juice like

科研民工

[Agingdoc🩺Dr David Barzilai🔔MD PhD MS MBA DipABLM @agingdoc1]

"The reprogramming thesis rests on the idea — articulated in the seminal 2013 paper on the hallmarks of aging — that biological decline is driven less by accumulated damage than by lost epigenetic information that can, in principle, be rewritten. Mouse studies have shown tissue healing, vision restoration, and cognitive improvement. Whether the biology translates to humans is the empirical question. Whether the funding model survives if it doesn’t is the institutional one."

[NonsparseOncologist @5_utr]

🧵
A Nature Medicine paper trained a transformer with ~1M parameters on immunotherapy cohorts as small as n=16, dichotomized every outcome twice, “beat” 22 rival methods with zero rank uncertainty, and skipped covariate adjustment.

Textbook bad biomarker research, gift-wrapped.

[墓碑科技 @mubeitech]

MIT 历史上一门被疯传的经典公开课，居然跟 AI、算法、物理一点关系都没有。

这门课叫《How to Speak》。
主讲人 Patrick Winston 教授，在 MIT 每年一月雷打不动地讲这门课，讲了整整 40 年，直到 2019 年去世。
几代 MIT 的天才学生，在做任何项目、搞任何创业之前，都会被前辈叮嘱：先去把这门课看了。

这位教授一开场，就丢出了一个甚至有点刺耳、极其粗暴的暴论：
“你这辈子的成功，很大程度上取决于这三件事：你的表达能力，你的写作能力，最后，才是你想法的质量。”

在最硬核的科技殿堂 MIT，表达，被排在了想法前面。

这个结论对很多“技术思维”的人来说，简直是当头一棒。
我们总觉得，只要我的技术够硬、想法够天才，别人就应该买单。
但现实极其残酷。
你空有一脑子改变世界的想法，如果讲不明白，在别人眼里，你和没有想法没有任何区别。

到底怎么把想法讲到别人愿意听、听得懂、还忘不掉？
这门课的核心逻辑，其实可以被提炼成几个极度实用的表达规则：

第一，永远不要以笑话开场。
很多人觉得讲 PPT 前先讲个段子很暖场。
教授说：错。
听众刚坐下，脑子还没进入你的频道，这时候讲笑话，他们大概率反应不过来，现场只会剩下尴尬。
开场的黄金时间，应该用来抛出“承诺”，告诉他们听完这 60 分钟到底能带走什么。

第二，控制信息密度。
别把你的 PPT 塞满密密麻麻的字，这等于逼着人家在现场做阅读理解。
人类的大脑，一次只能处理一个核心焦点。
页面上的字越多，听众的注意力离你越远。

第三，在听众脑子里“画画”。
如果你在讲一个复杂的概念，不要直接上公式或黑话。
试着用最具体的、能让人瞬间产生画面感的类比，把听众带入情境。

这位教授用 40 年总结出的这套沟通机制，直到今天依然是很多精英的秘密武器。
甚至现在，有人直接把这套规则调教成 AI 提示词，在每一次面试、路演、汇报前，让 AI 用 Winston 教授的严苛标准来当自己的教练。

在这个信息垃圾成灾的时代，最稀缺的早就不是“想法”，而是“注意力”。
学会表达，不用去迎合谁，它能给你脑子里那颗珍贵的种子，穿上一件在现实世界里落地的防弹衣。

[Ran Blekhman @blekhman]

Claude Science is incredible. I gave it some sequencing data, and in 8 hours it did a full analysis, generated figures, wrote a paper, submitted it for publication, got rejected, revised and resubmitted, got rejected again, it is now applying for positions in industry

[Aging Science News @AgingBiology]

Why billionaires could reverse ageing for everyone – but keep refusing to
https://t.co/Q5iZ8QMDzW

[0xCrypto Tied @CryptoTied]

🚨卧槽！iOS 网络定位也能一键切位置了

wloc 能让你一键把 iPhone 的网络定位改到全世界任意地方（香港、东京、纽约、巴黎等）。

操作极其简单：在地图 App 长按选点 → 共享 → 一键切换位置。支持 Surge、Quantumult X、Loon、Stash、小火箭等主流工具。

核心亮点：

→地图上直接点选，无需手动输入经纬度
→完美支持苹果地图 + 高德短链，自动处理坐标转换
→快捷指令集成度极高，操作流程非常顺畅
→无需越狱，纯 MITM 脚本，相对安全
→支持超长时长稳定运行

适合 App 测试、切换城市模式、出差打卡、室内定位测试等场景。

[鸟叔本鸟 @niaoshu]

中国内地用户最容易获得的正规海外银行卡，

就是：瑞士杜高斯贝银行的，这张万事达卡了，

还用担心怎么支付 Claude，chatgpt吗？NO！

它是正儿八经的海外实体银行的实体卡，

我的已经UPS寄出了，说 7 月 3 号就能拿到。

[Ming "Tommy" Tang @tangming2005]

1/ How to quantify a transgene (e.g., TdTomato, GFP) in scRNAseq data? You need to add the transgene to the reference. See my post https://t.co/7lnIW6kYMF However, it is not always working 👇

[黄赟 @huangyun_122]

品行好的老实人，在职场怕是最容易被拿捏的。老板绝对不会因你老实给你升职加薪，相反你最容易被拿捏，那就晾着你，等你开口

这两天看《四渡》，我突然明白了“运动”这个策略的重要性。换到职场来，老实人最应该做的事：不断跳槽

每次带着新理念回到老东家，反而最好涨薪

[Avi Roy @agingroy]

The clearest explainer yet of why the richest people alive are pouring billions into reversing aging. The bet is shakier than it looks.

Their wager is that your old cells aren’t broken, only badly configured, so you reload the settings and they run young again. That’s the whole thesis behind @altos_labs and its $3 billion.

The proof that lit the fuse came in 2020. @davidasinclair’s lab gave blind mice their sight back, with nothing but a short pulse of the Yamanaka factors, the proteins that rewind a cell toward its newborn state. The cells never changed jobs. They just read their DNA like young ones again.

But reloading only works where aging is a settings problem, and a lot of it isn’t. Your DNA picks up mutations, the letters themselves rewritten, and you can’t reload your way out of a typo. Other cells are already dead, so there is nothing to reload.

So here’s what the article leaves out. How much of aging is bad settings you can reset, and how much is damage you can’t undo? Two labs are fighting that out in @CellCellPress right now.

The whole bet rides on who wins.

[Science News @SciencNews]

How to Write Clearly

[Science News @SciencNews]

The importance of stupidity in scientific research

[Aging Science News @AgingBiology]

Methionine-supplemented longevity diet increases growth hormone, GLP-1, and FGF21; reduces frailty; and promotes healthspan
https://t.co/HGbQHe4thW

[Aging Science News @AgingBiology]

Unapproved gene therapy for boosting longevity is set to go on sale
https://t.co/4gfK8Vnltp

[Elliot Hershberg @ElliotHershberg]

CSHL: AI in Biology

https://t.co/MpE3HpDSV9

Since 1933, @CSHL has hosted an annual Symposium on Quantitative Biology.

At first, "quantitative biology" meant the use of chemical, physical, and mathematical techniques. This conference became the Schelling point for the pioneers of molecular biology. Watson first presented the structure of DNA at the 1953 Symposium.

This year, the topic for this famous conference was AI in Biology.

For five days, the top researchers in this field gathered from around the world to present their latest work. I went, and have done my best to summarize some of the major themes and results from the Symposium.

It was a lot of fun attempting to synthesize ideas from superstars including @pushmeet, @Avsecz, @zhou_jian, @Micro_Yunha, @pkoo562 @anshulkundaje, @Prof_Lundberg, @recursus, @ZhongingAlong, @marinkazitnik, @lecong, and more!

I hope you enjoy reading—it was one of my favorites conferences I've ever been to. Some truly beautiful research on display.

[Changyang Linghu @ChangyangLinghu]

What if you could record cytosolic activities or archive proteins on an intracellular tape across time? 📼
Introducing CyTRACE: a genetically encoded, modular framework built on protein tape recorders for scalable post hoc spatiotemporal analysis and molecular interrogation.
🔍 CytoSense records protease & kinase signaling activities
📦 CytoCatch archives cytosolic proteins along CytoTape
Preprint out: https://t.co/JWZt8VdVUt
🧵Thread (1/7)

[Arc Institute @arcinstitute]

Imagine programming protein, DNA, and RNA systems like you would write computer code, or even by natural language prompting of an AI agent. @brianhie and team just made this a reality with Proto: a high-level programming language for generative biology.

[Charles Brenner, PhD @CharlesMBrenner]

i love transcription factors as drugs & targets as much as the next guy but there's no special magic in which old or dysfunctional tissues go back in time in an orderly process on the basis of reaccessing an epigenetic state

investors who were told this were sold hype

[Biology+AI Daily @BiologyAIDaily]

Emyx: Fast and efficient all-atom protein generation

1 Emyx is a 140M-parameter conditional flow matching model for all-atom protein generation aimed at enzyme active-site scaffolding, where the generator must satisfy tight catalytic geometry while still exploring diverse global folds.

2 The paper’s key practical result is AME benchmark performance under a stricter “self-consistency” definition: 13.4% strict sc-RMSD success (39/41 targets solved), outperforming Proteína-Complexa (8.8%, 37/41) and RFdiffusion3 (6.7%, 35/41), despite being the smallest model.

3 The authors argue current all-atom generators are overbuilt by inheriting expensive structure-prediction components (dense pair representations, heavy embedding stacks). For generators conditioned mainly on sparse geometric constraints (motif + ligand), Emyx concentrates capacity in standard transformer blocks and uses sparse connectivity to avoid quadratic overhead.

4 Architecture summary: a two-level token/atom representation (Rep14: 14 atom slots per residue or ligand token), lightweight atom-level encoder blocks, gated cross-attention to pool atoms into token embeddings, a deep token-level transformer trunk with sparse self-attention + learned pair bias on sparse edges, then cross-attention back to atoms for decoding and a velocity head.

5 Sparsity is enforced via a fixed edge budget per node with priority rules (sequence neighbors, chemical bonds, ligand kNN, motif connectivity, and kNN fill). This preserves critical motif/ligand interactions while improving scaling versus dense O(N^2) designs.

6 Training efficiency: Emyx trains on experimental PDB-derived structures (≈629K chains) in 681.6 GPU-hours (3.55 days on 8 H200s), reported as ~4× less compute than RFdiffusion3 (≈2,688 GPU-hours). The model uses a linear flow-matching interpolant for lower-variance gradients and faster convergence.

7 A major methodological contribution is the strict sc-RMSD protocol: it requires (i) global backbone RMSD < 2.0 Å after full-backbone alignment, (ii) motif tip-atom RMSD < 1.5 Å relative to the input motif atoms, and (iii) no ligand clashes. This reveals the older AME heavy-atom sc-RMSD metric can overstate success by ~2× because many designs preserve local motif geometry but fail global fold recovery.

8 Sampling contribution: an exact, model-agnostic reparameterization of the linear flow-matching interpolant into the EDM noise-level framework, enabling Karras-schedule sampling and stochastic churn without retraining. In a 6-target sampler study, EDM sampling beats Euler–Maruyama SDE sampling, while deterministic Euler ODE sampling fails (0% on all 6 targets).

9 Quality beyond success rate: among strict-success designs, Emyx shows higher novelty and diversity (median TM-score to closest PDB hit 0.475; 441 Foldseek clusters) and strong ligand sterics (90.7% pass ligand-clash filter vs 52.3% for Proteína-Complexa). Geometry validity checks (PeptideBuilder ideal bonds/angles) are also notably higher for Emyx (82.2%) than the baselines under their evaluated pipelines.

10 Analysis insight: spectral/weight-utilization comparisons suggest a generator–predictor gap—predictors initialized with rich MSA/template context show higher effective rank than generators initialized from noised coordinates. Among generators, Emyx exhibits the lowest fraction of “overtrained” trunk layers, correlated with aggressive conditioning bottlenecks in adaLN modulation.

📜Paper: https://t.co/Tb4MGGvXrC
#ComputationalBiology #ProteinDesign #EnzymeDesign #GenerativeModels #DiffusionModels #FlowMatching #Transformers #StructuralBiology #MachineLearning

[风渡fengdu @fengdu2077]

分享一个免费用 Claude 、Gemini、GPT的方法，都是可用他们的最新旗舰模型。

除了御三家，还有国内的明星模型：Kimi-2.6、GLM-5.1、MiniMax-M3

只要下载一个美团的「tabbit 国际版」，记住要用国际版的才有御三家。
用国内版的只有国内的模型。