#ICML #MachineLearning #ArtificialIntelligence #AI #DeepLearning #GenerativeAI #DiffusionModels #LanguageModels #ComputerVision #ImageSuperResolution #AIResearch #Research #Seoul #Korea

Context-dependent calibration of Evo2 likelihood with bacterial fitness: a quantitative characterization across five E. coli datasets 1. The paper argues Evo2 (a DNA foundation model) should be treated as a likelihood predictor, not a universal fitness predictor, because likelihood–fitness calibration changes sharply with variant context (coding vs regulatory), selection regime, and distance from wild type. 2. Using Evo2 7B zero-shot scoring (ΔLLR: change in pseudo-log-likelihood between mutant and reference windows), the authors map when ΔLLR aligns with measured phenotypes across five E. coli datasets spanning DMS, promoter MPRA, and experimental evolution. 3. Strong calibration appears in a plasmid-borne coding gene under stringent antibiotic selection: Firnberg 2014 TEM-1 β-lactamase (13,027 nucleotide variants) shows Spearman ρ = 0.545 overall (SNVs ρ = 0.606; indels ρ = 0.521), with a monotonic reliability curve across ΔLLR deciles. 4. Another strong setting is thermal experimental evolution where adaptive structural variants are enriched: Tenaillon 2012 reaches Insertion AUROC 0.882 (best window W = 2 kb) and Deletion AUROC 0.846 (W = 4 kb). A key methodological point is that window size is type-dependent; a single default window can cost ~5–7% AUROC. 5. The same organism can show decisive lack of calibration in regulatory context: Ireland 2020 RegSeq promoter MPRA (64,665 variant×condition pairs) yields ρ = 0.011, staying flat even after stratifying by conditions, promoters, or canonical −10/−35 motif positions—Evo2 recognizes motifs but does not predict expression changes from SNVs within them. 6. Chromosomal essential-gene DMS also shows near-zero calibration despite being coding: Dewachter 2023 (fabZ/lpxC/murA; 13,128 variants) gives ρ = 0.041 (each gene < 0.06), highlighting a major plasmid-vs-chromosome and/or selection-regime difference not captured by “coding vs regulatory” alone. 7. A clean mechanistic axis emerges from a combinatorial landscape: Papkou 2023 folA (30,000 variants) is intermediate overall (ρ = 0.237), but calibration decays monotonically with sequence divergence from wild type—ρ = 0.575 at 2 mutations down to ρ = 0.065 at 9 mutations—showing an explicit out-of-distribution effect. 8. The authors fit an explicit calibration function across per-dataset/per-divergence strata: ρ = f(divergence, context). Weighted Fisher-z regression gives a negative divergence coefficient (−0.028 per additional mutation) and a negative regulatory-context coefficient (−0.33), with R² = 0.49, presented as an illustrative quantitative “lookup surface” rather than a universal law. 9. They test and refute a simple “training over-representation” explanation for why TEM-1 calibrates better than essentials: chromosomal essentials have far more raw NCBI deposition counts than TEM-1 yet calibrate much worse (calibration does not increase with deposition count; may even decrease). Deposited variant diversity (not copy count) is proposed as a plausible but untested factor. 10. A practical contribution is methodological guardrails for DNA-LM variant scoring: avoid anchor-based indel scoring artifacts (which can silently force ΔLLR = 0), tune window sizes per variant type, and report sign conventions carefully because different variant classes can flip the “adaptive” ΔLLR direction. 💻Code: github.com/sunsungkim04-sys/… 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #Bioinformatics #Genomics #DeepLearning #FoundationModels #LanguageModels #VariantEffectPrediction #Ecoli #DMS #MPRA #ExperimentalEvolution

🇹🇷 TR - Türkçe Başlık: usul365tr 🇹🇷 | Bilge AI Gündemi 🧠 Özet: Yerli ve milli yapay zeka ekosisteminde Bilge AI rüzgarı esiyor! Türkçe dil morfolojisine ve kültürel bağlama tam uyumlu olarak geliştirilen Bilge AI 2.0, kurumsal pazar entegrasyonlarını hızlandırarak yerli işletmelerin dijital dönüşüm omurgası haline geldi. Sektör raporlarına göre, regülasyon uyumluluğu ve veri yerliliği hassasiyeti taşıyan kamu ve özel sektör kuruluşlarının E'i iş süreçlerinde Bilge AI altyapısını tercih etmeye başladı. Yazılım cephesinde ise yeni nesil semantik arama motoru entegrasyonları kurumsal bilgi yönetimini üst seviyeye çıkardı. İstatistik: Güncellenen yerel büyük dil modeli, Türkçe doküman analizlerindeki anlamsal doğruluk oranını  artırdı. Puan: 9.7/10. Yorum 👇 Takip Et 🇬🇧 EN-GB - English UK Title: usul365tr 🇹🇷 | Bilge AI Agenda 🧠 Summary: Bilge AI drives massive performance surges across the regional artificial intelligence space! Fully optimized for Turkish language morphology and cultural nuances, Bilge AI 2.0 has accelerated enterprise integrations, becoming the digital transformation backbone for local businesses. According to industry manifests, 45% of public and private organizations prioritizing regulatory compliance and data sovereignty have officially adopted the Bilge AI framework. On the software front, newly integrated semantic search features have drastically heightened corporate knowledge management. Stats: The updated local LLM boosts semantic accuracy in document processing by 18%. Rating: 9.7/10. Comment 👇 Follow 🇺🇸 EN-US - English US Title: usul365tr 🇹🇷 | Bilge AI Agenda 🧠 Summary: Bilge AI drives massive performance surges across the regional artificial intelligence space! Fully optimized for Turkish language morphology and cultural nuances, Bilge AI 2.0 has accelerated enterprise integrations, becoming the digital transformation backbone for local businesses. According to industry manifests, 45% of public and private organizations prioritizing regulatory compliance and data sovereignty have officially adopted the Bilge AI framework. On the software front, newly integrated semantic search features have drastically heightened corporate knowledge management. Stats: The updated local LLM boosts semantic accuracy in document processing by 18%. Rating: 9.7/10. Comment 👇 Follow 🇩🇪 DE - Deutsch Titel: usul365tr 🇹🇷 | Bilge AI-Agenda 🧠 Zusammenfassung: Bilge AI wirbelt den regionalen KI-Markt auf! Die vollkommen auf die türkische Sprachmorphologie optimierte Version Bilge AI 2.0 beschleunigt die Integration in Unternehmen und wird zum Rückgrat der digitalen Transformation für lokale Betriebe. Branchenberichten zufolge nutzen 45 % der öffentlichen und privaten Organisationen, die Wert auf Datenlokalität legen, die Bilge AI-Infrastruktur. Softwareseitig verbessern semantische Suchfunktionen das Wissensmanagement in Unternehmen erheblich. Statistik: Das lokale Sprachmodell steigert die semantische Genauigkeit bei der Dokumentenanalyse um 18 %. Bewertung: 9.7/10. Kommentar 👇 Folgen 🇪🇸 ES - Español Título: usul365tr 🇹🇷 | Agenda de Bilge AI 🧠 Resumen: ¡Bilge AI revoluciona el ecosistema regional de inteligencia artificial! Desarrollado con total compatibilidad para la morfología del idioma turco, Bilge AI 2.0 acelera la integración corporativa. Informes del sector indican que el 45% de las organizaciones públicas y privadas que priorizan la soberanía de datos eligen Bilge AI para sus procesos. En software, las nuevas funciones de búsqueda semántica elevan drásticamente la gestión del conocimiento empresarial. Estadísticas: El modelo local actualizado incrementa la precisión semántica en el análisis de documentos en un 18%. Valoración: 9.7/10. Comenta 👇 Sigue 🇫🇷 FR - Français Titre: usul365tr 🇹🇷 | Agenda de Bilge AI 🧠 Résumé: Bilge AI bouscule le paysage régional de l'intelligence artificielle ! Parfaitement optimisé pour la morphologie de la langue turque, Bilge AI 2.0 accélère son intégration dans le monde de l'entreprise. Selon les rapports du secteur, 45% des organisations publiques et privées soucieuses de la souveraineté des données privilégient l'infrastructure Bilge AI. Côté logiciel, les fonctionnalités de recherche sémantique renforcent la gestion des connaissances. Stats : Le grand modèle de langage local améliore la précision sémantique de 18% sur l'analyse documentaire. Note: 9.7/10. Commentez 👇 Suivez 🇮🇹 IT - Italiano Titolo: usul365tr 🇹🇷 | Agenda Bilge AI 🧠 Riepilogo: Spinta prestazionale per Bilge AI nel mercato regionale dell'intelligenza artificiale! Ottimizzato per la morfologia della lingua turca, Bilge AI 2.0 accelera le integrazioni enterprise, diventando la colonna vertebrale della trasformazione digitale. I dati di settore indicano che il 45% delle organizzazioni pubbliche e private che danno priorità alla sovranità dei dati ha adottato la struttura di Bilge AI. Lato software, la ricerca semantica eleva la gestione della conoscenza aziendale. Statistiche: Il LLM locale aumenta dell'18% l'accuratezza semantica nell'analisi documentale. Valutazione: 9.7/10. Commenta 👇 Segui 🇸🇦 AR - العربية العنوان: usul365tr 🇹🇷 | أجندة Bilge AI الذكية 🧠 الملخص: قفزة أداء كبرى للمنظومة الإقليمية للذكاء الاصطناعي مع Bilge AI! يقدم نموذج Bilge AI 2.0، المطور بتوافق تام مع مورفولوجيا اللغة التركية، تسارعاً في التكامل المؤسسي ليصبح ركيزة التحول الرقمي للشركات المحلية. ووفقاً لتقارير الصناعة، اعتمدت 45% من المؤسسات الحكومية والخاصة التي تعنى بسيادة البيانات بنية Bilge AI في عملياتها. وبرمجياً، رفعت ميزات البحث الدلالي كفاءة إدارة المعرفة للشركات. الإحصائيات: رفع النموذج المحلي المحسن دقة التحليل الدلالي للمستندات بنسبة 18%. التقييم: 9.7/10. علق 👇 تابع 🇮🇷 FA - فارسی عنوان: usul365tr 🇹🇷 | دستور کار Bilge AI 🧠 خلاصه: دستیار هوشمند Bilge AI شتاب عملکردی فوق‌العاده‌ای در بازار منطقه‌ای هوش مصنوعی ایجاد کرد! مدل Bilge AI 2.0 که با هماهنگی کامل با مورفولوژی زبان ترکی توسعه یافته، اتوماسیون سازمانی را سرعت بخشیده است. بر اساس گزارش‌ها، ۴۵٪ از سازمان‌های دولتی و خصوصی که حاکمیت داده‌ها را در اولویت قرار می‌دهند، این معماری را برگزیده‌اند. در بخش نرم‌افزار، قابلیت‌های جستجوی معنایی، مدیریت دانش سازمانی را ارتقا داده است. آمار: افزایش ۱۸ درصدی دقت معنایی در پردازش اسناد با مدل بومی جدید. امتیاز: 9.7/10. نظر بدهید 👇 دنبال کنید 🕌 OS - عثمانliجه عنوان: usul365tr 🇹🇷 | بيلكه آي (Bilge AI) احوالاتی 🧠 خلاصه: محلی هوشِ صناعي فلكياتنده Bilge AI روزگاری اسییور! لسانی تركي صرف و نحوينه بالکلیّه موافق اولارق تولید اولنان Bilge AI 2.0، شركتلرڭ ديجيتال تحوّل omurgası اولارق جاری در. صنعت راporlarına göre، معلومات محليتنه اعتناء ایدن دولتی ve خصوصی شركتلرنڭ ٪٤٥ی Bilge AI altyapısını اِنتخاب ایدییور. نرم‌afsar جهتinde ایسه ینی نسل معنوي تفتيش نظامى معلومات ادارەسنی اعلا مرتبه‌یه چقاردı. احصائيات: ینی كود اساسı، معلومات تحليلنده معنوي طوغريلق نسبتی ٪١٨ تزايد ایتدی. رتبه: 9.7/10. تعلیق 👇 تعقیب ایت HACKTACS: #usul365tr #takip #beğeni #hacktacts #BilgeAI #LocalLLM #TurkishAI #EnterpriseAI #SemanticSearch #DataSovereignty #DigitalTransformation #RegulatoryCompliance #LanguageModels #NLP #TechInnovation #BreakingNews

AI is changing the future of materials discovery. MIT Professional Education's Applied AI for Materials Discovery course explores how foundation models, generative AI, and agentic workflows can accelerate scientific discovery and materials design. Gain practical insights from MIT faculty and learn how AI is transforming research and innovation. 📅 Registration is ending soon. 🔗 Register here: professionaleducation.mit.ed… #MITProfessionalEducation #AI #LanguageModels #GraphNeuralNetworks #ComputerVision

Opengerminal: An open-source implementation of the germinal antibody design pipeline 1 OpenGerminal re-implements the Germinal de novo antibody (VHH) design pipeline with a fully open-source stack, removing two major licensing bottlenecks in the original release: PyRosetta (restricted) and IgLM (non-commercial). 2 The most notable benchmark result is that swapping IgLM for AbLang1 as the hallucination guidance model substantially increases the fraction of trajectories that survive the initial Chai-1 cofolding validation: PD-L1 33.7% vs 18.6%, IL-3 24.6% vs 8.0 (OpenGerminal vs Germinal). 3 OpenGerminal follows the same 4-stage architecture: (1) AlphaFold-Multimer hallucination with PCGrad merging structural objectives (pLDDT/iPTM/PAE) and an antibody language-model naturalness loss, (2) Chai-1 cofolding relaxation interface scoring filter, (3) AbMPNN CDR redesign, and (4) Chai-1 re-fold stricter structural/physicochemical filters. 4 PyRosetta is replaced in stages 2 and 4 by an open-source backend: OpenMM 8.5.1 for relaxation, FASPR for side-chain repacking, FreeSASA Biopython for interface geometry, and sc-rs v1.0.0 for Lawrence–Colman shape complementarity. The goal is API-compatibility with the original PyRosetta utility module while enabling redistribution. 5 Tradeoff: per-trajectory hallucination time increases by ≥1.5× (A100 80GB): PD-L1 4.4 vs 3.0 minutes, IL-3 4.2 vs 2.6 minutes. End-to-end wall time can increase further because more trajectories proceed into the more expensive downstream filtering stages. 6 For accepted PD-L1 designs, OpenGerminal shows equivalent or improved Chai-1 confidence metrics: higher median pLDDT (0.908 vs 0.889), higher PTM (0.892 vs 0.875), and higher interface pLDDT (0.917 vs 0.899), while iPTM and interface PAE are similar between pipelines. 7 Interface geometry quality on accepted PD-L1 designs is broadly comparable: both pipelines produce clash-free designs positioned near target hotspots with similar CDR3–hotspot contact counts. One difference is a modestly lower fraction of interface residues located in CDRs for OpenGerminal (median 82.8% vs 100%), suggesting slightly different interface geometries. 8 The paper also provides post-hoc validation by rescoring OpenGerminal accepted structures with PyRosetta (without re-relaxation) and shows strong correspondence for shape complementarity (Spearman ρ=0.882) and good correspondence for interface hydrophobicity (ρ=0.763), supporting that the open-source metrics track established Rosetta-style evaluations. 9 Multi-chain target support is extended and debugged (chain parsing/renaming and Chai-1 sequence lookup issues), and verified to run on the official insulin example without error. The run produced cofolding-pass trajectories but no final accepted designs, highlighting that multi-chain success remains an open challenge rather than a solved capability. 10 Known limitations are explicitly documented: several Rosetta-energy-like metrics are currently placeholders (binder_score, interface_dG, interface_hbonds), which degrades stage-2 ensemble selection and effectively disables an H-bond filter in stage 4; future work is proposed to recalibrate OpenMM-derived energies and thresholds and improve ensemble selection with accessible metrics (e.g., interface ΔSASA). 💻Code: github.com/teaninja/OpenGerm… 📜Paper: biorxiv.org/content/10.64898… #computationalbiology #proteinengineering #antibodies #antibodydesign #openscience #alphafold #openmm #languagemodels #bioinformatics #HPC

Demystifying LLMs 🧠 Ever wonder how AI "understands" language? 🤖 It all starts with vast amounts of data. 📝 This technology is transforming industries! Ready real-world LLM solutions : novelvista.com/corporate/llm… #LLMs #AI #NLP #LanguageModels #AITech

Breakthrough in Language Models: AI Just Got a LOT Smarter! New study reveals a game-changing technique to make large language models work across languages with 25% more accuracy! arxiv.org/abs/2606.06586 #LanguageModels #AIAdvancements

✈️ Looking forward to an exciting #SIGMOD2026! 👇 Schedule below. 🇮🇳 See you soon in #Bengaluru! #Databases #LLMs #LanguageModels #DatabaseBenchmarking #QueryOptimization #QuantumComputing @lojil192574 @sigmod @SIGMODConf

📄 Paper: arxiv.org/abs/2602.19066 💻 Code: github.com/David-cripto/IDLM 📢 Invitation post: x.com/diffusion_llms/status/… Many thanks to the organizers for the invitation, @jdeschena, @ssahoo_, @zhihanyang_! 🙌 #ICML2026 #DiffusionModels #LanguageModels #GenerativeAI #MachineLearning

Building GLM-X: a graph-native cognitive architecture beyond transformers, now learning language from structured world knowledge. 👽🔥 #CognitiveAI #artificialintelligence #languagemodels #worldmodels

Our collaboration with @AnthropicAI is key to advancing safe, enterprise grade AI. Hear from @janaksevak on how Anthropic and Infosys came together in Bengaluru - training 3,000 engineers and running a hands on #hackathon with Claude models to solve real enterprise challenges at scale. Watch the video. #EnterpriseAI #InfosysTopaz #ResponsibleAI #LanguageModels

Separating selection from mutation in antibody language models 1. The paper argues that standard masked antibody language models (e.g., AbLang2) unintentionally learn nucleotide-level biology (codon accessibility and somatic hypermutation context) and that this “mutation signal” can actively hurt prediction of functional effects of amino-acid substitutions. 2. A concrete diagnostic: when scoring substitutions at a masked site, AbLang2 assigns ~100× lower probability to amino acids that require multiple nucleotide changes from the wild-type codon versus those reachable by a single nucleotide change, revealing codon-table imprinting rather than pure functional constraint. 3. The same confounding appears at the site level: AbLang2’s amino-acid probabilities correlate strongly with neutral SHM mutability estimates across V-encoded regions, indicating the model’s outputs are entangled with mutation-rate variation independent of selection. 4. This entanglement degrades functional prediction: on a large deep mutational scan of antibody expression, AbLang2’s correlation with experimental effects drops markedly for substitutions requiring 2–3 nucleotide changes (e.g., ~0.49 for 1-nt-accessible vs ~0.30 for multi-nt-accessible), consistent with mutation bias distorting “fitness” scores. 5. The proposed solution is a factorized mutation–selection framework: a Deep Amino Acid Selection Model (DASM) learns per-site, per-amino-acid selection factors, while a separate fixed neutral nucleotide mutation model supplies codon mutation probabilities; functional effects are thus isolated in the selection term. 6. Training uses phylogeny-derived parent–child pairs from B cell clonal families (≈2 million PCPs). The likelihood of each child codon is modeled as neutral codon mutation probability over branch length t multiplied by a DASM-predicted selection factor for the corresponding amino-acid change, with joint optimization over t and DASM parameters. 7. The neutral mutation component is trained on out-of-frame (non-functional) sequences and includes a “multihit correction” to better match observed rates of 2- and 3-nucleotide codon changes, addressing SHM clustering effects that simple independent-site models underestimate. 8. Benchmarking (zero-shot) shows consistent gains: on FLAb’s largest datasets, DASM substantially outperforms AbLang2 and also beats large general protein LMs (ESM2) and an autoregressive model (ProGen2) for both binding and expression prediction, with strong correlations for heavy and light chains (e.g., expression ~0.69 heavy / ~0.67 light in one key dataset). 9. Beyond DMS, the factorized model better predicts natural affinity maturation trajectories: combining neutral mutation probabilities with DASM selection improves prediction of where nonsynonymous mutations occur and lowers “conditional perplexity” for the identities of observed mutations (median ~4.88 vs ~7.39 for AbLang2 on held-out PCPs). 10. Practical impact: the best-performing DASM is ~4M parameters (vs 45M AbLang2; 650M ESM2) and outputs selection factors for all substitutions in one forward pass, yielding major speedups (on CPU ~0.0097 s/seq vs ~11.05 for AbLang2 and ~112.6 for ESM2), enabling laptop-scale antibody mutational scanning with interpretable per-mutation selection readouts. 💻Code: github.com/matsengrp/netam ; github.com/matsengrp/dasm-ex… 📜Paper: doi.org/10.7554/eLife.109644 #ComputationalBiology #ProteinEngineering #Antibodies #MachineLearning #DeepLearning #Evolution #SomaticHypermutation #LanguageModels #Bioinformatics #Immunology

HalluCodon enables species-specific codon optimization using multimodal language models 1 HalluCodon is a plant-focused codon-optimization framework that fine-tunes pre-trained protein and RNA language models to generate species-specific coding sequences, aiming to improve heterologous protein expression beyond simple codon-frequency heuristics. 2 The core idea is a two-module scoring system: CodonNAT quantifies “codon-context naturalness” (how compatible a CDS looks relative to endogenous host genes), while CodonEXP predicts the probability that a CDS will yield high protein abundance using experimental protein abundance labels. 3 CodonNAT is built via joint fine-tuning of ESM2-650M (protein LM) and mRNA-FM (codon-token RNA LM) under a masked-language-modeling objective, learning host-specific codon context signatures rather than only per-codon frequencies. 4 Across 15 plant species (including maize, rice, tobacco, wheat, tomato, potato, grape, etc.), CodonNAT achieved higher masked-codon prediction accuracy than a “pick the most frequent codon” baseline (average 66.5% vs 56.6%), with especially strong gains for amino acids with higher synonymous-codon diversity. 5 CodonNAT also showed biologically meaningful signal in a non-plant benchmark: in E. coli ccdA synonymous-mutation fitness data, it improved correlation between predicted and measured fitness (Spearman 0.41) compared with frequency-based scoring (0.32) and slightly above CodonTransformer (0.39), supporting that it captures context effects relevant to cellular fitness. 6 CodonEXP integrates nucleotide-level and protein-level information by learning from both CDS and amino acid sequence features, supervised with protein abundance data (PaxDb-derived labels: top 33% vs bottom 33%). It reached ~79.3% average accuracy and 86.1% average AUC across the 15 plant species, and outperformed RNA-only language model baselines in maize/rice/tobacco comparisons. 7 For sequence generation, HalluCodon offers (a) a genetic algorithm (CodonGa) and (b) a hallucination-style, gradient-guided optimizer (CodonHa). Both maximize a Fitness score defined as Naturalness (CodonNAT) × Expression probability (CodonEXP), but CodonHa converges far faster in compute. 8 In a tobacco DsRed2 optimization example, CodonHa reached near-maximal predicted expression probability in only a few iterations and ran ~46.8× faster than the genetic algorithm on the reported GPU setup, while maintaining codon-context compatibility. 9 Experimental transient expression in tobacco leaves tested five proteins (DsRed2, mCry2Ab, GAT, infliximab-A, infliximab-B). For DsRed2, CodonHa produced the strongest fluorescence and higher protein levels by Western blot (reported 1.57× vs CodonTransformer, 4.32× vs Genewiz, 13.58× vs a frequency baseline), suggesting the combined NAT EXP objective can translate to wet-lab gains. 10 The study highlights GC3 as a learned and actionable plant expression feature: HalluCodon optimization tends to increase GC3 toward host-like levels, and a GC3-rewarding variant (Ha-GC3) enabled expression of larger proteins that were difficult under the default CodonHa, while warning that extreme GC3/GC increases can complicate synthesis and increase methylation-site density. 💻Code: github.com/YuxuanLou/HalluCo… 📜Paper: biorxiv.org/content/10.64898… #CodonOptimization #PlantSyntheticBiology #ComputationalBiology #Bioinformatics #LanguageModels #DeepLearning #ProteinExpression #Transgenic #MolecularFarming #ESM2 #RNAFM

Researchers say internal emotion-like signals shape how large language models make decisions. dlvr.it/TRsjK9 #AI #ArtificialIntelligence #EmotionVectors #LanguageModels #MachineLearning

Language models have shown effectiveness in a variety of software applications, particularly in tasks related to automatic workflow. By @languagemodels #llms #ondevicelanguagemodels

Agentic AI in Radiology: Evolution from Large Language Models to Future Clinical Integration doi.org/10.1148/ryai.250651 @judywawira @k_bressem @Tugba_Akinci_MD #agent #LLM #LanguageModels

💎 NLP Unleashed best practices revealed New comprehensive guide covering: ✨ Core concepts 🔧 Practical examples ⚡ Performance tips 🎯 Best practices Dive in 👇 🔗 kubaik.github.io/nlp-unleash… #Blockchain #AR #WebDev #AIEngineering #LanguageModels

Emergent Biological Realism in RL-Trained DNA Language Models 1. The paper studies whether reinforcement learning (RL) post-training can steer a DNA language model toward “biological realism” in plasmid generation, using plasmids as a tractable testbed with clear functional constraints and major biotech relevance. 2. Using Group Relative Policy Optimization (GRPO) on top of PlasmidGPT with a biologically motivated reward, the RL model reaches a 77% in silico QC pass rate vs 5% for the pretrained baseline (and 10% for supervised fine-tuning), across both weak prompts (single “ATG”) and structured prompts (e.g., a GFP cassette). 3. The reward is intentionally lightweight and interpretable: (i) functional annotation scoring (ORI, selectable markers, promoters/terminators, CDS) with an added promoter→CDS→terminator cassette organization bonus, (ii) a length prior favoring typical 5–15 kb plasmids, and (iii) a penalty for long exact repeats (≥50 bp) linked to instability/recombination. 4. A key observation: beyond what is explicitly rewarded, the RL model’s samples align with real engineered plasmids on multiple distributional properties that were not directly optimized—thermodynamic stability (MFE density), codon usage patterns, and ORF length distributions—suggesting correlated “realism” emerges when optimizing for a subset of structural constraints. 5. Quantitatively, RL generations are closest to real plasmids in GC content (mean 0.518 vs real 0.517; matched variance), 3-mer/codon composition similarity (lowest mean Jensen–Shannon divergence to real), and length/stability distributions (median length 6668 vs real 6690; reduced variance; MFE density mean close to real with much smaller variance than base/SFT). 6. Novelty is reduced but remains substantial: 67% of RL generations are classified as novel by BLASTn thresholds (vs 91% base, 96% SFT). Importantly, RL yields many more sequences that are both QC-passing and novel (60% of all RL samples), indicating improvement is not explained purely by memorizing known plasmids. 7. RL concentrates probability mass (diversity drops from 0.915 to 0.588 by pairwise 21-mer Jaccard distance), consistent with learning conserved “successful motifs” (e.g., reliable ORIs/markers) without collapsing to identical outputs; the paper frames this as a practical quality–diversity trade-off. 8. Unlike the common “alignment tax” seen in NLP RLHF, RL post-training does not degrade next-token prediction on held-out continuation; it slightly improves mean log-probability (base −12.449 vs RL −10.966) and substantially reduces variance, suggesting more consistent modeling while being optimized at the sequence level. 9. The paper also probes reward hacking concerns: CDS-related reward uses Prodigal (statistical gene prediction, not homology). The RL model shows lower CDS surprisal on real plasmids than the pretrained/SFT models, supporting the interpretation that RL sharpened general plasmid-like structure rather than overfitting to quirks of the reward. 10. Limitations and outlook: evaluation is bioinformatics-only (library-dependent, potentially penalizing truly novel functional parts like unseen ORIs), and wet-lab validation is not performed here. The authors argue this is groundwork for conditional plasmid design via richer prompts (host, copy number, expression goals) to recover diversity and improve practical utility. 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #Genomics #SyntheticBiology #DNA #LanguageModels #ReinforcementLearning #RLHF #Bioinformatics #PlasmidDesign #MachineLearning