Introduction: Natural and Artificial Intelligence Interplay

• Transformative regenerative medicine requires understanding the body’s natural intelligence. This involves an interplay between understanding natural biological systems and developing artificial intelligence.
• Living systems are multiscale, not just structurally, but functionally. Each level (cells, tissues, organs, organisms) solves problems with a form of “collective intelligence.”
• Levin’s lab uses machine learning to understand/control biological endpoints (for medicine) and uses biology to inspire new AI architectures (non-neural).

Key Concepts: Anatomical Homeostasis and Bioelectricity

• Anatomical homeostasis: Maintaining a correct body structure despite perturbations (injury, mutations, etc.). This involves *feedback loops*, not just feed-forward processes.
• Current focus in biology is primarily manipulating genetics, cells and proteins (hardware). There’s much needed on *form*,*function* and controlling *decisions*.
• Current focus of computer science has moved toward manipulating data. The analogy to biomedicine is this *form* can be created with a biological compiler (analogy: electrical compuational systems don’t need to rewire transistors for diff tasks; they can simply program data, we are very behind on this journey.)
• Cells make decisions, and non-neural bioelectricity is a key medium for this computation, a crucial “software layer” between the genome and the body. It’s a kind of “epigenetics.”
• Regenerative medecine’s final form would have to be: sit in front of computer, type in body plan you want, push go.
• How do a collection of cells KNOW to produce an adult organism.
• the planarian: even with hundreds of science/nature papers there still has not ever been an experiment created yet.
• There’s an issue in current biology which is its very focused on the lowest level building blocks but this isn’t what we need, we need whole level organization.
• The bioelectric code: Decoding this will lead to “electroceuticals” for regenerative medicine, cancer, and synthetic bioengineering.
• Body tissues, like the brain, form electrical networks that make decisions about dynamic anatomy. AI/machine learning tools help target this system.

Examples and Model Systems

• Morphogensis is *flexible*.
• Axolotls: Regenerate limbs, eyes, jaws, spinal cords, etc. The regeneration is *context-sensitive* and *goal-directed*.
• Planaria (flatworms): Extreme regeneration (any body part), “immortal” (no aging). Demonstrate flexible regeneration and the role of bioelectric “set points.”
• Experiments moving frog tadpole facial features (Picasso tadpoles): Show that anatomical structures are *not hardwired*, but achieved through error minimization.
• Frog tadpoles can compensate for a *variety of things*, where all tissues migrate to their correct spot even after it is not on where it supposed to be.
• Frog Eye: frog is not ‘wired’, and a specific bio-electrical patter says ‘make an eye’, *anywhere*.
• Planaria Heads: You can set how many heads it should create.

Bioelectric Manipulation and Control

• Analogy to Thermostat: Don’t need to rewire things! Instead *manipulating electrical activity by* of rewriting “set point” information.
• During evolution *size* of the thing organisms operate is flexible and scales up/down with goals. Cancer being an example where single-cellular organisms revert to small, simpler goal.
• Methods: Voltage-sensitive fluorescent dyes to *visualize* bioelectric activity. Computational modeling to simulate electrical networks.
• Bioloectricals do not *need* to equal *now*: It does not need to equal the *present*, it’s often stored *before* that.
• Can re-create head patterns.
• Manipulations: Controlling ion channels and gap junctions (like in neuroscience) using drugs, mutations, optogenetics (light). *No external electric fields*.
• Single-cell level: Preventing tumor formation by restoring electrical connection to neighboring cells (overriding oncogene effects).
• Organ level: Inducing ectopic (out-of-place) organs (eyes, hearts, limbs) by imposing specific bioelectrical states. Like a “subroutine call.”
• Whole-body level: Controlling planarian head number (one vs. two) by altering bioelectric patterns. Can even create head shapes of *different species* without genetic changes.
• It can rewrite ‘set point’ to the anatomy! which changes the *form* of an organism: The way you rewrite is a biological intervention, can use electrical-based drugs. This will provide for ‘ionoceuticals’.
• Limb regeneration in frogs (non-regenerating species): A 24-hour bioelectric treatment triggers long-term leg regrowth, without further intervention.

AI’s Role in Understanding and Intervention

• “Full stack” approach: Modeling transcriptional circuits, bioelectric dynamics, and large-scale patterning to derive *algorithmic* descriptions for intervention.
• Betsy, is a software designed to do *circuit models* using individual cells on the tissue/anatomy to ‘simulate’ it.
• AI’s role is two-fold: 1) *Inferring models* from experimental data. 2) *Inferring interventions* (which channels to target, how) based on a desired outcome.
• Example: Evolutionary computation used to infer a model of planarian regeneration. The AI “guessed” a human-understandable model.
• There is no model that we know so far that gives a prediction on shapes.
• Problem is current model for regnenerative biomedicine would requrie that if we wanted a specific part of anatomy is make many many many mutations but a intractable reverse problem we simply don’t have solutions for.
• Using *evolution* in AI to design biology *models*. *software*, it discovers model based on human understanding that could only before, only be created from a very good human mind, yet even so no models that can give prediction (e.g. with Planaria experiments on changing bio-electric field shapes, for example) have ever yet to have existed.
• Software ‘Elektra’ has ability to: take database of how things *should* function, how *does* function, with all various data, use an evolutionary computation system. (in Plenaria case it got 800 inputs, where most don’t work at all, so had to infer using functional info)
• The inference system gave useful information with model *without* large amounts of input data.
• AI to *model*, but ALSO AI to design new interventions (how you create a new medicine).
• The “code” metaphor: Genome defines the *hardware* (ion channels), but the resulting electrical network (excitable medium) has emergent properties (software), storing *patterns* and *memories*. Like a flip-flop.
• Example: Editing the planarian “memory” of head number (software) *without changing the genome* (hardware). The new pattern is stable and *heritable* (through cutting).
• Editing bio-patterns allows editing for: the *shape*, it also controls *growth* in adult organism too! e.g. with *Leg* (frog). This is a big example with the two heads on an Plenaria organism. It even *keeps going*: you cut up head into many *and* can use a different bio-signal *to rewrite* it, again!
• Machine Learning Connections: Connecting bioelectric circuits to concepts from connectionist machine learning (pattern completion, energy minimization).

Future Vision and Conclusions

• Bioelectricity: Key role of biology and it’s relation to software.
• Key goal: use computer simulation, not to replace experiment *but* it tells us *which* intervention will get us the result we want.
• Example: Rational design of an “electroceutical” to rescue brain defects. A model predicted a specific ion channel (hcn2) to correct the bioelectric pattern, *not* a trial-and-error screen.
• “Electroceutical Design Environment”: A future system where you specify cells, tissues, and a desired pattern, and it tells you which ion channel drugs to use.
• Rational design of drugs based on pattern completion *with no human trial and error* in frogs! e.g. drug hcn2 (discovered from *electrical modelling*)
• The ultimate Vision of using models for biology/biomed: is that this AI system would output specific and useful outputs from AI to help guide with which *bioelectrical* *and biochemical* changes would have to take place based on all known scientific inputs (from a database).
• Conclusions: Bioelectricity is a tractable “software layer” for regenerative medicine. Evolution uses electrical signaling for large-scale coordination. We can read/write pattern memories to reprogram shape. Machine learning helps infer models *and* interventions. This could revolutionize medicine and inform new AI architectures.
• Two Big Outcomes: fantastic regeneration medecine AND give inspiration to design new kinds of AI that uses different principles of cognition.

导言：自然智能与人工智能的相互作用

• 变革性的再生医学需要理解身体的自然智能。这涉及理解自然生物系统和开发人工智能之间的相互作用。
• 生命系统是多尺度的，不仅在结构上，而且在功能上。每个层次（细胞、组织、器官、生物体）都以一种“集体智慧”的形式解决问题。
• 莱文的实验室利用机器学习来理解/控制生物终点（用于医学），并利用生物学来激发新的AI架构（非神经）。

关键概念：解剖稳态和生物电

• 解剖稳态：尽管存在扰动（损伤、突变等），仍保持正确的身体结构。这涉及反馈回路，而不仅仅是前馈过程。
• 目前生物学的重点主要是操纵遗传、细胞和蛋白质（硬件）。 在形式、功能和控制决策方面，还有很多需要做的工作。
• 当前计算机科学的重点已转向操纵数据。与生物医学的类比是，这种形式可以用生物编译器创建（类比：电子计算系统不需要为不同的任务重新连接晶体管；它们可以简单地对数据进行编程，我们在这方面还很落后。）
• 细胞做出决策，而非神经生物电是这种计算的关键媒介，是基因组和身体之间至关重要的“软件层”。这是一种“表观遗传学”。
• 再生医学的最终形式必须是：坐在电脑前，输入你想要的身体计划，然后按下开始键。
• 一群细胞如何知道要产生一个成年生物体？
• 涡虫：即使有数百篇科学/自然论文，仍然没有进行过任何实验。
• 当前生物学存在一个问题，它非常关注最低层次的构建块，但这不是我们需要的，我们需要整体层面的组织。
• 生物电代码：解码这将导致用于再生医学、癌症和合成生物工程的“电疗药物”。
• 身体组织，像大脑一样，形成电网络，对动态解剖结构做出决策。人工智能/机器学习工具有助于针对该系统。

实例和模型系统

• 形态发生是灵活的。
• 蝾螈：再生四肢、眼睛、下巴、脊髓等。再生是上下文敏感的和目标导向的。
• 涡虫（扁虫）：极度再生（任何身体部位），“永生”（不衰老）。展示了灵活的再生和生物电“设定点”的作用。
• 移动青蛙蝌蚪面部特征的实验（毕加索蝌蚪）：表明解剖结构不是硬连接的，而是通过错误最小化来实现的。
• 青蛙蝌蚪可以补偿各种情况，即使组织不在它应该在的位置，所有组织都会迁移到它们正确的位置。
• 青蛙眼睛：青蛙不是“接线”的，特定的生物电模式表示“制造一个眼睛”，可以在任何地方。
• 涡虫头部：您可以设置它应该创建多少个头部。

生物电操纵与控制

• 与恒温器的类比：不需要重新接线！而是通过重写“设定点”信息来操纵电活动。
• 在进化过程中，生物体操作的物体的大小是灵活的，并且可以根据目标进行放大/缩小。癌症就是一个例子，其中单细胞生物恢复到小的、简单的目标。
• 方法：电压敏感荧光染料以可视化生物电活动。计算模型来模拟电网络。
• 生物电的不一定需要现在。它并不需要与现在相等，它通常是在那之前存储的。
• 可以重新创建头部图案。
• 操纵：使用药物、突变、光遗传学（光）控制离子通道和间隙连接（就像在神经科学中一样）。没有外部电场。
• 单细胞水平：通过恢复与邻近细胞的电连接来防止肿瘤形成（覆盖癌基因效应）。
• 器官水平：通过施加特定的生物电状态诱导异位（异位）器官（眼睛、心脏、四肢）。就像一个“子程序调用”。
• 全身水平：通过改变生物电模式控制涡虫头部数量（一个与两个）。甚至可以创建不同物种的头部形状，而无需基因改变。
• 它可以重写解剖结构的“设定点”！ 这会改变生物体的形式：您重写的方式是一种生物干预，可以使用基于电的药物。 这将提供“离子药物”。
• 青蛙（非再生物种）的肢体再生：24小时的生物电治疗触发了长期的腿部再生，无需进一步干预。

人工智能在理解和干预中的作用

• “全栈”方法：对转录回路、生物电动力学和大规模模式进行建模，以导出用于干预的算法描述。
• Betsy，是一种软件，旨在使用组织/解剖结构上的单个细胞进行电路模型“模拟”。
• 人工智能的作用是双重的：1）从实验数据中推断模型。2）根据期望的结果推断干预措施（要针对哪些通道，如何）。
• 示例：进化计算用于推断涡虫再生的模型。 人工智能“猜到”了一个人类可以理解的模型。
• 到目前为止，我们还没有任何模型可以预测形状。
• 问题是，当前再生生物医学的模型要求，如果我们想要解剖结构的特定部分，就必须进行许多突变，但这我们根本没有解决办法。
• 利用人工智能中的进化来设计生物学模型。 软件，它发现了基于人类理解的模型，而以前，这些模型只能由非常好的人类思维创造出来，即便如此，还没有任何模型可以给出预测（例如，关于改变生物电场形状的涡虫实验）。
• 软件“Elektra”具有以下能力：获取事物应该如何运作、如何运作的数据库，以及所有各种数据，使用进化计算系统。（在涡虫案例中，它获得了800个输入，其中大多数根本不起作用，因此必须使用功能信息进行推断）
• 推理系统提供了有用的信息，即使没有大量的输入数据,依旧提供有效模型信息。
• 人工智能用于建模，也用于设计新的干预措施（如何制造新药）。
• “代码”隐喻：基因组定义了硬件（离子通道），但由此产生的电网络（可激发介质）具有涌现特性（软件），存储模式和记忆。就像一个触发器。
• 示例：编辑涡虫头部数量的“记忆”（软件）而不更改基因组（硬件）。新模式是稳定的并且可遗传（通过切割）。
• 编辑生物模式允许编辑：形状，它也控制成年生物体的生长！例如，腿（青蛙）。这是涡虫生物体上两个头的一个大例子。它甚至可以持续下去：你把头切成很多，并且可以使用不同的生物信号再次重写它！
• 机器学习连接：将生物电电路与连接主义机器学习的概念（模式完成、能量最小化）联系起来。

未来愿景和结论

• 生物电：生物学的关键作用及其与软件的关系。
• 关键目标：使用计算机模拟，不是为了取代实验，而是告诉我们哪种干预措施将得到我们想要的结果。
• 示例：合理设计“电疗药物”以挽救脑部缺陷。一个模型预测了一个特定的离子通道 (hcn2) 来纠正生物电模式，而不是试错筛选。
• “电疗药物设计环境”：一个未来的系统，您可以在其中指定细胞、组织和所需的模式，它会告诉您使用哪种离子通道药物。
• 基于模式完成的药物合理设计，无需在青蛙身上进行人工试错！例如药物 hcn2（从电模型中发现）
• 使用生物/生物医学模型的最终愿景是，该人工智能系统将根据所有已知的科学输入（来自数据库）输出具体的、有用的结果，帮助指导需要进行哪些生物电和生化变化。
• 结论：生物电是再生医学的一个易于处理的“软件层”。进化利用电信号进行大规模协调。我们可以读/写模式记忆来重新编程形状。机器学习有助于推断模型和干预措施。这可能会彻底改变医学并为新的人工智能架构提供信息。
• 两大成果：神奇的再生医学，并为设计利用不同认知原则的新型人工智能提供灵感。