Core Insights - Google DeepMind's new AI model, AlphaGenome, expands the predictive capabilities of AI into the complex realm of the human genome, marking a significant advancement in genomic research [1] Group 1: AlphaGenome's Capabilities - AlphaGenome can simultaneously predict 11 different gene regulatory processes, accurately capturing complex interactions within genes [3][9] - The model analyzes intricate gene splicing mechanisms, identifying how a single gene can produce multiple proteins and when errors in this process lead to diseases [4] - It has successfully reconstructed pathogenic mutations related to leukemia, predicting changes in regions up to 8000 base pairs away from the gene [6][19] Group 2: Performance Metrics - AlphaGenome has achieved state-of-the-art (SOTA) performance in various tests, surpassing existing models in the field of genomic prediction [8][12] - In 24 evaluations related to genomic trajectory prediction, it secured 22 SOTA results, demonstrating its precision in capturing the effects of small genetic variations [12] - The model's predictions have been validated through rigorous benchmark tests, showcasing its ability to outperform competitors like Borzoi and Enformer in multiple rounds [12] Group 3: Technological Framework - AlphaGenome employs a hybrid architecture combining CNN and Transformer technologies, allowing it to extract local DNA sequence features while capturing long-range dependencies [23][30] - The model's input window has been expanded to 1 million base pairs, enabling comprehensive coverage of interactions between remote enhancers and promoters [28] - A two-phase training strategy was implemented, including pre-training with strict cross-validation and a distillation strategy to enhance generalization and inference efficiency [30] Group 4: Applications and Implications - AlphaGenome's ability to predict molecular phenotypes from DNA sequences enhances the understanding of non-coding regions, addressing challenges in genome-wide association studies (GWAS) [17] - The model has successfully identified regulatory directions for 49% of GWAS-related variants, significantly exceeding traditional methods [17] - Its findings provide actionable insights into the biological functions of non-coding region variations, potentially leading to breakthroughs in disease understanding and treatment [23]

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 谷歌Alpha家族，再登Nature封面！ 这次推出的全新成员 AlphaGenome ，将AI的预测疆域拓展到了最为宏大且神秘的 人类基因组图谱 。 AlphaGenome能够 同时对11种不同的基因调控过程进行综合预测 ，准确捕捉基因深处的复杂互动。 它能 深入分析复杂的基因剪接机制 ，识别细胞如何从单个基因生成多种蛋白质，以及这一过程何时会出错导致疾病。 例如，AlphaGenome对白血病相关基因TAL1的致病突变进行了精准还原，准确预测出8000个碱基之遥的区域发生的突变引起病变。 这能让人类更进一步了解免疫细胞失控增殖引发癌症的深层成因，同时也证明了该模型不仅能处理已知数据，更能对从未见过的DNA片段及 其未知突变做出准确预测。 综合成绩方面，其预测性能在各项测试中均持平或超越现有程序，成为当前基因组预测领域的SOTA模型。 目前，Google DeepMind已面向非商业研究开放AlphaGenome API。 同时预测11种基因调控过程 以及在直接关联疾病研究的变异效应预测任务中， 它与Borzoi、Enformer等现有模型进行了26轮 ...