Games have played a pivotal role in advancing artificial intelligence, withAI agents using sophisticated techniques to compete. Despite the success ofneural network based game AIs, their performance often requires significantcomputational resources. In this paper, we present Rapfi, an efficient Gomokuagent that outperforms CNN-based agents in limited computation environments.Rapfi leverages a compact neural network with a pattern-based codebookdistilled from CNNs, and an incremental update scheme that minimizescomputation when input changes are minor. This new network uses computationthat is orders of magnitude less to reach a similar accuracy of much largerneural networks such as Resnet. Thanks to our incremental update scheme,depth-first search methods such as the alpha-beta search can be significantlyaccelerated. With a carefully tuned evaluation and search, Rapfi reachedstrength surpassing Katagomo, the strongest open-source Gomoku AI based onAlphaZero's algorithm, under limited computational resources where acceleratorslike GPUs are absent. Rapfi ranked first among 520 Gomoku agents on Botzone andwon the championship in GomoCup 2024. <<<

Sebastian Soyk, Zachary H. Lemmon, Matan Oved, Josef Fisher, Katie L. Liberatore, Soon Ju Park, Anna Goren, Ke Jiang, Alexis Ramos, Esther van der Knaap, Joyce Van Eck, Dani Zamir, Yuval Eshed, Zachary B. Lippman <<<

Ann C Huang, Tsung-Han S Hsieh, Jiang Zhu, Jackson Michuda, Ashton Teng, Soohong Kim, Elizabeth M Rumsey, Sharon K Lam, Ikenna Anigbogu, Philip Wright, Mohamed Ameen, Kwontae You, Christopher J Graves, Hyunsung John Kim, Adam J Litterman, Rene V Sit, Alex Blocker, Ci Chu <<<

AbstractThe rapid expansion of massively parallel sequencing technologies has enabled the development of foundation models to uncover novel biological findings. While these have the potential to significantly accelerate scientific discoveries by creating AI-driven virtual cell models, their progress has been greatly limited by the lack of large-scale high-quality perturbation data, which remains constrained due to scalability bottlenecks and assay variability. Here, we introduce “Fix-Cryopreserve-ScRNAseq” (FiCS) Perturb-seq, an industrialized platform for scalable Perturb-seq data generation. We demonstrate that FiCS Perturb-seq exhibits high sensitivity and low batch effects, effectively capturing perturbation-induced transcriptomic changes and recapitulating known biological pathways and protein complexes. In addition, we release X-Atlas: Orion edition (X-Atlas/Orion), the largest publicly available Perturb-seq atlas. This atlas, generated from two genome-wide FiCS Perturb-seq experiments targeting all human protein-coding genes, comprises eight million cells deeply sequenced to over 16,000 unique molecular identifiers (UMIs) per cell. Furthermore, we show that single guide RNA (sgRNA) abundance can serve as a proxy for gene knockdown (KD) efficacy. Leveraging the deep sequencing and substantial cell numbers per perturbation, we also show that stratification by sgRNA expression can reveal dose-dependent genetic effects. Taken together, we demonstrate that FiCS Perturb-seq is an efficient and scalable platform for high-throughput Perturb-seq screens. Through the release of X-Atlas/Orion, we highlight the potential of FiCS Perturb-seq to address current scalability and variability challenges in data generation, advance foundation model development that incorporates gene-dosage effects, and accelerate biological discoveries. <<<