非影像深度学习与DNA：92篇论文逐图解读
每篇论文有独立深读文章，按原始文献清单编号排列。原图之后解释各子图，再把研究问题、模型设计、训练任务和验证证据串起来。文章底部可切换上一篇和下一篇。
阅读时可以沿着四个问题往下走：作者缺少哪种证据？怎样把问题变成训练任务？各张图排除了什么解释？这一范式可以怎样复用？关于作者构思过程的重建标为推测，个人感悟与论文直接结论分开。
先分清模型在学习什么
这些论文并非都在做疾病分类。DNA模型把一段碱基序列当作输入，学习结合、开放、剪接或表达等分子读出；细胞模型把表达状态当作输入，学习基因之间的上下文；临床模型则从病历、波形等信息预测患者终点。输入对象不同，能够回答的问题也不同。把它们放在同一专题，是为了比较怎样把生物学困难转换为可以训练、可以检验的任务。
| 研究范式 | 为什么这样设计 | 最关键的检验 | 本专题的例子 |
|---|---|---|---|
| 用分子任务搭桥 | 疾病变异标签稀缺，先用大量分子实验学习序列规律，再比较两个等位序列 | 对同一位点的等位差异进行独立验证，而不只评价区域分类 | DeepSEA、SpliceAI、ExPecto |
| 从天然序列学习约束 | 不先给致病标签，利用蛋白或DNA序列中的进化变化学习哪些改动不寻常 | 区分序列似然、功能损伤和疾病风险，并在独立功能数据中检验 | EVE、GPN |
| 预训练后迁移 | 目标组织或疾病样本少，先从大量细胞或序列学习可复用的表示 | 检查供体、位点与任务留出，比较相同数据下的简单基线及未预训练模型 | Geneformer、scGPT、DNABERT |
| 预测扰动响应 | 需要知道细胞受到药物或基因干预之后怎样变化 | 留出真正未见的组合、剂量或细胞背景，以实测响应评价预测 | CPA、GEARS |
| 生成并实验筛选 | 已知“怎样预测”之后，反过来寻找满足目标的序列 | 测试新设计在独立实验中的成功率与失败模式 | 调控序列设计相关论文 |
| 检验工具改变了什么 | 预测分数有用，仍要确认医生或研究者据此采取行动后产生实际收益 | 前瞻性流程评价或随机试验，明确结局与观察单位 | EAGLE |
怎样从论文设计重建一个idea
下面是依据论文任务与实验结构作出的编辑推测，用来学习研究设计。它们不能代替作者对真实构思过程的陈述。
DeepSEA：把缺少的标签换成可以获得的标签。 每个非编码变异都做功能实验很难，但大量参考序列已有染色质实验。于是先学“什么序列对应什么分子状态”，再比较同一位置的两个等位版本。真正巧妙的环节，是随后用独立等位读出检验这个转换是否成立。研究起点是标签与问题之间的桥，而不仅是选择CNN。
Geneformer：把大数据经验转给小数据任务。 目标组织和疾病样本少，公开细胞数据却多。表达排序和掩码任务提供一种学习基因上下文的方式，然后用少量标签检验迁移。计算删除一个基因后表示怎样移动，可以帮助产生干预候选；实测扰动及功能读出才把候选接回生物学。
EAGLE：在模型之外寻找下一道瓶颈。 有了能预测低射血分数的心电图模型，接下来缺少的是它能否改变真实诊疗的证据。将结果是否提供给医疗团队进行随机比较，才能评价信息带来的诊断增量。这里新的研究问题来自模型进入流程之后，而不是继续提高一个回顾性分数。
从模型结果走向疾病结论
可以把证据链写成“序列变化 → 分子读出 → 细胞状态 → 患者表型”。每篇文章通常只验证其中几段。高AUC说明某个任务上的排序能力；解释图提供候选线索；等位实验、干预实验或独立队列才检验后续连接。读图时要找出已经验证的箭头，也要保留尚未验证的空白。
这里的个人感悟着重讨论可借用的研究动作：需要补什么数据、设置什么对照、什么结果会让设想失败。对于AD、转座子或演化方向，方法迁移始终需要目标细胞和独立功能证据；论文没有研究AD时，不把它的结果改写成AD机制。
阅读时常见的几个词
| 词语 | 可以怎样理解 |
|---|---|
| motif／基序 | 序列中反复出现、可能被某种蛋白识别的短模式；发现模式后仍要验证它在哪种环境中起作用 |
| TF／转录因子；CRE／顺式调控元件 | 前者是参与调控转录的蛋白，后者是启动子、增强子等DNA片段；二者分别是识别者和可能被识别的序列对象 |
| one-hot／独热编码 | 用四个位置记录某个碱基是A、C、G还是T，让网络读取明确的碱基身份 |
| embedding／嵌入 | 网络用一组数字表示序列、基因或细胞的上下文；数字之间相近不自动代表生物学机制相同 |
| 自监督与微调 | 前者从数据自身构造训练目标，如补回被遮住的碱基；后者再用特定任务的数据调整模型 |
| AUROC与AUPRC | 都评价排序与阈值表现，AUPRC尤其受阳性比例影响；应结合任务分母和真实使用阈值来读 |
| 消融与独立验证 | 消融问去掉某个组件后增益是否消失；独立验证问模型是否能经受未参与开发的数据或实验检验 |
本专题只讲正文主图。第29篇依据所提供的2025年预印本解读，并同时链接2026年正式发表记录；第91、92篇的预印本版本也在各自文章中注明。
92篇独立深读目录
| 编号 | 论文缩写与独立解读 | 论文题名 |
|---|---|---|
| 1 | DeepSEA | Predicting effects of noncoding variants with deep learning–based sequence model |
| 2 | HumanSplicingCode | The human splicing code reveals new insights into the genetic determinants of disease |
| 3 | ExPecto | Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk |
| 4 | ASDNoncodingBurden | Whole-genome deep-learning analysis identifies contribution of noncoding mutations to autism risk |
| 5 | SpliceAI | Predicting Splicing from Primary Sequence with Deep Learning |
| 6 | Enformer | Effective gene expression prediction from sequence by integrating long-range interactions |
| 7 | EVE | Disease variant prediction with deep generative models of evolutionary data |
| 8 | AlphaMissense | Accurate proteome-wide missense variant effect prediction with AlphaMissense |
| 9 | PrimateAI-3D | The landscape of tolerated genetic variation in humans and primates |
| 10 | Borzoi | Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation |
| 11 | GET | A foundation model of transcription across human cell types |
| 12 | AlphaGenome | Advancing regulatory variant effect prediction with AlphaGenome |
| 13 | NucleotideTransformer | Nucleotide Transformer: building and evaluating robust foundation models for human genomics |
| 14 | Evo2 | Genome modelling and design across all domains of life with Evo 2 |
| 15 | scGen | scGen predicts single-cell perturbation responses |
| 16 | Geneformer | Transfer learning enables predictions in network biology |
| 17 | GEARS | Predicting transcriptional outcomes of novel multigene perturbations with GEARS |
| 18 | scGPT | scGPT: toward building a foundation model for single-cell multi-omics using generative AI |
| 19 | scFoundation | Large-scale foundation model on single-cell transcriptomics |
| 20 | CPA | Predicting cellular responses to complex perturbations in high-throughput screens |
| 21 | biolord | Disentanglement of single-cell data with biolord |
| 22 | State | Predicting cellular responses to perturbation across diverse contexts with State |
| 23 | P-NET | Biologically informed deep neural network for prostate cancer discovery |
| 24 | DrugCell | Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells |
| 25 | DeepDEP | Predicting and characterizing a cancer dependency map of tumors with deep learning |
| 26 | CODE-AE | A context-aware deconfounding autoencoder for robust prediction of personalized clinical drug response from cell-line compound screening |
| 27 | Sturgeon | Ultra-fast deep-learned CNS tumour classification during surgery |
| 28 | crossNN | crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors |
| 29 | MutationProjector | Translating clinical gene sequencing into a foundational representation of tumor subtype |
| 30 | DeepTCR | DeepTCR is a deep learning framework for revealing sequence concepts within T-cell repertoires |
| 31 | DeepTCR-Immunotherapy | Deep learning reveals predictive sequence concepts within immune repertoires to immunotherapy |
| 32 | Halicin | A Deep Learning Approach to Antibiotic Discovery |
| 33 | Abaucin | Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii |
| 34 | ExplainableAntibiotics | Discovery of a structural class of antibiotics with explainable deep learning |
| 35 | GENTRL-DDR1 | Deep learning enables rapid identification of potent DDR1 kinase inhibitors |
| 36 | TNIK | A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models |
| 37 | Rentosertib | A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial |
| 38 | TxGNN | A foundation model for clinician-centered drug repurposing |
| 39 | AntitoxinDesign | De novo designed proteins neutralize lethal snake venom toxins |
| 40 | CARGrammar | Decoding CAR T cell phenotype using combinatorial signaling motif libraries and machine learning |
| 41 | CellTargetingCRE | Machine-guided design of cell-type-targeting cis-regulatory elements |
| 42 | AKIPrediction | A clinically applicable approach to continuous prediction of future acute kidney injury |
| 43 | PancreaticCancerPrediction | A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories |
| 44 | NYUTron | Health system-scale language models are all-purpose prediction engines |
| 45 | Delphi2M | Learning the natural history of human disease with generative transformers |
| 46 | ArrhythmiaDNN | Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network |
| 47 | AI-ECG | Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram |
| 48 | EAGLE | Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial |
| 49 | ParkinsonBreathing | Artificial intelligence-enabled detection and assessment of Parkinson’s disease using nocturnal breathing signals |
| 50 | SleepFM | A multimodal sleep foundation model for disease prediction |
| 51 | RL-DITR | Optimized glycemic control of type 2 diabetes with reinforcement learning: a proof-of-concept trial |
| 52 | SCFM-ZeroShot | Zero-shot evaluation reveals limitations of single-cell foundation models |
| 53 | PerturbationLinearBaselines | Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines |
| 54 | CalibratedPerturbationMetrics | Deep learning perturbation models can outperform baselines on calibrated metrics |
| 55 | DeepBind | Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning |
| 56 | DeepSTARR | DeepSTARR predicts enhancer activity from DNA sequence and enables the de novo design of synthetic enhancers |
| 57 | Sei | A sequence-based global map of regulatory activity for deciphering human genetics |
| 58 | Akita | Predicting 3D genome folding from DNA sequence with Akita |
| 59 | Orca | Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale |
| 60 | scBasset | scBasset: sequence-based modeling of single-cell ATAC-seq using convolutional neural networks |
| 61 | scooby | scooby: modeling multimodal genomic profiles from DNA sequence at single-cell resolution |
| 62 | CREsted | CREsted: modeling genomic and synthetic cell-type-specific enhancers across tissues and species |
| 63 | TaskiranEnhancerDesign | Cell-type-directed design of synthetic enhancers |
| 64 | YeastRegulatoryDNA | The evolution, evolvability and engineering of gene regulatory DNA |
| 65 | PARM | Regulatory grammar in human promoters uncovered by MPRA-based deep learning |
| 66 | SegmentNT | Annotating the genome at single-nucleotide resolution with DNA foundation models |
| 67 | GPN-MSA | A DNA language model based on multispecies alignment predicts the effects of genome-wide variants |
| 68 | gReLU | gReLU: a comprehensive framework for DNA sequence modeling and design |
| 69 | Basset | Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks |
| 70 | Basenji | Sequential regulatory activity prediction across chromosomes with convolutional neural networks |
| 71 | BPNet | Base-resolution models of transcription-factor binding reveal soft motif syntax |
| 72 | DanQ | DanQ: a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences |
| 73 | Puffin | Sequence basis of transcription initiation in the human genome |
| 74 | seq2PRINT | Multiscale footprints reveal the organization of cis-regulatory elements |
| 75 | C-Origami | Cell-type-specific prediction of 3D chromatin organization enables high-throughput in silico genetic screening |
| 76 | DeepMEL | Cross-species analysis of enhancer logic using deep learning |
| 77 | DeepMEL2 | Interpretation of allele-specific chromatin accessibility using cell state-aware deep learning |
| 78 | DeepCpG | DeepCpG: accurate prediction of single-cell DNA methylation states using deep learning |
| 79 | DNABERT | DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome |
| 80 | GPN | DNA language models are powerful predictors of genome-wide variant effects |
| 81 | Evo | Sequence modeling and design from molecular to genome scale with Evo |
| 82 | EvoSemanticDesign | Semantic design of functional de novo genes from a genomic language model |
| 83 | DeepC | DeepC: predicting 3D genome folding using megabase-scale transfer learning |
| 84 | TissueEnhancerDesign | Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo |
| 85 | BE-Hive | Determinants of Base Editing Outcomes from Target Library Analysis and Machine Learning |
| 86 | inDelphi | Predictable and precise template-free CRISPR editing of pathogenic variants |
| 87 | PersonalExpressionBenchmark-Sasse | Benchmarking of deep neural networks for predicting personal gene expression from DNA sequence highlights shortcomings |
| 88 | PersonalTranscriptome-Huang | Personal transcriptome variation is poorly explained by current genomic deep learning models |
| 89 | SegmentNT-ContextBias | Systematic contextual biases in SegmentNT potentially relevant to other nucleotide transformer models |
| 90 | PersonalExpressionFineTuning | Fine-tuning sequence-to-expression models on personal genome and transcriptome data |
| 91 | ChromBPNet | ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants |
| 92 | ProCapNet | Dissecting the cis-regulatory syntax of transcription initiation with deep learning |