←
BioAI Frontier67 posts
BioAI Frontier avatar

BioAI Frontier

@BioAIFrontier

Making sense of the frontier where AI meets biology.

◷ Joined November 2026▧ 67 Photos and videos
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·6h

🧬 FoldMapper brings evolutionary information into protein inverse folding—without needing the sequence it is supposed to design. On TED domains, it achieved 60.0% best-of-100 sequence recovery versus 46.6% for ProteinMPNN under the same evaluation setting. The November 17 preprint by Haelyn Kim and Juyong Lee addresses a circular dependency: evolutionary profiles usually come from sequence alignments, but inverse folding starts with a backbone and asks which sequence should fit it. 🔍 FoldMapper starts with shape. → Retrieve structurally similar proteins using Foldseek, without a query-sequence homology search. → Aggregate their amino-acid identities into a position-specific frequency profile: which residues occur at corresponding structural positions? → Combine that profile with backbone geometry through a multiscale graph encoder and a structure-conditioned message-passing network. The model therefore uses known proteins’ sequences, but retrieves them through structural similarity rather than requiring a starting sequence for the design. 📊 Reported best-of-100 sequence recovery: → TED domains: 60.0% vs ProteinMPNN’s 46.6%. → PDB: 64.9% vs 50.3%. → Non-peptide binders: 51.3% vs 46.2%. → Peptide binders: 53.1% vs 51.0%. FoldMapper ranked first on TED and peptide binders, and second on PDB and non-peptide binders. It also sat on the recovery–diversity Pareto frontier, offering competitive trade-offs between recovering native residues and generating varied sequences. 🧩 Beyond native-sequence recovery, its predicted structure self-consistency surpassed all tested baselines on RFdiffusion-generated monomer and binder backbones, while matching the strongest baselines on natural proteins. 💡 The contribution is a sequence-independent route to retrieve evolutionary context for backbone-conditioned design. It gives inverse folding an additional source of information beyond geometry alone. ⚠️ Preprint; computational benchmarks, without wet-lab validation. Best-of-100 recovery is not single-sample accuracy or an experimental success rate. Predicted structural consistency does not establish binding or biological function. Code, checkpoints and benchmark resources are publicly available. The code is Apache-2.0; the preprint is CC BY-NC 4.0. 🔗 Paper: biorxiv.org/content/10.64898… 💻 Code: github.com/haelyn-kim/foldma… 📦 Model checkpoints: doi.org/10.5281/zenodo.24186… #ProteinDesign #InverseFolding #AIforScience #BioAI

Post attachmentPost attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·11h

🧬 Only 29 of 107 literature-reported AI nanobody binders showed binding in Spiritein’s retest—a 27.1% hit rate. Separate design campaigns yielded 9 binders from 221 IL-7Rα candidates and none from 105 PD-L1 candidates. Spiritein’s assessment asks a practical question: how well do published binding results and computationally selected designs hold up in a standardized experimental pipeline? 🔬 Published binders: 29 positives out of 107. Spiritein retested literature-reported sequences through standardized expression, purification and characterization. Seventy-eight showed no detectable binding under the retest conditions. An important qualification: among 27 confirmed binders with paired KD measurements, the report found no significant overall difference between published and retested affinities. The main discrepancy was whether binding was detected—not uniformly weaker affinity among confirmed binders. 🧪 New designs: approximately 1.16 million sequences generated. For IL-7Rα, the team compared 13 workflows combining structure-generation and sequence-design methods. Approximately 960,000 sequences underwent computational screening, with final ranking by AlphaFold3 ipTM. A third-party CRO tested 221 selected candidates: → Nine binders overall: a 4.1% hit rate. → Eight came from BoltzGen paired with ProteinMPNN or SolubleMPNN. Each route yielded 4/20 hits; the best KD across the two routes was 8.82 nM. → One came from PPIFlow paired with SolubleMPNN. For PD-L1, four workflows—BoltzGen, Germinal, OpenDDE-harness and HalluDesign—yielded no detectable binding hits among 105 candidates, despite nearly all being successfully expressed and purified. 🧩 Binding was only the first checkpoint. Confirmed binders were also assessed for thermal stability, heat-induced aggregation and binding to baculovirus particles, an indicator of nonspecific interactions. Results varied widely, and some candidates showed elevated nonspecific-binding signals. 🧭 The report draws three conclusions: → Reproducibility: some previously reported high-affinity binding results could not be reproduced under this round’s testing conditions. → Preclinical readiness: binding alone does not establish a development candidate. Binding kinetics, stability, aggregation and specificity remain important hurdles. → Performance on established targets: current AI workflows still struggle to reliably deliver experimentally validated binders on targets with mature antibody-development pathways, as illustrated by the PD-L1 results. ⚠️ These conclusions come from a company-reported assessment under specific design, selection and assay conditions. They should not be treated as universal model failure rates or proof that all original experiments were invalid. Spiritein says it plans to release the code, workflows and raw experimental data through AbGround, supporting closer examination and independent reproduction of the results. 🔗 AbGround / Spiritein: spiritein.com/ #AntibodyDesign #ProteinDesign #AIforScience #BioAI

Post attachmentPost attachmentPost attachmentPost attachment

Spiritein @SpiriteinTech 13h Article AI Antibody Design Meets Reality—and It’s Not Pretty. The Anthropic × Adaptyv competition and the Changjiang Cup put wet-lab testing at the center of protein design—a timely reminder to ask how well AI-designed nanobodies (VHHs) perform in experiments.

BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·13h

🧬 Anthropic × BigHat brought Claude into antibody optimization to tackle a blind spot in automated design: a model can improve an experimental score without improving the property researchers actually need. Across two campaigns, researchers and Claude examined raw instrument curves, uncovered misleading thermal-stability metrics and tested revised interpretations in new experiments. Their preprint, “Reasoning in the loop,” extends the optimization loop beyond choosing better molecules to questioning—and experimentally checking—how “better” is measured. 🔬 Seven rounds, approximately 1,010 design variants. On BigHat’s Milliner platform, the team optimized an anti-albumin nanobody and an anti-mesothelin scFv, a linked antibody fragment. Claude helped investigate antibody lineage, structural evidence and experimental curves, and propose hypotheses for testing. Researchers set the strategy, reviewed designs and performed the experiments. 📉 Two campaigns exposed two measurement problems: → Nanobody: aggregation began around 40°C, well below its melting temperature. Raising the melting point did not necessarily reduce aggregation. → scFv: the reported apparent “melting temperature” reflected protein precipitation rather than a clean unfolding transition. The team fitted physical models to the curves, adopted more suitable metrics, and varied protein concentration and heating rate to test preregistered predictions. 📊 Both final leads improved over their in-campaign parents: → Nanobody: KD improved from 10.3 to 1.15 nM—about 9× tighter binding. Confirmed melting temperature rose from 62.2°C to 70.6°C, with no aggregation onset detected up to the 85°C measurement limit. → scFv: KD improved from 3.7 to 0.63 nM—about 6× tighter binding. Solution-collapse temperature rose from 56.3°C to 74.2°C, approximately +18°C. 💡 The contribution is bringing scientific reasoning about measurement into a live optimization campaign. The agent helped researchers investigate why a readout behaved unexpectedly, formulate a physical explanation and test it—alongside designing improved molecules. ⚠️ Preprint; two human-directed lead-optimization campaigns. Lead measurements had limited replication, some affinities approached the assay’s resolution limit, and not all preregistered predictions were fully confirmed. The results do not isolate Claude’s contribution or establish autonomous drug development. 🔗 Paper: arxiv.org/abs/2611.19812 (arxiv.org/abs/2611.19812?utm…) #AntibodyEngineering #AIforScience #LabAutomation #BioAI #Anthropic #Claude

Post attachmentPost attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·17h

💰 Alphabet’s Isomorphic Labs has raised $5 billion in a single round—bringing its announced funding to $7.7 billion in roughly 20 months, while its first drug candidates are still heading toward human trials. The next test: whether its AI drug-design engine can turn computational gains into medicines that work in patients. Announced on November 17, the Series C was co-led by Greenoaks and Sequoia Capital, with a16z and other new investors joining. Alphabet invested again, alongside all existing backers: Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund. The final valuation was not disclosed. 📈 Three rounds, with financing accelerating: → March 2025: $600 million in its first external funding round → May 2026: $2.1 billion Series B → November 2026: $5 billion Series C That is $7.1 billion raised in approximately six months. 🧬 What is the money scaling? Founded in 2021 and led by Demis Hassabis, Isomorphic is building IsoDDE—the Isomorphic Labs Drug Design Engine. It combines predictive models to help researchers understand molecular interactions, select candidates for synthesis and testing, and refine their properties. Its February technical disclosure described three capabilities that extend beyond predicting a protein’s structure: → Structure prediction: more than twice AlphaFold 3’s accuracy in the hardest generalisation category of Runs N’ Poses, involving unfamiliar protein pockets and ligands. → Binding affinity: predicting how strongly a molecule binds, helping researchers rank and optimise candidates. → Pocket identification: finding potential binding sites from protein sequences, including hidden pockets that are difficult to identify in an unbound structure. 🤖 The latest announcement puts proprietary AI agents alongside these models. Isomorphic says the agents reason over complex biological systems, supporting its drug-development programmes. It reports positive preclinical results in internal oncology and immunology projects and partnered programmes. The new capital will also fund translational research and clinical-development infrastructure, supporting the work needed to move candidates from the laboratory into human testing. 💡 My read: investors are backing the possibility of a reusable drug-development platform that can support many programmes. Collaborations with Lilly, Novartis and Johnson & Johnson provide real projects on which to test its practical value. ⚠️ Clinical validation remains ahead. Accurate binding predictions do not establish human safety or efficacy, and positive preclinical findings do not guarantee clinical success. IsoDDE’s published benchmarks are company-reported; this announcement does not report clinical entry. The milestones to watch now: which candidate reaches human testing first, what disease it targets, and whether the resulting data support the promise of the platform. 🔗 Official funding announcement: prnewswire.com/news-releases… (prnewswire.com/news-releases…) 🔗 IsoDDE technical overview: isomorphiclabs.com/articles/… (isomorphiclabs.com/articles/…) #AIDrugDiscovery #AIforScience #Biotech #Alphabet #Isomorphic Labs

Post attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·20h

🧬 Claude Fable 5.2 outperformed the tested open protein language models—including ProGen3-3B—at predicting the next amino acid in little-studied proteins. An external analysis points to substantial protein-sequence learning in Claude. In a November 16 analysis, Xinming Tu and Leo Tianlai Chen investigated a question behind LLM performance on protein tasks: do these models learn sequences themselves, or mainly recall descriptions from papers and webpages? 🧪 The test: unfamiliar proteins, with shuffled controls. The researchers selected 668 bacterial and archaeal proteins from 167 Pfam families of unknown function. These families had no experimentally resolved structures and at most three associated papers each. Each protein was cut at five positions, creating 3,340 prediction tasks. Models received a sequence prefix and generated ten amino acids; the primary metric scored only the first. They then repeated the test with shuffled sequences, preserving amino-acid composition while destroying sequence order. 📊 Real-sequence accuracy minus shuffled-sequence accuracy: → Claude Fable 5.2: +45.1 percentage points → ProGen3-3B: +32.5 → ProtGPT3-10B: +27.7 → Claude Opus 5.5: +27.4 Fable’s accuracy was 54.1% on real sequences versus 9.0% on shuffled controls. The tested GPT, Gemini and seven open-weight models showed little comparable separation. 🔍 Three findings strengthen the sequence-learning interpretation: → The effect persisted across 128 families with no associated papers, weakening the explanation that models simply recalled published descriptions. → Performance increased with protein-family size—a pattern also observed in dedicated protein language models. → High accuracy did not depend on reported reasoning tokens, suggesting that extended deliberation was not driving the result. The signal extended beyond one residue: Fable correctly continued all ten amino acids at 88 cut points; Opus did so at 12. Both scored zero on shuffled controls. 💡 Some general-purpose LLMs may already contain substantial protein-sequence knowledge. Whether that knowledge can support useful mutation-effect prediction, folding or functional design remains an open experimental question. ⚠️ Public analysis with open code; no peer review or wet-lab validation. This benchmark tests sequence continuation, not overall protein-design capability. Prefix-only inputs disadvantage bidirectional ESM models, and the inferred training exposure has not been confirmed by Anthropic. 🔗 Original analysis (xinmingtu.cn/blog/2026/claud…) · Code and results (github.com/XinmingTu/LLM-bio…) #AIforScience #ProteinLanguageModels #BioAI #Claude #ClaudeFable5.2

Post attachmentPost attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 17

🧬 Nature Medicine has published OpenScientist, an open-source AI co-scientist evaluated across five biomedical case studies—including hidden-data testing, repeated runs and blinded comparison with independent human analysis. The study shows both the research workflows it can execute and the errors that still require expert correction. One concrete result: a six-protein survival score achieved a c-index of 0.826 in a hidden cohort of 176 participants, including 17 deaths. The model was fixed before evaluation on that cohort, testing whether it could rank survival risk in people whose data had not been used to build the score. Here’s what makes the paper worth reading: → It executes analyses, beyond proposing ideas. Given a scientific question and data, OpenScientist can search literature, write and run code, inspect results, iterate and produce a report with analysis artifacts available for review. → The evaluation goes beyond a single successful run. Across different cases, the authors examined performance on unseen data, repeated a tau-pathology analysis across 60 runs under different configurations, and compared AI and independent human proteomic analyses under a blinded, prespecified setup. → The failures are consequential. The evaluations identified issues involving duplicate observations, cohort definitions and statistical reporting. In one case, a stricter statistical correction was calculated but omitted from the final report. Executable code alone does not guarantee a sound conclusion. → Sensitive data can stay offline. The system is open source under Apache 2.0 and supports fully air-gapped operation with local open models—a useful option for clinical research environments. The contribution is a framework for executing and inspecting biomedical analysis workflows, with concrete tests of reliability. It also makes the remaining gap clear: generating an analysis and establishing a trustworthy conclusion are separate achievements. ⚠️ The survival result comes from one case with only 17 deaths in the hidden cohort; c-index 0.826 does not mean 82.6% diagnostic accuracy. These are computational biomedical studies, not autonomous wet-lab experiments or evidence of clinical readiness. 📄 Paper: nature.com/articles/s41591-0… 💻 Code: github.com/openscientist-io/… #AIAgents #AIforScience #ComputationalBiology #OpenSource

Post attachmentPost attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 17

🧬 David Baker and Anthony Green’s teams designed enzymes from scratch to hydroxylate different sites on the same molecule: estradiol. Separate designs targeted C6 or the unactivated C16 position, with 93% and >99% diastereomeric excess, respectively. The starting point was a predefined substrate and reaction geometry. The team then built new protein scaffolds around the catalytic machinery needed to perform the reaction. In this bioRxiv preprint, researchers at the University of Washington / HHMI and the University of Manchester report: → Reaction geometry as the design input. A cysteine-ligated heme, catalytic residues for hydrogen peroxide activation, and the target C–H bond positioned beside the reactive iron–oxo intermediate defined the active site. RFdiffusion2/3 generated scaffolds, followed by sequence design and all-atom filtering. → Different sites on the same complex substrate. The estradiol designs demonstrate control over both the position of hydroxylation and the stereochemistry of the product. The reported 93%/>99% d.e. values describe stereoselectivity—not conversion or yield. → Activity improved through further engineering. In the C16 route to estriol, substitutions increased yield from approximately 1% to 37% and total turnover number from below 1 to 37, while retaining >99% d.e. In a separate aliphatic-substrate campaign, an improved variant reached 250 turnovers, versus 12 for its parent design. → Crystal structures supported the designs. A 1.9 Å structure of an estradiol hydroxylase variant matched its AlphaFold3 model with a Cα RMSD of 0.65 Å. The advance is a concrete route toward custom biocatalysts: specify the substrate and target reaction geometry, then construct a protein around them. That could expand options for selective modification of complex molecules. ⚠️ Preprint with wet-lab and crystallographic validation. The demonstrated substrates and turnover numbers do not establish an industrially ready catalyst or the ability to target any C–H bond on demand. Further optimization was needed. 📄 Paper: biorxiv.org/content/10.64898… #ProteinDesign #EnzymeEngineering #Biocatalysis #AIforScience #SyntheticBiology

Post attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 16

🧬 In Nature Communications, an mRNA vaccine encoding two computationally designed influenza hemagglutinins gave mice 100% survival in five of six lethal influenza challenges—and 60% survival against H10N8. The work introduces PanFlu-IOA, a computational antigen-design toolkit built to extract shared antigenic features across influenza A Group 1 and Group 2. 🔬 The team designed two universal HA sequences: → G1-HA for Group 1 → G2-HA for Group 2 The two mRNA-LNPs were mixed into the vaccine candidate panFluA. Two weeks after the booster, female BALB/c mice were challenged with 10 LD50 of six divergent viruses. Reported survival: → 100%: pandemic H1N1, seasonal H1N1, PR8, H9N2, and H3N2 → 60%: H10N8 → Controls showed near-complete mortality by days 7–9 T-cell depletion and passive-serum-transfer experiments supported contributions from both cellular and humoral immunity. PanFluA serum alone provided about 80% protection against PR8. Why it matters: this connects cross-group computational HA design to multi-subtype lethal-challenge protection in mice—an end-to-end design-to-challenge validation rather than a purely computational antigen study. ⚠️ This is a mouse study, not a human vaccine result. Cross-protection was not uniformly matched by classical neutralizing or hemagglutination-inhibition activity outside H1 strains, so protection cannot be attributed to those antibody readouts alone. Recent isolates from the past five years were not tested, and pre-existing human immunity was not modeled. 📄 Paper: nature.com/articles/s41467-0… 💻 Code: github.com/xielisos567/PanFl… #Influenza #mRNA #VaccineDesign #ComputationalBiology #AIforScience

Post attachmentPost attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 16

🧬 MDClaw found 11 AI-run MD workflows that looked successful—but had skipped minimization or equilibration, or simulated the wrong structure. That is the failure mode this new system is designed to address: an agent can produce a trajectory and a convincing report without actually following the intended molecular-dynamics protocol. MDClaw is not another MD foundation model. It is an auditable constraint layer for general coding agents: validating command-line tools check inputs and stage dependencies before execution, while a directed acyclic graph (DAG) records every stage in a recoverable, branchable workflow. 🔬 On MDDataBench, a 31-system benchmark scored against deposited reference simulations: → Without MDClaw: 21 passes in 372 independent runs, or 5.6% pass@1 → With MDClaw tools: 81–94% pass³ across four open-weight models → With tools plus skills: 90–100% pass³ Here, pass³ means a system passes in all three independent runs—not merely one lucky successful attempt. Why it matters: as agents begin to simulate atomic-scale scientific hypotheses, the important question is not only “did it produce a trajectory?” It is whether a scientist can inspect what was actually simulated, which inputs were used, and whether essential stages were skipped. ⚠️ This is a bioRxiv preprint. MDDataBench measures correct execution against predefined simulation tasks and reference outputs; it does not establish sampling convergence or biological correctness. The main results cover four specific open-weight models in one agent setup. Its no-reference applications are traceable simulation studies, not experimentally confirmed biological discoveries. 📄 Preprint: biorxiv.org/content/10.64898… 💻 Code: github.com/matsunagalab/mdcl… 🧪 Benchmark: github.com/matsunagalab/MDDa… #MolecularDynamics #AIAgents #ScientificAI #ComputationalBiology #Reproducibility

Post attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 16

🧬 No experimental structure. About 3 million compounds screened. One AlphaFold-guided campaign produced QL07, a CLIP1–LTK inhibitor with an enzymatic IC50 of 524.27 nM. CLIP1–LTK is an oncogenic fusion kinase without an experimentally resolved structure. In a new Acta Pharmacologica Sinica paper, researchers first benchmarked how AlphaFold2- and AlphaFold3-derived structures perform in virtual screening, then applied a hierarchical workflow to this target. 🔬 After screening approximately 3 million compounds and conducting similarity-based analog searching, they identified QL07. The study reports that QL07: → inhibited downstream CLIP1–LTK signalling → induced apoptosis in Ba/F3 cells transformed with CLIP1–LTK → suppressed growth of CLIP1–LTK-positive xenografts Why it matters: this is not only an in silico docking exercise. It is a peer-reviewed wet-lab case study showing that AlphaFold-derived structures can help start a structure-based discovery campaign when a disease target has no experimentally determined structure. ⚠️ QL07 is an early research compound, not a clinical candidate. The paper does not establish superiority over existing LTK or ALK inhibitors, and its reported xenograft activity does not demonstrate human efficacy. No public code or model-weight release was stated in the abstract. 📄 Paper: nature.com/articles/s41401-0… #AlphaFold #DrugDiscovery #VirtualScreening #CancerResearch #AIforScience

Post attachmentPost attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 16

🧬 Given only the sequence of the EBV protein BHRF1—no structural template—LiteFold’s LiteMol-1 selected 19 linear peptides for testing. Nine bound by BLI. That is a 47.4% hit rate in the selected set. Two leads showed apparent KDs of 37.6 nM and 49.9 nM, lower than the natural ligand Bim at 75.6 nM in the same assay format. 🔬 LiteMol-1 is a masked diffusion language model for target-conditioned peptide generation. The workflow generated and cofolded roughly 500–600 candidates, then used Protenix v2, Boltz-2, interface scoring, and manual selection before sending 19 peptides for capture-format BLI testing at Adaptyv Bio. Results: → 9 BLI-positive peptides → 9 BLI-negative peptides → 1 expression failure → selected peptides were 18–27 residues long Why it matters: this is a useful early wet-lab readout for a sequence-conditioned peptide-generation model—not merely a structural prediction or in silico ranking result. ⚠️ This is a company technical report, not peer-reviewed research. The 47.4% is the hit rate after computational filtering, not the raw generation success rate. The reported values are apparent KDs from tagged capture-format BLI, not solution KDs; all nine hits carried N-terminal tags. There are no selectivity, binding-site, cellular-activity, or stability data, and no head-to-head comparison with other generators. LiteMol-1 weights and code have not been announced as public. 🔗 Technical report: litefold.ai/research/bhrf1 𝕏 Official post: x.com/try_litefold/status/21… #PeptideDesign #ProteinDesign #GenerativeAI #BLI #AIforScience

Post attachmentPost attachmentPost attachmentPost attachment

We experimentally tested 19 AI-designed peptides. 9 showed binding. 2 achieved lower apparent KD than a natural ligand. Today, we're sharing the first wet-lab validation results of LiteMol-1, our molecular diffusion language model. We targeted BHRF1, an anti-apoptotic protein encoded by Epstein–Barr virus, which is implicated in certain lymphomas. LiteMol-1 generated linear peptides conditioned only on the target sequence. We then prioritized candidates using Boltz-2, Protenix along with interface scoring. We collaborated with @adaptyv independently tested the selected peptides using biolayer interferometry (BLI). The results: > 9/19 candidates showed binding (47.4%). > Peptide_010: 37.6 nM apparent KD. > Peptide_012: 49.9 nM apparent KD. > Natural Bim control: 75.6 nM apparent KD. These are early, single-target biochemical results, but an important step toward connecting generative molecular design with experimental validation. Our broader goal is to close the loop between AI-driven generation, computational evaluation, and real-world experimental feedback. Technical report link in the comments.

Post attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 15

🧬 On 25 held-out computational biology tasks, the top AI agent scores just 40.3/100. FrontierBioML, released by Valen Labs, evaluates agents across antibody engineering, protein engineering, structure, binder design, and functional genomics. Agents write prediction scripts in a GPU-free sandbox with roughly six hours of runtime; submissions are then scored on held-out data. 🏆 Current leaderboard: → Claude Mythos 5.2 (Claude Code): 40.3 → Gemini 4 Carbon: 39.0 → Claude Opus 5.5: 36.5 → Baseline: 13.1 The important context: scores are normalized per task. The benchmark authors say 0 usually corresponds to a simple baseline, 100 to a perfect answer, and a score of 30 is often already strong. Scores therefore should not be read as a conventional “percentage correct.” No system solves all 25 tasks. The result is less “Mythos wins again” than a useful reality check: even leading agent setups remain far from reliable general-purpose computational biologists. ⚠️ This is a company-run benchmark and compares full model-plus-scaffolding systems, not base models in isolation. Results are not peer reviewed, and task scores are not directly comparable across tasks. Two GPT systems were also affected by biosecurity filtering: zero-scored blocked runs lower their overall averages, while their scored-only runs were reportedly closer to the leading group. 🔗 Benchmark: frontierbioml.com/ 📄 Details: valenlabs.com/blog/frontierb… #AIforScience #ComputationalBiology #ProteinEngineering #AIAgents #Bioinformatics

Post attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 15

A 35-million-parameter model has moved ahead of ESM2-3B on protein stability prediction. On 66 ProteinGym stability assays, DynamicsPLM reached a mean Spearman ρ of 0.528 — ahead of ESM2-3B at 0.509, and second only to the 3-billion-parameter VespaG at 0.533. The model, from the University of South Florida and George Mason University, adds molecular-dynamics heads to an ESM2-35M encoder and scores mutations zero-shot from sequence alone. Across 201 ProteinGym substitution sets, it posted a mean ρ of 0.398: 19th of 47 sequence models, and first among models with 100 million parameters or fewer. Why it matters: for stability engineering, a smaller model with explicit dynamics supervision can compete with sequence models many times its size. Bigger is not automatically better. ⚠️ This is a bioRxiv preprint, not a peer-reviewed paper, and it reports no new wet-lab experiments. The authors found no statistically significant advantage on the 12 Binding assays. Eleven of 201 wild-type sequences (5.5%) are identical to training sequences, and the paper does not report a full recalculation after removing them. 📄 biorxiv.org/content/10.64898… 💻 github.com/kabir-ai2bio-lab/… 📦 Weights: doi.org/10.5281/zenodo.23602… #ProteinLanguageModels #ProteinEngineering #ProteinGym #AIforScience

Post attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 15

⚡ LiteFold has released Jade, a GPU-accelerated homology search engine that builds co-folding-ready alignments for proteins, DNA, and RNA without a prebuilt index. In sensitive mode on the SCOP protein benchmark, Jade reportedly cut runtime from 506 seconds on 64 CPU threads to 45 seconds on a single A100 — about 11× faster, at the same ROC1 sensitivity. On a simulated bacterial-genome test, default mode reached 99.4% recall in 19.3 seconds, versus BLASTN at 99.8% in 501.5 seconds. Why it matters: cheap, repeatable MSA search means a design loop need not wait on a prebuilt database. ⚠️ This is a company research page, not peer review. The DNA test is simulated. Jade is already used on the LiteFold platform; an API and SDK are described as coming soon. 🔗 litefold.ai/research/jade #HomologySearch #ProteinDesign #MSA #AIforScience

Post attachmentPost attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 14

🧬 University of Michigan researchers introduced Perturb-Tags, a CRISPR-screening method that measures how a perturbation changes both the edited cell and its unedited neighbours—inside intact tissue. From the same spatially barcoded nucleus, it reads the CRISPR guide, gene expression, chromatin accessibility, and spatial position at single-cell resolution. 🔬 The team applied Perturb-Tags in human embryoid bodies, systematically perturbing ligands, receptors, and all expressed kinases. They report that ligand knockdowns produced distance-dependent loss of signalling, compensatory activation of alternative pathways, and changes in neighbouring cells’ gene expression. Receptor and kinase knockdowns also produced responses that depended on the local microenvironment and altered spatial cell-state organization. 🧠 A graph-attention network then jointly modeled the spatial effects of perturbations on RNA and chromatin accessibility, including counterfactual guide substitutions in held-out embryoid bodies. Why it matters: conventional Perturb-seq is powerful for cell-autonomous effects, but dissociating tissue removes the neighbourhood where many ligand–receptor effects actually happen. Perturb-Tags brings perturbation biology closer to an in situ, multiomic “virtual tissue” setting. ⚠️ This is a bioRxiv preprint. The reported biological effects are from embryoid bodies, not patient tissue or clinical models. The authors claim higher sensitivity than previous spatial perturbation methods, but the full numerical comparison still needs careful review. 💻 Code: github.com/welch-lab/perturb… 📄 Preprint: bioRxiv, posted Nov 13, 2026 #SpatialOmics #CRISPR #PerturbSeq #SingleCell #AIforScience

Post attachment
BioAI Frontier avatar
BioAI Frontier@BioAIFrontier·Nov 14

🧬 Deep Genomics has published DeepRNAi, a model designed to predict both siRNA potency and transcriptome-wide off-target knockdown before experiments. In an author-reported genome-wide scan of 24.4 million siRNAs across 7,946 genes, 90.1% of the top 100 most potent candidates carried at least one predicted off-target liability. 🔬 DeepRNAi predicts RISC-mediated knockdown for any siRNA–transcript pair—not just fully complementary on-targets, but also near-perfect and seed-mediated matches. It uses transcript-context representations from biological foundation models rather than relying only on sequence-matching or thermodynamic heuristics. The preprint was trained on 2,031 siRNAs and 16 million siRNA–transcript pairs. The authors say they will release a 1.84-million-pair training and benchmark dataset spanning 240 siRNAs. Why it matters: siRNA discovery is not only about finding the strongest silencer. It is about finding one that silences the intended transcript without unintentionally silencing many others. ⚠️ This is a company-authored bioRxiv preprint. The definition of an “off-target liability” behind the 90.1% result needs full-method review; no public code repository was identified. The model’s prior company announcement is not independent validation or clinical evidence. 📄 Preprint: biorxiv.org/content/10.64898… 🔗 Background: deepgenomics.com/blog/introd… #RNAi #siRNA #RNAtherapeutics #AIforScience

Post attachment