BioAgent Daily

iGC models germinal-center affinity maturation as two stages: learn somatic hypermutation first, then RL-select for binding without throwing away that grammar. - Stage 1 mutation-site pos@1: 11.929% vs best baseline DASM 5.307% — more than doubles; joint accuracy 4.864% vs 2.153% (Tang test; paper) - After RL: 57.0% of predicted mutation positions sit in WRC hotspots vs 2.0% random background; mutability still correlates with S5F (Pearson 0.474) - Under 2–4 mutation budgets, exactly rediscovers 110 of 799 affinity-maturation sequences from the evaluation set - Reward–Tax Meta-Gradient (RewardNet + TaxNet) is there to stop the usual RL cheats — repeating one high-reward sub, or parking mutations in tolerant framework sites Not a wet-lab campaign. It’s a mutation policy that stays SHM-shaped while chasing affinity.

PharmAgents2 runs end-to-end drug discovery as parallel agents with a decision knowledge base — then hands molecules to humans for wet-lab checks. - PharmAgents-Bench v0.1: 121 questions / 9 decision scenarios + 3 end-to-end tasks; highest process scores vs Claude Code, Codex, and LLM-only (same DeepSeek V4.1 Flash base; paper) - LpxH: hypothesized an unreported allosteric site; 63 purchased candidates → AIPher-L1 IC50 8.38 μM; SPR 1.40 μM; MST 0.87 μM (human-run assays in the paper) - JAK1: prioritized covalent Cys817 strategy; 34 purchased → AIPher-J1 matched positive-control pathway inhibition at 10 μM; 3 more >50% in both pathways - Self-Scaling Discovery also nominated a new binding target for a known bioactive; interaction later confirmed experimentally (paper) The agent wrote the plan and shortlist. Humans bought and assayed. Binding-site ID for LpxH is still ongoing.

The UK launched the AI Biosecurity Lab — first dedicated AI×biosecurity org in the government's framing — with £100M public money and up to £200M in company commitments. - OpenAI: £100M compute/model credits + trusted access via Rosalind Biodefense - Anthropic: up to £100M credits, trusted model access, dedicated technical support - Also named: Google DeepMind research collab; Basecamp Research AI-biology models - Initial focus: earlier detection of biological threats + faster vaccines / treatments / MCMs - Structure: independent non-profit; partners UKHSA, APHA, Dstl, MHRA Policy-level bet: put frontier model access next to national biosecurity centres, not just another safety white paper.

Toronto's El Agente Fármaco turns a plain-language discovery goal into a multi-tool screen — and keeps every molecule, pose and metric in Fármaco Atlas with provenance. - 22-task bench vs plain tool calls and Codex: highest mean model-judged score; lower median cumulative tokens and peak context (authors) - TBXT retrospective: agent shortlist mean QED 0.69 vs expert-purchased 0.52; Lipinski violations 0.06 vs 0.54; max Tanimoto to purchased set 0.30 (common docking protocol) - Progressive disclosure: subset_id for whole libraries so the LLM never sees 200 SMILES at once - Tools span MolMIM / PocketXMol / DiffDock / Boltz-2 / ADMET-AI / RBFE — still computational; wet-lab pending on several campaigns The interesting bit isn't another chat wrapper. It's a drug-design-native database the agent can query without stuffing the context window.

ProteoCraft treats binder design as proposal–verification search: truncated diffusion gradients propose mutations; full ESMFold2 decides what's worth keeping. - 14 de novo settings / 8 targets; MCTS Pair iPTM 0.721 vs Best-of-N 0.660 (paper) - Strict yield (Pair iPTM ≥0.75 + binder pLDDT ≥0.9): 114/210 = 54.3% vs Best-of-N 40/210 = 19.0% - AF2-IG strict pass: 86.2% vs 78.6% (same sequences, outside the search verifier) - Wet lab PD-L1 miniproteins: 29/35 hits = 82.9%; best SPR KD 0.879 nM; IFNAR2 VHH 4/9 expressed hits, KD ~153 nM (no affinity maturation) Gradients steer where to look. Full folds decide what ships.

建模 好像掺进了梁圣

Axiom formalized all four Weil conjectures in Lean, including the Riemann hypothesis over finite fields. - Rationality, functional equation, Betti-number comparison, and |α| = q^(i/2) for Frobenius eigenvalues - Repo reports sorry-free, no axioms beyond Lean's standard ones (Axiom) - The key weight estimate comes from their Stacks Project formalization, via Laumon's route rather than Deligne's 1974 proof line by line - Next target named: Ramanujan's tau bound |τ(p)| ≤ 2p^(11/2) Deligne's theorem is now a library call other formal proofs can use.

The Aaronson–Ambainis conjecture, open since 2008, now has a Lean-checked proof found by an AI research system. - PSI's internal system, running a non-public Anthropic model, had a full proof on day 3 of the run (PSI) - Lean accepted it; the system then improved the result twice - Bound: some bit has influence at least c·v^(5/2)/d^6, c about 0.0025 - Meaning: no exponential quantum speedup on yes/no questions about random inputs without structure Stanford theorists who proved a special case in August are now writing the expert exposition.

Two DeepMind AI agents designed, ran and debugged a 1024-bit RSA root computation in under four days. - First root: 391 CPU core-years, 3.5x less than the original 1024-bit run (same algorithm, polynomials and bounds) - New affine-oracle cost: L[1/3, 1.8171] down to L[1/3, 1.5803] per root - Setting: a temporary signing oracle for older formats like PKCS#1 v1.5, not plain factoring - Authors: no threat to deployed 2048-bit+ RSA The attack didn't get scary. The research loop got fast.

X AI 算法更新 🚨 X 昨天更新了他们的开源算法，现在早期点赞变得更加重要。小号应该专注于：撰写人们会点赞并读到结尾的帖子（所以要写得值得一读） tl;dr： - X 的开源代码在 11 月 5 日获得了新分数，这是 3 周以来的首次变更 - 点赞：0.66 → 0.708（迄今最高） - 帖子停留时间：0.006 → 0.0075 - 点击进入帖子：0.3 → 0.274 - 打开视频：0.02 → 0 - 新规则：AI 回复可能在大号下被标记为“llm slop”，并且可能无法到达该人的通知 吸取的教训：点击不再是获胜关键。有人停留并点赞才是。 完整细节： - 新数字 上升：点赞 0.66 → 0.708，帖子停留时间 0.006 → 0.0075，转发 2 → 2.118。 下降：点击进入帖子 0.3 → 0.274，点击后停留 10+ 秒 0.4 → 0.385，打开视频 0.02 → 0。 “不感兴趣”变得更严厉了，-52.27 → -53.07。 复制链接、回复、引用、私信分享和关注没有变化。 X 没有解释这些。这是代码现在所说的。 （我每天用 Claude 检查开源算法，并与你分享更新） - 看看这在一个月内发生了什么 10 月初，一个点赞价值 0.5。现在是 0.708。 停留时间是 0.004。现在是 0.0075，几乎翻倍。 点击是 0.3，而且在下降。 每个变更都指向同一个方向。X 继续为停留的人支付更多。 - 如果你是小号，为什么这很重要 我的解读：你无法靠大数字获胜，但你可以靠看到你帖子的人获胜。如果他们停留并点赞，那今天比上周更有价值。而且那些早期点赞也决定哪篇新帖获得额外推送。 - 视频和你的个人资料 X 还在其排名中启用了两个新猜测：某人打开视频后继续观看多久，以及点击后在个人资料上停留多久。 我无法告诉你这些价值多少。唯一列出它们的文件的日期已过时，所以我不会猜测一个数字。 我的解读：打开视频在同一天降到 0，因为 X 更关心继续观看的人。所以让前几秒钟引人入胜，并让你的个人资料值得停留。 - 大号下回复的新规则 X 对超过 25 万粉丝的账号下的回复运行 AI 检查。被标记的回复可能获得一个持续 14 到 30 天的标签。其中一个标签 literally 叫“llm slop”（AI 垃圾，正如我解读的）。 那部分不是新的。这是：X 的规则现在说，被标记的回复或提及不会到达该人的通知，除非他们关注你。 灰徽账号和 X 评为高度可信的账号跳过这个标签。小号没有这样的豁免。 我能看到规则，我看不到每个通知是否都通过它。无论如何，用你自己的话写回复。 - 没有变化的部分 你的新帖仍然在头 2 小时和前 200 次浏览中获得额外推送，而且它们仍然不会首先被削减。 以下是你应该做的 recap： - 点赞刚刚达到 0.708，这是它们有史以来价值最高的一次 - 某人花在你帖子上的每一秒比上周价值高 25% - 点击后停留 10+ 秒额外赚取 0.385（在点击基础上） - 用你自己的话回复，尤其是在大号下 - 让你的个人资料值得一访，X 现在猜测人们在那里停留多久 - 对于视频，早早抓住注意力，仅仅打开现在得分 0 - 被读到结尾也有助于 X 找到新读者 - 低于 5 万粉丝的账号的原创帖可以跳到大约第 16 位 - 早期观众点赞是赢得那个提升的关键 - 你的头 2 小时和 200 次浏览是窗口，所以在你的人在线时发布 - 最大的单一数字仍然是 20，对于某人复制你的链接 - 引用、私信分享和回复各赚 5 分 - 互关之间的来回价值 20，所以互关你的常客 - 陌生人只显示你的原创帖 - 第 1 个点赞和第 32 个点赞各打开一条通往新人的新路径 - 你在同一 feed 中的第二篇帖降到 62%，所以让一篇帖发挥作用 - 视频列表想要超过 10 秒的剪辑 - 当一篇帖很可能为你赚取新粉丝时 +4 - 你的置顶帖可以影响整个账号，所以看看它 你不应该做的： - 在大号下粘贴 AI 回复，“llm slop”是 X 的真实标记 - 复制粘贴回复，重复标签意味着它们永远不会被通知 - 追逐点击，它降到 0.274 并且在下降 - 偏题帖子，一个可能的“不感兴趣”成本 -53.07 - 仅被打开的视频为打开赚取零分 - 在凌晨 3 点发布会烧掉你的 2 小时窗口 - 依赖回复和转帖到达新人，它们留在你的粉丝中 - 获得侮辱标签，非粉丝完全看不到那篇帖 - 低质量和虚假信息标签会带来警告，非粉丝无法点赞或分享 - 你最近 5 篇帖中的 3 篇有成人媒体会让整个账号被标记 - X 在 64 和 512 个点赞时重新检查视频的隐藏内容 - 在大号帖下堆积你的回复可能沉没你自己的分数 - 账号上一个标签，非粉丝就停止看到你 - 置顶帖中链接到低评分的网站风险 7 天标签 - 逐字重复你自己的措辞风险 copypasta 标记 - 将低质量链接与陌生人标签配对在 X 看来像垃圾邮件 - 互动诱饵无论你多大都会被标记 - 连续发几篇，它们会为同一 feed 互相竞争 - 标签填充什么都不做 - 指望书签，它们价值 0 - 期待 For You 在 48 小时后显示一篇帖

OpenAI and Ligo built a benchmark from messy, real bench data, not curated paper figures: BioBasicsBench. - 100 routine wet-lab tasks: raw gels, blots, colony plates, microscopy, flow/plate-reader exports, sequencing reads, bench videos - GPT-6 Astra 66.7% vs GPT-5.6 Sol 43.8% at xhigh (OpenAI-reported mean reward) - 7 models, 37 settings, 18,500 graded attempts - Astra still misreads controls and overcounts one colony plate by ~12% Reading a gel is easy. Knowing when the data can't support a conclusion is the hard part.

DeepMind put Cheminformatics Skills into Google Antigravity — RDKit plus ChEMBL/PubChem under an agent workbench. - Demo on lung-cancer TKIs: literature triage → compound sets → matched molecular pairs → potency models → rigidified macrocycle proposals with routes - Authors say that loop used to take multiple days; the walkthrough finishes in about an hour - Offering 10,000 scientists the top Antigravity usage tier free for 6 months (eligibility/capacity; rolling onboard) Designs are hypotheses — still need synthesis and assay. Soft speed claim, but the tooling stack is concrete.

USTC's structRFM hit Nature — a structure-guided RNA foundation model, with weights and 21M pretraining pairs open. - Pretrained on 21 million RNA sequence–secondary-structure pairs - Pair-matching MLM + multi-source structure ensemble (MUSES) - Tertiary adapter Zfold: competitive with AlphaFold3 on CASP15-natural, CASP16, RNA-Puzzles (authors) - Also lifts function tasks; dataset/weights/code public RNA FMs that bake in base-pairing, not just sequence. Exact CASP score tables live in the paper/SI.

CellCanon is an agentic cellular world model — predict single-cell responses, explain with citations, then update from new measurements without retraining weights. - 339 held-out SciPlex3 conditions - DE AUPRC / Macro-F1: +63.3% / +17.3% over the best baseline for each metric; direction Macro-F1 +7.2% - Token use vs gene-wise LM baseline: −99.77% Predict, −98.41% Explain - With one measured source cell per drug, Update beats fine-tuned baselines on DE AUPRC/F1 without changing weights Prediction + source-grounded explain + observation update in one loop. SciPlex3 only so far.

Firecrawl opened its life-sciences Research Index inside ChatGPT and Claude — free. - 41M+ papers across drug discovery, clinical trials, and biology (Firecrawl) - The agent searches the index and pulls papers/passages instead of copy-paste abstracts - Same index also reachable via API / MCP / CLI Useful if your agent keeps bouncing between PubMed and chat. Corpus size is Firecrawl's own claim.

Edison published a Kosmos Target Assessment case study with Incyte — packaging target reviews for committee, not replacing the biologist. - Average package across categories: under one workday (Edison's numbers) - Retrospective vs human packages: broader coverage, biggest gains outside core biology - On 6 targets with known committee decisions: proposed derisking experiments matched perfectly or conceptually in 83% of cases; partial 16%; miss 1% (Edison) Retrospective only — doesn't show better go/no-go on new targets. Still a concrete ops number for AI in target selection.