English
Computer science and biology meet in three distinct ways: computational models help explain biological processes, algorithms help analyze biological observations, and biological mechanisms suggest methods or physical substrates for computing. Keeping these roles separate prevents a useful analogy from becoming an unsupported claim about how life works.
A sequence can be represented as a string, a reaction network as a graph, and a population as a distribution. These representations make particular questions tractable. Each leaves out features of the biological system, so its usefulness depends on the question and the evidence.
Background: Algorithms (§2.6), probability (§2.5), information theory (§3.8), machine learning (§5.2–§5.3), and scientific computing (§6.5). Basic molecular biology is needed to interpret the data and conclusions.
The Relationship: Life, Computation, and Their Mutual Illumination
Life as Information Processing
DNA stores sequence information using four standard bases. This discrete alphabet is useful for explaining copying, mutation, and sequence comparison. The molecules and cellular processes are also chemical: concentrations, reaction rates, spatial organization, and random molecular events matter. Discrete sequence information and continuous physical behavior coexist.
Translation makes the idea of a code concrete. Under the standard genetic code, the RNA codons AUG GCU UAA, read in that frame during translation, specify methionine, alanine, and then a stop signal. Multiple codons can specify the same amino acid. The reading frame and cellular machinery are necessary parts of the explanation; a bare sequence does not interpret itself. The example illustrates codon meanings, not all the requirements for initiating translation in a cell.
The familiar DNA-to-RNA-to-protein diagram describes an important route of sequence information. It is not a claim that every transfer must follow only those arrows: reverse transcription transfers sequence information from RNA to DNA. In Crick’s formulation, the central dogma concerns the restriction on transferring residue-by-residue sequence information out of protein into nucleic acid or protein. Distinguishing this claim from a simplified diagram prevents a common misunderstanding of what reverse transcription changed.
Gene regulation also matters. Cells with essentially the same genome can produce different sets of molecules and develop different functions. A genome sequence is therefore not a complete description of a cell’s current state or an organism’s development. Models of regulation need interactions, timing, and environmental conditions as well as sequences.
Information-processing models become explanatory when their variables correspond to measurable biological quantities and their predictions can be tested. A model of feedback in gene regulation, for example, should specify what is sensed, what changes production, and what response follows a perturbation. Calling a cell a computer adds little unless the proposed representation and operations explain something specific.
Computational Biology: Computation in the Service of Life Science
Computational biology starts from measurements, not already-correct strings and labels. Sequencing produces observations with errors and incomplete coverage. Assembly, alignment, statistical estimation, and data management help turn those observations into evidence about biological questions.
Sequence alignment connects directly to dynamic programming. With unit costs for insertion, deletion, and substitution, ACGT and AGT have edit distance one: delete C. No zero-edit transformation works because the strings differ. Longer sequences require systematic reuse of solutions for shorter prefixes. Real biological alignment often uses substitution scores and gap penalties chosen to reflect biological assumptions; the numerically best alignment depends on that model.
Similarity alone does not establish identical function. A sequence match can support a hypothesis about shared ancestry or function, but interpretation also depends on the extent of the match, statistical significance, biological context, and other evidence. Likewise, an association between a gene and an observed trait is not automatically a causal explanation.
AlphaFold2’s structure-prediction results illustrate how learned models can contribute to science. The system achieved high accuracy on many targets in the 2020 CASP14 assessment, with its method published in 2021. A predicted structure can guide further investigation, but it does not itself show the full folding trajectory, every functional state, or the effect of every mutation. Read local confidence and the relevant modeling assumptions before treating a region as reliable.
Evaluation must reflect the intended generalization. If closely related proteins appear in both training and test sets, apparent performance may overstate the ability to handle distant families. If the intended use concerns a new species or experimental setting, an appropriate split may be more informative than a random split. Experimental follow-up can test whether a useful prediction also supports the proposed biological mechanism.
Biologically Inspired Computation: Life’s Ideas in Computing
Biological inspiration can supply a search strategy without reproducing biology. An evolutionary algorithm maintains candidate solutions, evaluates them with a fitness function, selects candidates, and generates variations through operators such as mutation and recombination. The encoding and operators determine which changes are easy to explore; the fitness function determines what is rewarded.
For example, a candidate schedule can encode assignments of jobs to time slots. Its fitness may penalize conflicts and completion time. Mutation changes an assignment; recombination combines parts of two schedules. These operations may also create invalid schedules, so constraints need explicit handling. The biological terminology does not guarantee a good solution or an advantage over a simpler optimization method. Compare results and evaluation costs against suitable baselines.
Artificial neural networks similarly borrow selected ideas while using abstractions and training procedures that need not match brains (§9.3). In both cases, useful performance and biological fidelity are separate claims. A simplified mechanism can be a good engineering tool without being a faithful biological explanation.
DNA computing is a different connection: molecules serve as the computing substrate. Adleman’s 1994 experiment used DNA to solve a small Hamiltonian-path instance. Molecular parallelism does not make combinatorial growth disappear; representing candidates, preparing material, and reading results all consume physical resources. A demonstration on a small instance does not provide an efficient solution for arbitrary instances of an NP-complete problem.
Synthetic biology can also use ideas such as modularity and feedback to design biological behavior. The comparison with programming is useful for organizing a design, but biological components interact with their host and environment. A component’s behavior must be characterized in context; it cannot simply be assumed to have the isolation or reproducibility of an ideal software function.
What This Perspective Changes
When reading a biological computing claim, first identify its role: a model of life, a tool for analyzing observations, an algorithm inspired by life, or computation implemented with biological material. Then ask what observation or comparison would support the specific claim.
For practice, implement a small sequence-alignment example and vary the scoring assumptions. Explain which differences are consequences of the algorithm and which come from the biological model. This makes domain knowledge part of interpreting the result rather than an optional addition after the computation.
Resources
Foundational and Bridging Texts
Schrödinger’s What Is Life? provides historical context, not a substitute for modern molecular biology. Alberts and colleagues supply that biological background. Compeau and Pevzner and the Rosalind exercises connect sequence problems to algorithms; Durbin and colleagues develop probabilistic sequence models. Crick’s 1970 central-dogma paper clarifies the original information-transfer claim.
| Resource | Role | Type |
|---|---|---|
| Schrödinger, What Is Life? (1944) | Historical questions about physical explanations of life | Depth |
| Compeau & Pevzner, Bioinformatics Algorithms; Rosalind exercises | Sequence problems and algorithmic exercises | Practice |
| Durbin et al., Biological Sequence Analysis | Probabilistic models for biological sequences | Depth |
| Alberts et al., Molecular Biology of the Cell | Cell biology needed to interpret computational results | Reference |
Computational Biology, AI, and Biologically Inspired Methods
Read the AlphaFold2 paper alongside EMBL-EBI’s explanation of its strengths and limits. Inspect a prediction and its confidence before drawing a biological conclusion. Mitchell introduces evolutionary search; Adleman’s DNA-computing experiment illustrates a molecular implementation. Dasgupta’s The Second Age of Computer Science places genetic algorithms in their historical search-and-adaptation context. For each method, compare the biological idea with the actual computational mechanism.
| Resource | Role | Type |
|---|---|---|
| Jumper et al., AlphaFold2 paper (2021) | AlphaFold2 methods and structure-prediction evaluation | Depth |
| Mitchell, An Introduction to Genetic Algorithms | Population-based heuristic search | Entry |
| Adleman, DNA computing paper (1994, free) | A small molecular computation experiment | Depth |
| ML-in-biology literature (current) | Task-specific methods and evidence in biology | Reference |
| AlphaFold Protein Structure Database (free) | Predicted structures with associated confidence information | Practice |
| EMBL-EBI AlphaFold practical guide (free) | Interpreting predictions and their limitations | Practice |
| Rosalind platform (free) | Exercises connecting biological questions and algorithms | Practice |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Equating a genome with a complete program for an organism | Regulation, cellular state, chemistry, and environment affect what happens. | State which biological process the representation explains. |
| Treating a prediction as a complete biological explanation | Similarity or a predicted structure does not establish every function or causal mechanism. | Check assumptions, confidence, generalization, and experimental evidence. |
| Assuming biological inspiration removes computational limits | Heuristic search and molecular parallelism still consume resources. | Compare solution quality, total cost, and scaling under stated conditions. |
中文
计算机科学与生物学有三种不同联系：计算模型帮助解释生物过程，算法帮助分析生物观测，生物机制为计算提供方法或物理载体。区分这些作用，可以避免把有用的类比变成缺少依据的生命解释。
序列可以表示为字符串，反应网络可以表示为图，种群可以表示为分布。这些表示让某些问题更容易处理，但也各自省略了生物系统的部分特征。是否有用，取决于所问的问题与证据。
背景知识：算法（§2.6）、概率（§2.5）、信息论（§3.8）、机器学习（§5.2–§5.3）与科学计算（§6.5）。解释数据与结论还需要分子生物学基础。
这种关系：生命、计算及其相互照亮
生命作为信息处理
DNA 用四种标准碱基存储序列信息。这个离散字母表有助于解释复制、突变与序列比较。分子和细胞过程同时具有化学性质：浓度、反应速率、空间组织与随机分子事件都会产生影响。离散的序列信息与连续的物理行为可以同时存在。
翻译过程让“编码”变得具体。在标准遗传密码下，RNA 密码子 AUG GCU UAA 若在翻译过程中按这个阅读框读取，就依次指定甲硫氨酸、丙氨酸与终止信号。多个密码子可以指定同一种氨基酸。阅读框与细胞中的分子机制，都是解释的必要部分；序列本身不会解释自己。这个例子说明密码子的含义，并没有列出细胞启动翻译所需的全部条件。
常见的 DNA 到 RNA 再到蛋白质示意图，描述了一条重要的序列信息传递路径，并不意味着所有传递都只能沿这些箭头发生：逆转录可以把序列信息从 RNA 传递到 DNA。在 Crick 的表述中，中心法则限制的是从蛋白质向核酸或蛋白质逐个残基传递序列信息。区分这一主张与简化示意图，可以避免误解逆转录究竟改变了什么。
基因调控同样重要。基因组基本相同的细胞，可以产生不同的分子组合并形成不同功能。因此，基因组序列不是细胞当前状态或个体发育的完整描述。调控模型除了序列，还需要相互作用、时间与环境条件。
当模型变量对应可测量的生物量、预测可以接受检验时，信息处理模型才具有具体解释力。例如，基因调控的反馈模型应说明感知什么、什么因素改变产量，以及受到扰动后会出现什么响应。如果没有明确表示与操作能解释的现象，把细胞称为计算机并没有增加多少信息。
计算生物学：服务生命科学的计算
计算生物学从测量开始，而不是从已经正确的字符串与标签开始。测序产生的观测存在误差，覆盖也不完整。组装、比对、统计估计与数据管理，帮助把这些观测转化为有关生物问题的证据。
序列比对直接联系到动态规划。若插入、删除与替换的成本均为一，ACGT 与 AGT 的编辑距离就是一：删除 C 即可。两者不同，因此零次编辑不可能完成转换。更长的序列需要系统地复用较短前缀的解。实际生物序列比对往往采用体现生物学假设的替换得分与空位罚分；数值上最优的比对，取决于所选模型。
相似性本身不能证明功能相同。序列匹配可以支持共同祖先或功能方面的假设，但解释还取决于匹配范围、统计显著性、生物背景与其他证据。同样，某个基因与观测性状相关，也不自动构成因果解释。
AlphaFold2 的结构预测结果说明了学习模型怎样帮助科学研究。系统在 2020 年 CASP14 评估的许多目标上取得很高的准确度，方法于 2021 年发表。预测结构可以指导后续研究，但不能单独展示完整折叠轨迹、所有功能状态或每个突变的影响。应先查看局部置信度与相关建模假设，再判断某个区域是否可靠。
评估应对应预期的泛化要求。如果训练集与测试集包含亲缘关系很近的蛋白质，表面成绩可能高估处理远缘家族的能力。如果预期用途涉及新物种或新的实验条件，适当的数据划分可能比随机划分更有意义。后续实验可以检验：一个有用的预测，是否也支持提出的生物机制。
受生物启发的计算：生命思想进入计算
生物启发可以提供搜索策略，而不必复制真实生物过程。演化算法维护候选解，用适应度函数进行评价，选择候选解，再通过变异、重组等操作产生变化。编码与操作决定哪些变化容易探索，适应度函数决定奖励什么。
例如，一个候选调度方案可以编码任务与时间段之间的分配。适应度可以惩罚冲突与较长的完成时间。变异改变一次分配，重组组合两个方案的部分内容。这些操作也可能产生无效方案，因此需要明确处理约束。采用生物学术语，不能保证得到好解，也不能保证优于更简单的优化方法。应与合适的基线比较结果及评价成本。
人工神经网络同样只借用部分思想，其抽象与训练过程未必对应大脑（§9.3）。两种情况下，性能有用与生物机制逼真，都是不同主张。简化机制可以成为好的工程工具，而不是忠实的生物解释。
DNA 计算属于另一种联系：分子充当计算的物理载体。Adleman 在 1994 年的实验中，用 DNA 解决了一个小型哈密顿路径实例。分子并行不会消除组合数量的增长；表示候选解、制备材料与读取结果，都消耗物理资源。小实例上的演示，不能提供解决任意 NP 完全问题实例的高效方法。
合成生物学还可以借用模块化、反馈等思想，设计生物行为。与编程的比较有助于组织设计，但生物组件会与宿主和环境相互作用。必须在具体背景下描述组件行为，不能直接假定它具有理想软件函数的隔离性或可重复性。
这种视角会改变什么
阅读生物计算相关结论时，先识别其作用：生命过程模型、观测分析工具、受生物启发的算法，还是以生物材料实现的计算。再判断，什么观测或比较能够支持这个具体主张。
可以实现一个小型序列比对例子，改变评分假设，说明哪些差异来自算法，哪些来自生物模型。这样，领域知识就成为解释结果的一部分，而不是计算结束后可有可无的补充。
资源
基础与桥接文本
Schrödinger 的《What Is Life?》提供历史背景，不能替代现代分子生物学；Alberts 等人的教材补充相应生物学基础。Compeau、Pevzner 的教材与 Rosalind 练习联系序列问题和算法；Durbin 等人发展概率序列模型。Crick 的 1970 年中心法则论文澄清了原来的信息传递主张。
| 资源 | 作用 | 类型 |
|---|---|---|
| Schrödinger，What Is Life?（1944） | 有关生命物理解释的历史问题 | 深入 |
| Compeau & Pevzner，Bioinformatics Algorithms；Rosalind 练习 | 序列问题与算法练习 | 实践 |
| Durbin 等，Biological Sequence Analysis | 生物序列的概率模型 | 深入 |
| Alberts 等，Molecular Biology of the Cell | 解释计算结果所需的细胞生物学 | 参考 |
计算生物学、AI 与受生物启发的方法
阅读 AlphaFold2 论文时，可以结合 EMBL-EBI 对其能力与限制的说明。先检查预测及其置信度，再作出生物学结论。Mitchell 介绍演化搜索；Adleman 的 DNA 计算实验展示分子实现。Dasgupta 的《The Second Age of Computer Science》把遗传算法放在搜索与适应研究的历史背景中。对每种方法，都应比较生物启发与实际计算机制。
| 资源 | 作用 | 类型 |
|---|---|---|
| Jumper 等，AlphaFold2 论文（2021） | AlphaFold2 方法与结构预测评价 | 深入 |
| Mitchell，An Introduction to Genetic Algorithms | 基于种群的启发式搜索 | 入门 |
| Adleman，DNA 计算论文（1994，免费） | 小型分子计算实验 | 深入 |
| 机器学习在生物学中的文献（当前） | 生物学中特定任务的方法与证据 | 参考 |
| AlphaFold Protein Structure Database（免费） | 预测结构及相关置信度信息 | 实践 |
| EMBL-EBI AlphaFold practical guide（免费） | 解释预测及其限制 | 实践 |
| Rosalind 平台（免费） | 联系生物问题与算法的练习 | 实践 |
陷阱
| 误区 | 为什么会误导 | 更好的做法 |
|---|---|---|
| 把基因组等同于个体的完整程序 | 调控、细胞状态、化学过程与环境都会影响结果。 | 说明这个表示解释的是哪种生物过程。 |
| 把预测当作完整的生物解释 | 相似性或预测结构，不能证明所有功能或因果机制。 | 检查假设、置信度、泛化要求与实验证据。 |
| 认为生物启发能够消除计算限制 | 启发式搜索与分子并行仍然消耗资源。 | 在明确条件下比较解的质量、总成本与规模增长。 |