<?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 Lynakhermane on Medium]]></title>
        <description><![CDATA[Stories by Lynakhermane on Medium]]></description>
        <link>https://medium.com/@lynakhermane2?source=rss-f28a973ae43f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*QR-SgJLSvx2Xpfll</url>
            <title>Stories by Lynakhermane on Medium</title>
            <link>https://medium.com/@lynakhermane2?source=rss-f28a973ae43f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 21:08:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lynakhermane2/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[Automated Neoantigen Discovery in Modern Immunotherapy]]></title>
            <link>https://medium.com/@lynakhermane2/automated-neoantigen-discovery-in-modern-immunotherapy-34a77587023c?source=rss-f28a973ae43f------2</link>
            <guid isPermaLink="false">https://medium.com/p/34a77587023c</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Lynakhermane]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 23:24:57 GMT</pubDate>
            <atom:updated>2026-11-12T12:44:16.192Z</atom:updated>
            <content:encoded><![CDATA[<p>TL;DR:</p><p>Personalized cancer vaccines rely on identifying tumor neoantigens that trigger real immune responses. While tools like NetMHCpan quickly filter basic binders, advanced deep learning models like DeepNeo-v2 bridge the gap by predicting actual T-cell activation. Together, combining these pipelines offers a powerful strategy for precision immunotherapy.</p><p><strong>Abstract</strong>:</p><p>About 9,6 to 10 million people die from cancer each year (or approximately 1 out of six deaths worldwide), making cancer an unrelenting biological burden. But can we use modern technology or AI to solve this problem?</p><p>To realize this potential, neoantigen discovery serves as a vital pathway. By integrating advanced computational frameworks , specifically convolutional neural network (CNN) architectures inspired by models like DeepNeo. This article explains how tools like pVACtools and DeepNeo rank neoantigens. The proposed pipeline optimizes the translation of raw peptide sequences into precise immunogenicity predictions, offering a robust computational workflow to enhance personalized cancer immunotherapy and bridge the gap between machine learning and translational oncology.</p><blockquote>Author: Lyna Khermane</blockquote><blockquote>Thesis question: How are Deep Learning models help design customised cancer vaccines?</blockquote><blockquote>Content:</blockquote><blockquote>1.. The shift from chemotherapy to immunotherapy</blockquote><blockquote>2. The followed pipeline “ Neoantigen discovery”</blockquote><blockquote>3. From simple ranking to AI powered selection</blockquote><blockquote>4. DeepNeo vs NetMHCpan</blockquote><blockquote>5. Putting it together: “Automated cancer vaccine” pipeline</blockquote><blockquote>6. The clinical reality: promise, progress and remaining gaps</blockquote><blockquote>7. Future Horizons: What AI will unlock next in immunotherapy</blockquote><h3>1. The shift from chemotherapy to immunotherapy:</h3><p>Cancer is a mutational breakdown from within, where DNA damage causes normal cells to divide uncontrollably and hide from your immune system because they originate from the body’s own tissue. While traditional treatments like chemotherapy kill fast-dividing cells, the drugs travel throughout the whole body and target healthy cells, causing side effects like hair loss or fatigue. More importantly, chemotherapy can fail because cancer cells quickly mutate, build resistance, and survive to grow back stronger.</p><p>This is why medicine has shifted toward immunotherapy. Unlike chemotherapy, immunotherapy acts like a coach that trains your body’s own living white blood cells to hunt down only the bad cells. Once your immune system learns what the cancer looks like, it saves that memory forever to destroy recurring cells instantly, using living immune cells that adapt as the cancer changes. The engine driving this entire process is a targeted pipeline called Neoantigen discovery.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/549/1*EyGn0-6vqf6RksMMaBhLYA@2x.jpeg" /><figcaption>https://www.asbestos.com/treatment/immunotherapy-vs-chemotherapy/</figcaption></figure><h3>2. The followed pipeline “Neoantigen discovery”:</h3><p>Think of your body as a large ecosystem made up of billions of normal, healthy cells wearing specific molecular markers that show they belong, while cancer begins when healthy cells develop genetic mutations from multiplying and mutating rapidly to produce abnormal, mutated proteins called neoantigens that act as unique flags distinguishing diseased cells from normal ones.</p><p>Neoantigen discovery is the process of hunting down these specific, unique mutant flags so we can train a patient’s immune system to attack only the cancer.</p><p>The process generally works in three steps:</p><h4>Sequencing ( the detective work &amp; NGS ) :</h4><p>Scientists start by taking a biological sample of both the cancerous tumor biopsy and the patient’s healthy tissue, chopping their DNA or RNA into millions of tiny, manageable pieces.</p><p>Tiny chemical adapters are attached to these fragments, and they are copied millions of times so the machine can clearly detect them. The prepared sample is then loaded onto an NGS machine for massive parallel sequencing, where instead of reading one long DNA strand sequentially, the machine reads millions of fragments simultaneously.</p><p>Finally, computational assembly uses computer algorithms to stitch all the million tiny pieces back together by matching their overlapping letters. By comparing the complete genetic sequence of the tumor against the healthy tissue dataset, scientists pinpoint the exact somatic mutations (unique genetic alterations) that distinguish the cancer cells from normal ones.</p><p>Sources [5,6,7]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hWAlP8vLUtmJGIotLAHkIw.jpeg" /><figcaption><a href="https://genomemedicine.biomedcentral.com/articles/10.1186/gm524/figures/1">https://genomemedicine.biomedcentral.com/articles/10.1186/gm524/figures/1</a></figcaption></figure><h4>In-Silico Binding and Presentation Predictions</h4><p>Once scientists have identified mutated peptides from cancer cells, complex AI models determine their biological presentation potential. For a mutation to trigger an immune response, it has to be chopped up, transported, and successfully displayed on the cell surface like a flag. Computers simulate this entire journey step by step:</p><ul><li>Chopping the Protein: Inside a cell, proteins are constantly cut into smaller pieces by natural cellular machinery. AI models simulate this chopping process to predict whether a mutated protein will naturally break apart in a way that keeps the unique mutation intact.</li><li>The Transport Check: Once chopped into small pieces, these fragments must be moved from the inside of the cell to the right compartment where surface displays are assembled. The computer algorithms estimate how easily these mutated fragments can make it to the surface.</li><li>The Display Tray Fit: Using AI models like NetMHCpan, scientists calculate a binding score to predict how securely the fragment will lock onto the patient’s HLA display trays for the immune system to spot.</li></ul><p>Sources [8,9,10]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*Pq2G4SkR8YBsv0by9cjSmw.png" /><figcaption><a href="https://commons.wikimedia.org/wiki/File:MHC_Class_I_processing.svg">https://commons.wikimedia.org/wiki/File:MHC_Class_I_processing.svg</a></figcaption></figure><h4>Vaccine Candidate Prioritization and Filtering :</h4><p>Just because a peptide binds to an HLA molecule doesn’t make it a great vaccine target. Scientists run candidates through a prioritization framework using key biological features to select the final elite candidates:</p><ul><li>Gene Expression: Evaluates ifthe cancer cell produces enough of this mutated protein, low expression won’t trigger an effective response.</li><li>Clonal Status: Prioritizes mutations present in every cancer cell (clonal) rather than a small patch of cells (subclonal). Vaccines prioritize clonal targets so the cancer cannot easily mutate and escape destruction.</li><li>Sequence Divergence: Assesses how structurally different the mutated peptide is from normal proteins. The more abnormal it looks to the immune system, the stronger the T cell attacks.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mevicEnZOlMBd3e3wBgiOQ.png" /><figcaption><a href="https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.836821/full">https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.836821/full</a></figcaption></figure><h3>3. From simple ranking to AI‑powered selection:</h3><p>Here is where the real integration of AI is , scientists can’t do all the work manually, it would take too much time but also missing a lot of potential candidates making the customised vaccine more theoretical than practical, so they use computational tools, like pVACtools which is a framework that tells the computer the rules it should follow to predict the best candidates, saving scientists a significant amount of time.</p><p>However, it only automates heavy data processing workflows, relying on deterministic, rule based filtering and pre-existing scoring formulas, which keeps it from generating better results</p><p>To overcome the rigid rules traditionale pipelines, modern ones replace the manual work with specialized neural networks trained directly on on biological sequence data.</p><p>Tools like NetMHCpan integrate Artificial Neural Networks trained on vast binding and mass spectrometry datasets, mapping peptide sequence features to physical MHC binding pockets to predict presentation affinity.</p><p>As for DeepNeo, it expands this neural architecture even further into a two branch deep learning pipeline: DeepNeo-MHC uses a hybrid CNN and Bi-LSTM network to model presentation dynamics, while DeepNeo-TCR employs a cross attention transformer to evaluate T-cell receptor engagement.</p><p>By using data driven probability scores, these neural networks can make neoantigen selection a high precision predictive engine</p><p>Sources[11,12,13]</p><h4>How DeepNeo works</h4><p>Predicting if a peptide will successfully trigger an immune response require to go through two steps, will a peptide bind tightly to the MHC molecule and display on the cell’s surface? and will a TCR receptor recognize it as a foreign and attack it?</p><p>Previous tools like NetMHCpan only solved the first half. But DeepNeo solves both by utilizing a specialized dual-branch prioritization framework:</p><blockquote>DeepNeo‑MHC,</blockquote><p>which employs a hybrid 1D convolutional Neural Network paired with a Bidirectionale Long Short-Term Memory network (Bi-LSTM). The CNN layers scans small groups of amino acids side by side to catch specific local patterns like the exact shape needed to fit into the pocket of the body’s MHC cell marker</p><p>The Bi-LSTM reads the amino acid chain forwards and backwards instead of just looking at small clusters. This helps it understand how distant parts of the chain affect each other, capturing the full context of the sequence.</p><blockquote>DeepNeo‑TCR,</blockquote><p>It employs a Cross-Attention Transformer which shines a spotlight on how specific parts of that complex interact directly with the T-cell receptor ( the immune system ultimate trigger).</p><p>By computing multi-head attention weights, one checks the physical shape, another checks the chemical charge, and a third checks flexibility. By looking at all these connections simultaneously across the sequence, it sees the complete picture of how they fit together.</p><p>This transformer layer makes sure the chosen peptides are not false alarms and actually wake the immune system up.</p><h4>Exact training datastes &amp; Benchmark metrics</h4><p>When DeepNeo first came out in February 2020, it introduced a dual-model framework that predicts both MHC binding and T-cell reactivity; however, it was initially limited to human MHC Class I alleles and built on legacy Python 2.7 and Theano libraries.</p><p>By April 2023, the researchers released DeepNeo-v2, transforming the prototype into a comprehensive, high-throughput web server.</p><p>Beyond migrating the codebase to modern Keras and Python 3.7 frameworks and replacing Negative Log-Likelihood with Binary Cross-Entropy loss for numerical stability, v2 expanded coverage to include both human and mouse MHC Class I and Class II alleles.</p><p>Crucially, v2 scaled its MHC binding dataset by 7.8-fold and fixed the “dying ReLU” problem in its T-cell reactivity convolutional layers by switching to linear activation functions, preventing small filter matrices from zeroing out critical amino acid interaction weights.</p><h4>1.Data Scale &amp; Expansion</h4><blockquote>DeepNeo-mhc (Presentation Module):</blockquote><ul><li>MHC Class I: Scaled to 340,987 samples – a 7.8-fold expansion over DeepNeo-v1’s 57,173 samples. The positive instance fraction was adjusted to 75.0% to reflect presentation likelihood.</li><li>MHC Class II: Added a dedicated dataset of 106,908 samples (57.2% positive) to extend prediction capabilities to CD4+ helper T-cell epitopes alongside CD8+ targets.</li></ul><blockquote>DeepNeo-tcr (Immunogenicity Module):</blockquote><ul><li>MHC Class I: Expanded to 19,149 samples (42.6% positive). To mitigate severe class imbalance, all confirmed positive reactive pairs (N = 8,148) were paired with 11,001 randomly selected negative pairs.</li><li>MHC Class II: Expanded to 12,913 samples (42.4% positive) by incorporating negative binding pairs to balance the positive recognition set.</li></ul><h4>2. Strict Biological Filtering Criteria</h4><p>To reduce noise from heterogeneous experimental setups, the dataset was constrained by rigorous quality controls:</p><ul><li>Peptide Standardized Lengths: Fixed strictly at 9-mers for MHC Class I and 15-mers for MHC Class II.</li><li>Resolution: Restricted to entries with verified 4-digit HLA allele typing across human and mouse strains.</li><li>Assay Standardization for TCR: For DeepNeo-tcr, only biological experiments measuring Interferon-gamma (IFN-γ) release upon peptide-MHC challenge were included, establishing a clear gold standard for true functional T-cell activation over passive binding.</li><li>Conflict Resolution: If contradictory experimental outcomes existed across literature, the pair was conservatively labeled as positive if it demonstrated reactivity under at least one condition.</li></ul><h4>3. Benchmark Performance Metrics</h4><p>On independent test sets, the scaled training infrastructure yielded marked improvements over earlier versions and baseline predictors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/704/1*ClbRosSEmjiP-7TRfRNINw@2x.jpeg" /></figure><h3>4. DeepNeo vs NetMHCpan:</h3><blockquote>What Do They Measure?</blockquote><ul><li>NetMHCpan: Primarily predicts peptide-MHC binding affinity – it calculates how tightly a short piece of protein (peptide) will stick into an MHC molecule’s pocket.</li><li>DeepNeo: Goes a step further by evaluating T-cell immunogenicity and reactivity. It assesses whether a T-cell receptor (TCR) will actually recognize the complex and trigger an immune response (\text{IFN-}\gamma release).</li></ul><blockquote>Input Requirements &amp; Complexity</blockquote><ul><li>NetMHCpan: Requires the peptide sequence and the target HLA allele. It is exceptionally fast and optimized for high-throughput scanning of entire exomes to filter down thousands of mutated peptides to hundreds of potential binders.</li><li>DeepNeo: Requires more complex data pairing (matching peptide sequences, MHC alleles, and TCR characteristics) across its dual-model framework (DeepNeo-mhc and DeepNeo-tcr) to filter those binders down to true functional vaccine candidates.</li></ul><blockquote>Architecture &amp; Evolution</blockquote><ul><li>NetMHCpan: Built on artificial neural networks (ANNs) trained heavily on mass spectrometry peptidomics data and in vitro binding affinity measurements across thousands of HLA alleles.</li><li>DeepNeo: Relies on 2-layer Convolutional Neural Networks (CNNs) with specialized adjustments (like linear activation layers in DeepNeo-tcr to prevent the “dying ReLU” problem) specifically optimized to bridge presentation prediction with T-cell receptor recognition.</li></ul><blockquote>Where Each Fits in the Pipeline</blockquote><ul><li>NetMHCpan: Acts as the first-line filter in an mRNA vaccine pipeline – used early to discard thousands of non-binding mutations.</li><li>DeepNeo: Acts as the refining filter – applied later in the pipeline to weed out “false positives” that bind well to MHC markers but fail to activate actual T-cells.</li></ul><p>To reach the best results, scientists tend to use the intersection of DeepNeo and NetMHCpan, which means that they use NetMHCpan as a first filtering tool, then feed the resulting peptides into DeepNeo to take the best use for both tools.</p><h3>5. Putting It Together: The Modern Neoantigen Pipeline:</h3><p>Bringing these tools together creates a streamlined workflow for precision immunotherapy:</p><ul><li>Initial Screening: NetMHCpan acts as the first-line filter, rapidly scanning tumor mutations to screen out non-binders and isolate high-affinity peptide candidates.</li><li>Immunogenicity Validation: Those filtered peptides are then fed into DeepNeo to predict actual T-cell reactivity and rule out false positives.</li></ul><p>Ultimately, deep learning answers the biophysics question “Does this peptide bind the MHC pocket?”, while integrated toolsets handle the clinical validation “Will this trigger a real immune response?”. Together, they form an automated triage system that turns thousands of raw mutations into a precise, high-quality list of vaccine targets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*EvG0S4NTgh8LXnN5S5of_w@2x.jpeg" /></figure><h3>6. The Clinical Reality: Promise, Progress, and Remaining Gaps</h3><p>The fusion of next generation sequencing, deep learning binding predictors, and automated multi‑omics triage is not just a theoretical computational model , it is actively reshaping modern oncology.</p><p>Traditional cancer treatments like chemotherapy act as blunt instruments, attacking both cancerous and healthy cells and causing severe systemic toxicity. In contrast, AI‑driven neoantigen discovery shifts care toward a paradigm of extreme precision: instead of poisoning all rapidly dividing cells, it aims to train the immune system to recognize mutation‑specific “flags” that exist only on the tumor.</p><p>These technologies can achieve high clinical accuracy because they operate on a multi‑layered verification standard:</p><p>•	Cellular validation: By filtering candidates through binding %Rank, gene expression thresholds, and clonal frequency simultaneously, the pipeline ensures that selected neoantigens are genuinely present in the tumor, abundantly transcribed, and likely to be displayed on the cell surface.</p><p>•	Immune evasion avoidance: Machine learning based prioritization examines whether a mutation is clonal (present across most cancer cells) or subclonal (restricted to a subset), helping to select targets that are less likely to disappear as the tumor evolves.</p><p>•	Reduced trial and error: Automating candidate selection shortens the time from biopsy to vaccine design, making it more feasible to manufacture and administer customized mRNA or peptide vaccines while the disease is still manageable.</p><p>Landmark clinical developments ( such as positive late‑stage trial results for personalized mRNA cancer vaccines combined with immune checkpoint inhibitors ) show that targeting patient‑specific neoantigens can significantly improve recurrence free survival in certain cancers, turning theoretical immunology into real world benefit for some patients.</p><p>And yet, despite this progress, AI‑driven neoantigen pipelines are not yet routine in most hospitals, and scientists do not rely on deep learning models like DeepNeo as the final authority on vaccine design. The main reasons are not a lack of clever algorithms, but fundamental limitations in data, biology, and infrastructure:</p><p>•	Limited and biased training data: Models like DeepNeo learn from experimental datasets of peptide – MHC binding and T‑cell reactivity. Today, these datasets are still relatively small, skewed toward certain cancers (like melanoma) and common HLA types, and often pre‑filtered to “likely binders.” This limits how well the models generalize to new patients, rare HLA alleles, and under‑represented populations. In other words, the neural networks are only as good as the experiments they’ve seen, and we simply don’t have enough high‑quality, unbiased measurements yet.</p><p>•	Incomplete biological modeling: Current tools capture parts of the antigen pathway ( binding, sometimes stability, some aspects of T‑cell recognition) but not the full story from antigen processing to presentation to effective tumor killing. A peptide can look perfect on paper and still fail in the body if it isn’t processed correctly, isn’t stably presented, or encounters an immunosuppressive tumor microenvironment.</p><p>This is because biology is a continuous chain of events, but our technology treats it as a series of disconnected snapshots</p><p>•	Manufacturing, cost, and time: Even with perfect predictions, each vaccine is bespoke: it requires tumor sequencing, bioinformatics, GMP manufacturing, and release testing for a single patient. This makes the process slow (often weeks to months) and expensive, limiting access to major centers and clinical trials rather than community hospitals.</p><p>•	Evolving clinical evidence: Regulators and guidelines need large, randomized trials showing clear survival benefits across multiple cancer types. Early results are promising, but the field is still building the robust, long‑term evidence base needed to make personalized neoantigen vaccines standard of care.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*sIcPc3YQeIMkyFQsJEk7bw@2x.jpeg" /></figure><p>So the clinical reality is nuanced: AI‑driven neoantigen discovery is already delivering real, measurable benefits in cutting‑edge trials, but it is not yet a accessible solution for every cancer patient.</p><p>The next leap from promising precision to widely available cure will depend less on inventing entirely new models and more on generating larger, more diverse, and better annotated experimental datasets, integrating the full antigen pathway into AI models, and solving the practical challenges of speed, cost, and scale.</p><p>In that sense, tools like DeepNeo represent a powerful foundation, but the field is still building the data and infrastructure needed for scientists to rely on them with the same confidence they now place in established therapies.</p><h3>7. Future Horizons: What AI Will Unlock Next in Immunotherapy</h3><p>As computing power scales and specialized infrastructure expands, the next decade of AI-driven oncology will unlock entirely new capabilities that sound like science fiction:</p><ul><li>Generative AI for Custom Vaccine Design: Future large language and generative models will do more than just rank existing peptides they will design completely de novo synthetic mRNA sequences and optimize codon usage to maximize protein translation efficiency and stability inside the human body.</li><li>Real-Time Dynamic Triage: As tumors mutate and develop resistance mid-treatment, future AI models will process continuous single cell biopsy streams to dynamically redesign or update vaccine targets on the fly, blocking drug resistance pathways before they can take hold.</li><li>In Vivo Cellular Engineering &amp; Robotics: Combining AI orchestration platforms with automated robotic wet labs and in vivo delivery systems will shorten the “vein-to-vaccine” manufacturing timeline from weeks to days, drastically reducing production costs and making personalized immunotherapy globally accessible.</li><li>Predictive Toxicity Management: Advanced predictive configurations will map complex combination therapies – simulating how checkpoint inhibitors, vaccines, and cellular therapies interact across unique patient genetics to completely anticipate and prevent immune-related adverse side effects before treatment even begins.</li></ul><p>Ultimately, artificial intelligence is turning cancer immunotherapy from a generalized guessing game into a predictable, highly orchestrated digital science, opening the door to a future where personalized cancer vaccines are fast, routine, and curative.</p><h3>Sources:</h3><ul><li><a href="https://github.com/sysbioinfo/DeepNeo-MHC?utm_source=perplexity">GitHub - sysbioinfo/DeepNeo-MHC: peptide-MHC binding prediction tool</a></li><li><a href="https://link.springer.com/article/10.1186/s12929-026-01286-3">Client Challenge</a></li><li><a href="https://academic.oup.com/nar/article/51/W1/W134/7127215">DeepNeo: a webserver for predicting immunogenic neoantigens</a></li><li><a href="https://academic.oup.com/bib/article/26/6/bbaf582/8313934">A machine learning-based method to optimize the immunogenicity of human leukocyte antigen class I-restricted neoantigens</a></li><li><a href="https://doi.org/10.1186/s13073-019-0666-2">Client Challenge</a></li><li><a href="https://doi.org/10.1038/nrg.2016.49">Client Challenge</a></li><li><a href="https://doi.org/10.1186/gm524">Client Challenge</a></li><li><a href="https://doi.org/10.1038/nri3084">Client Challenge</a></li><li><a href="https://doi.org/10.1007/s00251-010-0441-4">Client Challenge</a></li><li><a href="https://doi.org/10.1093/nar/gkaa379">NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data</a></li><li><a href="https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.836821/full">Frontiers | Cancer Neoantigens: Challenges and Future Directions for Prediction, Prioritization, and Validation</a></li><li><a href="https://academic.oup.com/nar/article/51/W1/W134/7127215">DeepNeo: a webserver for predicting immunogenic neoantigens</a></li><li><a href="https://doi.org/10.1093/nar/gkad275">DeepNeo: a webserver for predicting immunogenic neoantigens</a></li><li><a href="https://doi.org/10.1038/s41467-020-14562-z">Client Challenge</a></li><li><a href="https://doi.org/10.1038/s41588-022-01273-y">Client Challenge</a></li><li><a href="https://pubmed.ncbi.nlm.nih.gov/36624345/">pubmed.ncbi.nlm.nih.gov</a></li><li><a href="https://doi.org/10.1093/nar/gky1006">The Immune Epitope Database (IEDB): 2018 update</a></li><li><a href="https://deepneo.co.kr/">DeepNeo</a></li></ul><p>https://github.com/jracle85/DeepNeo</p><p><a href="https://pubmed.ncbi.nlm.nih.gov/37070174/">pubmed.ncbi.nlm.nih.gov</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34a77587023c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>