https://rwin2.github.io/erwinpoussi.github.io/tag/computational-biology/Computational BiologyHugo Blox Builder (https://hugoblox.com)en-usSat, 01 Aug 2026 00:00:00 +0000https://rwin2.github.io/erwinpoussi.github.io/media/icon_hu0b7a4cb9992c9ac0e91bd28ffd38dd00_9727_512x512_fill_lanczos_center_3.pnghttps://rwin2.github.io/erwinpoussi.github.io/tag/computational-biology/https://rwin2.github.io/erwinpoussi.github.io/project/virtual-embryo/Sat, 01 Aug 2026 00:00:00 +0000https://rwin2.github.io/erwinpoussi.github.io/project/virtual-embryo/<h2 id="embryo-growth-simulation">Embryo Growth Simulation</h2> <p> <figure > <div class="d-flex justify-content-center"> <div class="w-100" ><img alt="Virtual embryo growth simulation" src="https://rwin2.github.io/erwinpoussi.github.io/erwinpoussi.github.io/project/virtual-embryo/embryo_growth.gif" loading="lazy" data-zoomable /></div> </div></figure> </p> <p>This project trains <strong>foundational 3D convolutional models</strong> to reproduce <strong>embryogenesis in silico</strong>, capturing both <strong>cell-level gene expression prediction</strong> and <strong>3D morphogenetic dynamics</strong>. The goal is to learn a generative model of embryonic development that jointly predicts how cells divide, differentiate, and spatially organize — bridging gene regulatory networks with tissue-scale morphology.</p> <p>The approach leverages volumetric deep learning architectures to model the spatiotemporal evolution of developing embryos, learning from multi-modal developmental biology datasets that combine single-cell transcriptomics with 3D imaging.</p> <p>I am one of the builders of the <strong>Virtual Embryo Challenge</strong> — a community benchmark for computational models of embryonic development.</p> <p><strong><a href="https://virtualembryo.ai/challenge" target="_blank" rel="noopener">Virtual Embryo Challenge</a></strong> — check it out and don&rsquo;t hesitate to participate!</p> <p><strong>Status:</strong> More to come soon!</p>https://rwin2.github.io/erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/Wed, 01 Oct 2025 00:00:00 +0000https://rwin2.github.io/erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/<h2 id="training-overview">Training Overview</h2> <p> <figure > <div class="d-flex justify-content-center"> <div class="w-100" ><img alt="SelectSmart gene panel optimization results" srcset=" /erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/RL_huf8a9f04deb0d9e4c8e9ab11372a3db2c_184981_35dfe2eb56b0435ff55209586fcb08d6.webp 400w, /erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/RL_huf8a9f04deb0d9e4c8e9ab11372a3db2c_184981_e724c9bfdab2c9f642dc825f89fdeb34.webp 760w, /erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/RL_huf8a9f04deb0d9e4c8e9ab11372a3db2c_184981_1200x1200_fit_q100_h2_lanczos.webp 1200w" src="https://rwin2.github.io/erwinpoussi.github.io/erwinpoussi.github.io/project/learning-optimal-gene-panels-with-reinforcement-learning/RL_huf8a9f04deb0d9e4c8e9ab11372a3db2c_184981_35dfe2eb56b0435ff55209586fcb08d6.webp" width="760" height="149" loading="lazy" data-zoomable /></div> </div></figure> </p> <p><strong>SelectSmart</strong> is a reinforcement-learning framework for <strong>gene panel selection</strong> in single-cell transcriptomics.<br> This project was developed as part of coursework with <strong>Prof. Mykel Kochenderfer</strong><br> (<a href="https://mykel.kochenderfer.com/" target="_blank" rel="noopener">AA228 / CS238 – Decision Making under Uncertainty</a>) and extended through research assistantship work with <strong>Prof. Xiaojie Qiu</strong><br> (<a href="https://www.devo-evo.com/people/xiaojie/" target="_blank" rel="noopener">Qiu Lab, Stanford</a>).</p> <p>The method combines <strong>meta-voted candidate genes</strong>, an <strong>actor–critic architecture</strong>, and a reward balancing <strong>clustering fidelity (ARI)</strong> with <strong>panel-size regularization</strong>.</p> <p>Trained on a <strong>30k-cell kidney dataset</strong> and evaluated on an <strong>independent CZ Kidney dataset</strong>, SelectSmart produces a <strong>500-gene panel</strong> that preserves transcriptomic geometry and <strong>outperforms classical gene panel selection methods</strong>.</p> <p><strong>Status:</strong> 🧪 Active development</p> <h3 id="report">Report</h3> <p>📄 <a href="report.pdf">Download the technical report (PDF)</a></p>