and-microbiology/amino-acid-composition">aacomp</a> (amino acid composition): Measures the distribution of each amino acid type in a sequence compared to UniProt sequences, with natural sequences scoring between 80–100. While this metric doesn’t consider the order of amino acids, it helps filter out overly repetitive sequences.</li><li><a href="https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/plddt-understanding-local-confidence/">plddt</a> (average predicted local distance difference test): Calculates the average per-residue local confidence for structure prediction (using <a href="https://www.biorxiv.org/content/10.1101/2022.07.20.500902v1">ESMFold</a>). A well-defined structure typically scores between 70–100. In many applications, a more structured protein is preferable.</li><li>nlmsim (ProtNLM function prediction similarity score): A text-based comparison between the input prompt and the ProtNLM output. Scores of 80–100 indicate exact or subset matches, while scores of 60–80 reflect similar-looking words. Note that this measure may miss synonyms or specific examples within a broader category, leading to false negatives.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6GJBK3N4GFMfLzwSS8dSEw.png" /></figure><p>The functional comparison, or nlmsim metric, is of most interest, as this is the self-consistency measure of how well protein sequences generated by MP4 align with the specified prompt. However, due to limitations in available protein function predictors and the challenges of text-to-text comparisons, this quality is the hardest to measure accurately. We look forward to improvements in the accuracy and standardization of this evaluation in the near future.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WErHoIeNf9UoNp3W7Wpvow.png" /><figcaption>Overview of repo quality</figcaption></figure><h3>Unlocking the World Beyond Nature</h3><p>In addition to generating “good” proteins, we are also interested in creating novel ones. Novel proteins are appealing because they can achieve functions or properties that are not constrained by the forces of natural selection acting on natural proteins. For example, natural proteins rarely encounter conditions of extremely high concentration, while high concentrations are often necessary for medications to be effective. In addition, novel proteins are intriguing from a research perspective, as they expand our understanding of what these molecules can achieve, and may refine our notions of their potential applications. Here, we use two standard metrics to evaluate sequence novelty and structure novelty.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TcqoCOWUZYJyGn3Yyz-HYg.png" /></figure><ul><li>seqdif: Sequence novelty score derived from the percentage of identical matches (<a href="https://www.ncbi.nlm.nih.gov/books/NBK279684/table/appendices.T.options_common_to_all_blast/">pident</a>), which assesses sequence similarity between an input query and sequences in the <a href="https://www.nlm.nih.gov/ncbi/workshops/2023-08_BLAST_evol/databases.html">NR/NT database</a> using <a href="https://www.nature.com/articles/s41592-021-01101-x">DIAMOND</a> (a faster alternative to <a href="https://blast.ncbi.nlm.nih.gov/Blast.cgi?PROGRAM=blastp&amp;PAGE_TYPE=BlastSearch&amp;LINK_LOC=blasthome">BLASTP</a>). We define seqdif as 100 — pident for consistency, with a cutoff above 50 indicating a novel sequence. Although this threshold is not universally established, it is commonly used in de novo design. Note that this metric is sensitive to parameter settings.</li><li>structdif: Structure novelty score based on structural similarity (<a href="https://zhanggroup.org/TM-score/">tmscore</a>) between an input query and structures in the Protein Data Bank (<a href="https://www.rcsb.org/?ref=nav_home">PDB</a>) via <a href="https://www.nature.com/articles/s41587-023-01773-0">Foldseek</a>. We calculate structdif as 100 — tmscore * 100 for consistency, with a cutoff above 50 indicating a novel structure, which is a standard threshold.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UQlb6Cj8eOCnU9-J.gif" /><figcaption>Examples that novel both by sequence and structure</figcaption></figure><p>We have found many notable examples of sequence and structural novelty, demonstrating that MP4 can explore beyond natural limits.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8r0Rfpy2T3XI8A-Pf-pOlw.png" /><figcaption>Overview of repo sequence and structural novelty</figcaption></figure><h3>The Generalist Approach</h3><p>While traditional biological research often emphasizes reductionism and narrow focus, a generalist approach has proved more successful in AI. Models concentrating on specific enzyme types or properties struggle to generalize to new inputs — an essential requirement for most applications. By developing models that encompass a wide range of potential inputs, we enhance the likelihood that our desired generalizations fall within the model’s input space. This allows for meaningful extrapolation, whereas attempts to generalize outside this space are typically unsuccessful.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RMpxeI51TCOGUGFKg-5fHA.jpeg" /></figure><p>We are seeing an explosion in the use of natural language text driven by AI technology, and we predict similar adoption in the scientific community. While the MP4 vocabulary has limitations, it is poised for continuous growth in size, scope, and complexity as more data — both public and private — are compressed and optimized.</p><p>MP4 is trained using 138K tokens and 3.2B unique datapoints across 70 synchronized tasks. This foundational model prioritizes molecule programmability and achieves state-of-the-art results with approximately 3,800 AMD-Instinct GPU-days.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X3tave5zAeVpe4ZU.png" /><figcaption>MP4 architecture</figcaption></figure><h3>The Next Horizon</h3><p>MP4 represents a thrilling leap in making molecules programmable, but this is just the beginning. We’re on the brink of adding groundbreaking new controllers that will reshape our understanding of proteins, DNA, RNA, and small molecules — unlocking both natural and novel functions. Together with our collaborators, we’re diving into specific application areas where AI-driven creations can catalyze disruption in health, environment, agriculture, and more.</p><p>Just a few years ago, AI felt like science fiction. Today, it’s at the forefront of innovation. Tomorrow, AI will reshape the future of biology!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Qw4gtA45suLArVALstp9-Q.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=870045a8c1ad" width="1" height="1" alt=""><hr><p><a href="https://medium.com/310-ai/mpm4-ai-text2protein-breakthrough-tackles-the-molecule-programming-challenge-870045a8c1ad">MP4: AI Text2Protein Breakthrough Tackles the Molecule Programming Challenge</a> was originally published in <a href="https://medium.com/310-ai">310 AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>