Watvina (@Biocheming) profile picture

structure based drug screening, drug design, pharmocology

France
Joined April 2012
Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)Photo by Watvina (@Biocheming)
🚨重磅!商汤科技正式发布并开源 SenseNova Loop-9B!9B参数循环架构赶超Qwen3.8-27B! 性能亮点 🔹架构黑科技:32层权重循环出56层深度,参数仅~9B,上下文最高512K,专攻长上下文+Agent任务 🔹全面超越Preview:AutomationBench 36.67、Terminal-Bench 59.55、GAIA 69.90、AIDABench 65.31、BIRD 72.34 🔹硬刚Qwen3.8-27B:AutomationBench持平、AIDABench微胜、BIRD&GAIA领先,Agent/代码/数据分析场景极具竞争力 🔹Loop Scaling验证:循环次数提升最高带来+16.93%收益 小参数深计算,效率拉满,已开源! 🔗 GitHub:github.com/OpenSenseNova/Sen…
🔗 Hugging Face:huggingface.co/sensenova #商汤科技 #SenseNova #Loop9B #开源大模型 #AI智能体 #Qwen #人工智能
🤯惊了!9B参数大模型,只需16G内存即可本地跑通,能力反超Qwen3.5-27B! 用计算深度(Looped Transformer + 参数共享循环)替代堆参数,打出27B级水准。 关键亮点 🔹9B顶27B:WorkBuddy、RaccoonPPT、AutomationBench全面超越Qwen3.5-27B 🔹mHC加持:GPQA Diamond +4.42,TeamBench +13.34 🔹多模态长上下文:支持图文输入,262K上下文,代码/办公/Agent全面强势 🔹极致轻量:量化后轻松跑在16G设备,Apache 2.0开源,支持MTP加速 操作步骤 1️⃣HF下载:sensenova/SenseNova-Loop-9B-Preview 2️⃣安装依赖:pip install "transformers==5.18.0" torch accelerate 3️⃣加载推理(trust_remote_code=True),支持thinking与多模态 4️⃣社区GGUF量化版适合16G纯CPU/低显存本地部署 
huggingface.co/sensenova/Sen… 
github.com/OpenSenseNova/Sen…
 #SenseNova #商汤 #开源大模型 #本地部署 #AI #9B顶27B
34
34
6
446
31,960
SoluProtMut: Siamese Deep Learning for Predicting Solubility Effects of Protein Mutations with Experimental Validation 1 SoluProtMut introduces a structure-based Siamese geometric neural network that predicts whether a mutation will be desolubilizing vs non-desolubilizing, using an anti-symmetric design by construction (swapping WT/mut flips the score), aiming for more reliable mutation-effect directionality. 2 A key practical result: on an independent benchmark of 44 proteins from low-throughput (non-DMS) studies, SoluProtMut achieved the highest informedness among compared tools, with the clearest advantage on multipoint mutants (ternary informedness 26.5%, 8.8 percentage points above the next-best method). 3 The model is trained exclusively on deep mutational scanning (DMS) solubility-proxy datasets (mostly GFP/YSD-based expression/solubility readouts), addressing the common issue that low-throughput solubility data are sparse, heterogeneous, and often biased by design choices (e.g., alanine scanning, selected sites). 4 Architecture details: each branch uses an IEConv-derived graph CNN operating on C-alpha backbones as ordered 3D point clouds; edges are formed on-the-fly within a radius (8→12→16 Å), with kernels parameterized by both Euclidean and sequence distance to distinguish spatial neighbors that are close vs distant along the chain. 5 Input strategy: the network uses WT backbone coordinates while encoding the mutant sequence, avoiding the need for mutant structures and reducing sensitivity to side-chain modeling; residues are represented with an 8D VHSE physicochemical descriptor (found essential for convergence vs one-hot/learned embeddings). 6 Training strategy: five-fold cross-validation yields a 5-model ensemble; class imbalance is handled via class weights (and protein-wise rebalancing for combined-DMS training); geometric augmentation adds small Gaussian noise to coordinates, applied identically to WT and mutant to preserve relative geometry. 7 The study highlights a consistent theme in mutational-effect prediction: strong within-protein performance does not guarantee protein-level generalization. Cross-protein transfer among DMS proteins (TEM, LGK, PKS) was generally weak, and external validation emphasized that generalization across unseen proteins is the main bottleneck rather than raw model capacity. 8 Data synergy was observed: mixing data from multiple proteins and even multiple assays for the same protein improved performance (e.g., combining LGK GFP + YSD improved performance vs GFP-only), suggesting that heterogeneity across assays can act as an implicit regularizer rather than purely a nuisance variable. 9 Experimental contribution: the authors generated a new split-GFP solubility dataset for 277 LinB haloalkane dehalogenase variants (E. coli expression; semi-automated robotic screening for throughput and standardization) and used it, together with independent myoglobin DMS data, for external validation—reinforcing how urgently broader, less imbalanced solubility datasets are needed. 💻Code: github.com/loschmidt/SoluPro… 📜Paper: biorxiv.org/content/10.1101/… #ProteinEngineering #DeepLearning #GeometricDeepLearning #GraphNeuralNetworks #DeepMutationalScanning #ProteinSolubility #ComputationalBiology #Bioinformatics #StructuralBiology
4
28
1,389
Watvina retweeted
🚨 Google resmen kabul etti. AI’lar frontend yazarken hâlâ 2019 seviyesinde takılı kalıyor. Sebep basit: Eğitim verileri eski miras kodlarla dolu. Çözüm olarak Chrome ekibi modern-web-guidance’ı çıkardı. • 212 modern web özelliğini modellere besliyor • Claude Code, Codex, Copilot ve Antigravity ile çalışıyor • Başarı oranı %56’dan %88’e çıkmış Kurulum tek satır: npx modern-web-guidance github.com/GoogleChrome/mode… Sizce sıradaki “kaşıkla düzeltilmesi gereken” alan hangisi olacak?
18
148
8
1,660
117,916
Watvina retweeted
对 PhD 最大的误解,是以为博士训练的终点是掌握某个领域的知识。 其实不是。 本科和硕士更多是在学习如何解决已经被定义的问题,而 PhD 真正困难的部分,是你必须学会在没有标准答案、甚至不知道问题是否成立的情况下,独立定义问题、设计实验并建立证据。 但这里存在一个很微妙的矛盾。 学术界一方面要求博士生成为 independent researcher,另一方面,博士生的研究方向、经费、算力、合作资源乃至毕业时间,往往都深度依赖导师。 我们要求一个人在高度依赖的组织结构中,训练出完全独立的研究能力。 所以我认为,衡量 PhD 训练是否成功,不能只看他能不能独立完成一篇 paper。 更应该看:当导师不再提供题目、资源和研究路线时,他是否还知道什么问题值得研究,以及如何找到实现它的路径。 学术独立性不是你能独自完成多少工作,而是你能否独立决定什么工作值得完成。 一个博士生可以非常擅长执行研究,却始终没有真正学会做研究
大学院で行われているのは、所謂「お勉強」ではなく、訓練に近い。特に博士課程では、研究者として自走するための訓練が行われる。研究テーマの設定や研究計画も自分で決める。指導教員の立場は、例えば未踏でいうところのPMに近い。
5
52
7
301
20,544
Watvina retweeted
Boltzinaのversion 1.1を公開しました🎉 NVIDIA BioNeMoを使って推論速度を向上させる機能をオプションに追加しました。 従来比約2倍速、特にBoltzina (UniDock2)はデフォルトのBoltz-2から最大24.4倍速、1リガンドあたり平均1秒未満でスコアリングできます。 (※H100 1枚、CPU 48コア環境で測定)
2
7
50
2,965
We’ve won our 4th consecutive OpenADMET Blind Challenge, coming ahead of 270 entrants, this time for predicting CYP direct inhibition. Why do we keep winning? The short answer is our data, but it goes deeper than that. Our goal at Inductive is to build superhuman chemical intelligence. In chemistry, the most valuable data remains orders of magnitude slower and more expensive to generate than in domains like coding and math, which have benefited from the near-infinite scalability of compute and reliable verification. Because of this cost, any strategy for building chemical intelligence that relies on brute force data generation alone – even if you have billions of dollars to spend – will lose to a strategy that also aggregates data already being generated as part of economically valuable work. That’s why, from the start, we’ve built products and services that generate immense value to drug discovery scientists while building a consortium to access large amounts of diverse, experimentally relevant chemistry data as a byproduct. That data enables us to build better models for the drug hunters we work with. Doing this effectively requires trust from scientists who have decades of scar tissue dealing with the unending complexities of biology while being told again and again by Silicon Valley that technology is about to change everything. Earning that trust has taken years, and we take it seriously. It’s also in the DNA of our team, where the best ML scientists and software engineers I know work with veteran drug discovery scientists. Together, we dream big while simultaneously thinking pragmatically about how to improve drug discovery today. Thank you to OpenADMET for organizing yet another well-run competition and thank you to our customers for continuing to trust us to support their drug programs.
14
22
7
187
17,283
Announcing LigParGen 3.0: OPLS-AA parameters for organic molecules, now a free app for Mac, Windows and Linux. Paste a SMILES or draw a molecule, and get files for OpenMM, GROMACS, AMBER, CHARMM, LAMMPS and more. jorgensenresearch.com/ligpar…
3
76
4
324
8,728
大規模なタンパク質収量測定と予測 約4万種類のタンパク質について、 大腸菌で発現・精製した際の単量体収量を大規模に測定し、そのデータを使ってアミノ酸配列から収量を予測するAIモデルを開発。人工タンパク質の選別精度の向上を実現 biorxiv.org/content/10.64898…
10
2
103
7,724
我可以自信地说,这是市面上最全的用大量数据、模型语义验证来测评 AI 写作特征的实验。我对 100多项特征做了 50 万次语义判断,最后做成 skill 方便大家使用。很多AI 产品也对写作能力做测评优化,但从未公布完整调试过程。当然,它也有很多不完善的地方,欢迎讨论。
263
445
90
3,530
407,649
Watvina retweeted
One force field for proteins, DNA, lipids and carbohydrates? Introducing SO3LR Tauri, a pretrained ML force field that couples a fast neural network with explicit repulsion, electrostatics and dispersion. It takes a step toward a unified model that describes all major classes of (bio)molecules at quantum-mechanical accuracy. arxiv.org/abs/2611.06237 1/12🧵
13
84
7
504
25,056
RT @jppiquem: #compchem I am thrilled to share our latest research published in @CommsChem. 🔬 Our new article, "Enhanced sampling enables b…
Watvina retweeted
英伟达开源了一个极其硬核的 AI 数据基座项目:auto-ontology。 当前几乎所有做企业级 RAG 的团队都在遭遇同一个天花板:靠传统的向量数据库切块做相似度检索,只要文档跨章节或者存在复杂的上下游依赖关系,大模型就开始一本正经地胡说八道。 解决幻觉的下一站从来不是盲目堆砌几百万的 Context Window,而是把混乱非结构化的私域文档,重新梳理成具备实体关系和确定性逻辑路径的本体论(Ontology)图谱。 Nvidia 这套工具链使用大模型自动化提取领域概念、实体层级与逻辑约束,把构建知识图谱这个过去需要资深领域专家耗费数月手动绘制的苦活,直接封装成了自动化流水线。 给大模型喂一万页 PDF 只是盲人摸象,把知识点编织成一张逻辑严密的拓扑网络,智能体才能在复杂的生产决策中踩在结实的确定性地面上。 开源地址:github.com/NVIDIA/auto-ontol…
6
8
34
2,161
Watvina retweeted
🧬 Design proteins with natural language. Pinal is now online in Nature! Pinal democratizes computational protein design by turning functional intent into protein candidates. Open weights + code: github.com/westlake-repl/Den… Paper: nature.com/articles/s41586-0…
8
88
7
372
28,080
Public preview of BioEmu2! @MSFTResearch AI for Science’s new foundation model for biomolecular structure, dynamics and function. Open weights, MIT license. github.com/microsoft/bioemu2 Movie: BioEmu2-simulated binding of intrinsically disordered peptide PUMA to MCL-1 receptor. 🧵⬇️
24
247
28
915
74,272
How should science address positivity bias? A new two-part Focus series explores its effects on the scientific literature. Learn more ⬇️ scim.ag/4xVQZuR scim.ag/4ry5yD8
1
1
12
2,429
おーん これいいっすね〜
6
46
1
332
18,379
Watvina retweeted
Large Drug Discovery Model is out! I am exited to introduce our new generative SBDD framework. We experimentally validated LDDM across 5 targets and successfully designed novel, validated hits with structural accuracy confirmed by X-ray. Preprint: biorxiv.org/content/10.64898… 🧵 1/7
20
109
7
686
121,478
💥炸了!满血Qwen 3.8 27B 被塞进 Mac 本地就能跑:体积只剩 5.9GB,智商几乎没掉! PrismML 发布 Ternary Bonsai 2 27B:Qwen3.8-27B 血统,三值量化到约 1.76bit,只有满精度的 1/9,综合基准保住 98.2%。不是云端演示,是笔记本直接开跑。 关键突破: 🔹权重仅 5.9GB,不是常见 Q4 的 17–18GB 🔹总分 83.9 vs Qwen3.8 的 85.4(保留 98.2%) 🔹指令跟随 102%,数学 99.5%,代码 99.3% 🔹智能密度 0.444/GB,甩开同级 Q4 / IQ2 / FP16 🔹262K 上下文 + 图文多模态 + Apache 2.0 🔹Apple MLX 可跑,M5 Max 约 45 tok/s;RTX 5090 最高 143 tok/s 以前 27B 要 32GB 统一内存才舒服。现在一台普通 Mac,就能装进接近满血的Qwen3.8! 链接模型:huggingface.co/collections/p… 查看博客:prismml.com/news/bonsai-2-27… #Qwen #Qwen3 #Bonsai2 #LocalLLM #MacAI #开源模型 #边缘AI #MLX
🚨最新:Qwen-Image 2.1确认将开源,发布前先放 50 个内测名额! 面向有经验的创作者和开发者,入选后邮件通知。 
9 月 28 日 23:59(北京时间)前,须发一篇原创作品或实测评测,好评差评都行。 现在还没放权重,先抢资格:alidocs.dingtalk.com/notable… #QwenImage #开源 #AI绘画 #AIGC #ModelScope #阿里巴巴
26
49
4
367
84,030
ColabFold 1.6.3 is out! 2.5x faster, pip-installable, ipSAE+pDockQ2 scores. Thanks Choonghwan Lee, Marielle Russo, Gyuri Kim, Martin Steinegger, Milot Mirdita 🐍 pip install colabfold[alphafold] (1/3)
1
113
4
510
22,188
一个 9B 模型,只加一套会持续进化的 Skill,平均表现就超过了没有 Skill 的 27B 模型。 Google Research 这篇 WikiSkill,研究的是一个大家都一直关注的问题: Agent 每轮任务都在产生经验,但这些经验如果一直散落在执行轨迹里,下一轮其实很难真正复用。 WikiSkill 给 Agent 加了一层持续积累的「Wiki」。 整个系统分成三层: Raw 保存原始成功/失败轨迹; Wiki 把这些轨迹整理成失败模式、有效策略和历史经验; Skill 再把稳定下来的知识编译成 Agent 可以直接执行的流程。 每轮 Skill 更新都会经过验证。有效的留下,效果变差就回滚;但 Wiki 不会清空,所以失败的尝试也会成为下一轮进化的经验。 实验覆盖数学、搜索、Excel、长文档和具身任务等 5 个 benchmark、5 个模型。 结果是 Qwen-3.5-9B + WikiSkill 平均达到 47.4%,超过没有 Skill 的 Qwen-3.6-27B 的 39.4%。 而且 Skill 还能跨模型迁移。有些由其他模型进化出来的 Skill,甚至比模型自己的 Skill 效果更好。 这会让我比较关注 Wiki 这一层。 Agent 自进化如果只有「轨迹 → Skill」,很容易反复重新理解过去。中间多一层持续整理的知识库,才能让成功、失败和被拒绝的修改真正积累下来。 Skill 开始有点像独立于模型权重的知识资产:可以积累、验证、版本化,还能跨模型迁移。 📎 arxiv: arxiv.org/abs/2608.27454
20
58
309
19,723