BioAI Frontier (@BioAIFrontier) profile picture

Making sense of the frontier where AI meets biology.

Joined November 2026
Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)Photo by BioAI Frontier (@BioAIFrontier)
🦠 Nature Communications introduces VHSeek: a tool that predicts potential viral hosts without candidate-host genomes. Using viral nucleotide and protein sequences, it expanded host-annotation coverage from 36.5% to 64.5% in a processed IMG/VR collection. Researchers at USC and MBZUAI developed the framework to address a practical bottleneck: sequencing reveals viruses faster than researchers can identify their hosts. 🧬 Two complementary views of each virus. AIDO.DNA extracts nucleotide features, while ESM2 extracts protein features. A supervised predictor combines them and learns from known virus–host pairs. Ablation experiments found that combining both representations outperformed using either alone. Predictions are organized along an existing host taxonomy, supporting species, genus and broader ranks. Inference uses viral sequences and known host taxonomy, without searching candidate-host sequence databases. 📊 Better predictions and broader candidate coverage. On the retained iPHoP benchmark, genus-level Top-1 accuracy reached 82.8%, versus 69.5% for iPHoP. Across nearly 487,000 processed IMG/VR records, VHSeek assigned high-confidence candidate hosts to 136,173 previously unannotated records. A validation-calibrated threshold allows uncertain predictions to remain unanswered. The practical value is prioritization: narrowing which virus–host associations researchers should investigate when host genomes are unavailable. ⚠️ This peer-reviewed computational study includes no new infection experiments. Predictions remain within a predefined host-taxonomy label space. Related viruses can occur across training and test sets; VHSeek did not outperform iPHoP on the most evolutionarily distant subset. Expanded annotation coverage does not establish experimentally confirmed hosts or cross-species transmission. 📄 Paper: nature.com/articles/s41467-0… (nature.com/articles/s41467-0…) 💻 Code listed in the paper: github.com/maovshao/VHSeek
8
🧪 Tsinghua researchers compared 13 released drug–target prediction systems: for the same target, 77.8% of pairwise Top-10 drug-list comparisons shared no candidates. Separately, retraining one architecture with different random seeds produced similar test scores—but only about two shared candidates per Top-10 list. The preprint asks a practical question: when experiments can test only a few candidates, what do benchmark scores tell us about which ones to choose? 🔬 Same candidates, different priorities. The systems ranked a common catalogue of 711 molecular structures representing FDA-approved drugs and 614 human proteins. In reverse screening—finding targets for a drug—70.1% of pairwise Top-10 comparisons were also completely disjoint. 🎯 Finding activity is not the same as identifying the right pairing. Using published measurements for 79 compounds against 65 kinases, the authors reassigned recommendation lists to different queries. Some systems retained almost the same hit yield, suggesting that broadly favoured candidates accounted for much of their enrichment. Others lost substantially more hits, showing stronger correspondence to the intended query. 🔁 Similar scores did not guarantee reproducible recommendations. In the controlled retraining experiment, the data, architecture, encoders and objective stayed fixed. Changing the seed substantially changed candidate identities, without identifying a seed with clearly superior mean hit yield on the measured overlap panel. The contribution is an evaluation framework connecting scores to experimental choices: activity yield, query-specific gain, selectivity and recommendation stability. ⚠️ Preprint; retrospective computational evaluation, not new wet-lab validation. Findings depend on the tested configurations and candidate pools. Disagreement alone does not establish which recommendations are correct. 📄 Paper: arxiv.org/abs/2611.26369 (arxiv.org/abs/2611.26369?utm…) #DrugDiscovery #VirtualScreening #AIforScience
1
2
9
324
🧬 Science Advances: GenNA generates DNA and RNA conditioned on natural-language annotations—connecting descriptions of biological function with nucleotide sequence generation. The model was pretrained on a joint sequence–text corpus spanning 2,221 eukaryotic species and approximately 416 billion characters. Its contribution is not simply scale: functional descriptions, species information and gene metadata become inputs that guide generation. 🧩 How it works. Developed by a team led by Zhejiang University’s Ming Chen, GenNA jointly models annotations, nucleotide sequences and structural tags in an autoregressive Transformer. The approximately 3.6B-parameter model was trained from scratch; a 0.36B version is also provided. 🔬 What was tested? The preprint describes targeted generation of tRNAs, ribosomal RNAs and histone-coding sequences. Evaluations examined RNA structural compatibility and whether translated histones resembled natural families in physicochemical properties. Swapping correct functional annotations for mismatched ones increased model perplexity, supporting sensitivity to the relationship between text and sequence. 🎯 The advance is a more direct interface for controllable nucleotide generation: biological descriptions can help specify the candidates a model produces. ⚠️ Computational evidence, not experimentally demonstrated function. The preprint reports no systematic wet-lab validation, and generated sequences underrepresented some natural diversity. This is not evidence that arbitrary instructions reliably produce functional DNA or RNA. 📄 Science Advances: science.org/doi/10.1126/scia… 📑 Preprint: biorxiv.org/content/10.64898… (biorxiv.org/content/10.64898…) 💻 Code: github.com/DrBlackZJU/GenNA (github.com/DrBlackZJU/GenNA?…) 🤗 Weights: huggingface.co/DrBlack/GenNA (huggingface.co/DrBlack/GenNA…) #GenerativeBiology #RNADesign #AIforScience
4
8
314
🧬 An LLM-guided antibody-design pipeline produced 38 experimentally detected binders from 80 synthesized designs. Aureka AI Research’s OpenDDE-Harness keeps the LLM’s weights fixed, using repeated sequence edits and structural feedback to guide the search. 🔁 The framework edits antibody CDRs—the regions that largely determine target recognition—within predefined constraints. OpenDDE evaluates predicted complexes, while the execution layer checks proposals and tracks candidate lineages. Accepted, rejected and failed designs remain available to inform later cycles. Adaptation happens through accumulated feedback, without retraining the LLM. 🔬 The authors report binding across three targets and four target–format cohorts: • RAGE: 17/20 • TfR1: 11/20 • CXCR4 monoclonal antibodies: 4/20 • CXCR4 VHHs: 6/20 Binding was measured using Carterra SPR or BLI. The overall 47.5% hit rate includes synthesized candidates that yielded insufficient material for characterization. 🧪 Separately, across six computational targets, mean structural-filter success was 25.0%, versus 12.6% for Germinal. These are computational pass rates, not experimental binding rates. The practical contribution is a traceable, iterative workflow connecting LLM proposals, molecular evaluation and experimentally tested candidates. ⚠️ This is a company technical report, without peer review. The pipeline also used SolubleMPNN refinement and internal developability filters. The experiments do not establish superiority over competing methods or isolate the contribution of LLM memory. 📄 Technical report: github.com/aurekaresearch/Op… (github.com/aurekaresearch/Op…) 💻 Code: github.com/aurekaresearch/Op… (github.com/aurekaresearch/Op…) #AntibodyDesign #ProteinDesign #AIforScience
2
9
35
1,206
✨ OpenAI has launched the medicinal-chemistry benchmark MedChemBench: GPT-6.2 Astra earned 84.5% of available credit at xhigh. It tests whether AI can combine molecular structures and experimental evidence to explain findings, select compounds and predict properties—but this score does not establish drug-discovery ability or improved laboratory decisions. 🔬 The benchmark pauses research at a decision point. Its 99 problems draw on 117 papers and patents, covering 761 molecular structures. Models receive project goals, compounds and selected findings; later results are withheld for grading. 🧩 A concrete improvement: preserving molecular identity. On one reconstruction problem requiring an 18-atom ring, GPT-6 Astra answered correctly in all ten responses. GPT-5.6 Sol and Luna failed all ten; illustrated outputs shrank the ring or disconnected the molecule. These errors change the compound being analysed. openai.com 📊 The earlier GPT-6 Astra’s xhigh component scores nevertheless reveal uneven capability: → Molecular reconstruction: 95.37%. → Scientific rationale: 91.05%. → Ranking and selection: 50.42%. → Numerical prediction: 47.02%. These are not GPT-6.2’s component results. They show progress in representing chemistry alongside weaker candidate selection and forecasting. Structure and rationale average roughly 73% of available points. Partial-credit rules mean “half the credit” is not “half the decisions correct.” ⚠️ Retrospective, single-turn evaluation without tools, feedback or new experiments. Withholding later findings from prompts cannot rule out training exposure. Prospective validation is still needed. cdn.openai.com 🔗 Official announcement: openai.com/index/introducing… 📄 Technical report: cdn.openai.com/pdf/09c4ec60-… #DrugDiscovery #MedicinalChemistry #AIforScience #OpenAI
MedChemBench is a new benchmark for AI reasoning in medicinal chemistry. It shows the progress our frontier models have made in their ability to help scientists use AI to identify promising drug candidates. openai.com/index/introducing…
1
4
26
1,276
🧬 How much evolutionary information does AlphaFold 3 need? AlphaInterp, published in PLOS Computational Biology, finds that just 5–10 appropriately divergent homologous sequences can largely recover full-alignment prediction performance. At the same depth, near-identical homologues provide less benefit. The authors also traced and intervened on AF3’s internal representations to investigate how evolutionary differences become structural judgments. The key is not simply how many sequences enter the model, but what information they provide. 🔬 AlphaInterp is an interpretability study, not a new structure predictor. Georgia Tech researchers Jonathan Feldman and Jeffrey Skolnick examined AF3 at four computational checkpoints, combining representation analysis, linear probes, internal interventions and controlled MSA perturbations. 🧬 Diversity mattered beyond sequence count. In a 183-protein analysis, ten homologues with 30–60% sequence identity produced a mean TM-score of 0.92 against full-MSA predictions—not experimental structures. Unrelated sequences failed to rescue prediction despite preserving the alignment-shaped input. 🧩 The study looked inside the computation. Structural properties became increasingly decodable during Pairformer processing. Intervening on late pair representations changed uncertainty in predicted residue-pair distances; this did not establish improved structural accuracy. 🎯 Why it matters: MSA composition deserves attention alongside depth. For protein design, confident structural predictions still require independent checks of physical foldability. ⚠️ Peer-reviewed computational research, primarily on monomers. Five to ten homologues is not a universal sufficiency threshold, and the proposed role of learned structural priors remains a mechanistic interpretation. journals.plos.org (journals.plos.org/ploscompbi…) 📄 Paper: journals.plos.org/ploscompbi… (journals.plos.org/ploscompbi…) #AlphaFold3 #ProteinStructure #MechanisticInterpretability
3
16
296
🧬 MIT × Sanofi: 15 antibody–antigen structure predictions pooled across five models beat 100 predictions from any single model at selecting an accurate binding pose. Their selector, AbPoolRank, reached 58.0% Success@1 versus 51.5% for the strongest single-model deep-sampling setup—without using the predictors’ native confidence scores. 🔬 Across 103 post-2021, cluster-disjoint test complexes, the team pooled three structures each from Boltz-2, Promera, Protenix-v2, ESMFold2-Fast and OpenDDE. Success meant selecting a structure with DockQ ≥ 0.49: medium-or-better docking quality, not experimental binding. 🧩 AbPoolRank ranks candidates using structural agreement, epitope overlap and source-balanced support. It knows which structures share a source, but not the predictor’s name. The main gain came from complementary candidates: different models explored useful binding poses that deeper sampling of one model missed. Effective ranking then converted that broader coverage into better selections. 🎯 The practical takeaway: allocate inference budget across complementary predictors, alongside reliable selection—not just more runs of one model. ⚠️ Preprint; computational benchmark only. Results are specific to antibody–antigen complexes. Adding another predictor did not automatically improve accuracy. The listed code repository currently returns 404. 📄 Paper: biorxiv.org/content/10.64898… #AntibodyDesign #ProteinStructure #AIforScience
5
23
1,096
🧬 Nature Biotechnology: RNA design agent RhoAgent used experimental feedback to iteratively improve hovlinc ribozyme cleavage efficiency, reaching up to 6× that of a human-expert-truncated construct. The ribozyme’s structure had not been experimentally characterized, making this a test of functional optimization with limited structural information. Developed by CUHK’s Yu Li group and collaborators including James J. Collins and Chun Kit Kwok, RhoAgent combines RNA generation tools with an agent-driven design workflow. 🧩 How does it work? RDiff, its diffusion-based generative tool, produces candidates conditioned on base interactions and functional motifs. RhoAgent organizes candidate design and refinement, using experimental feedback to inform subsequent iterations. The goal is to preserve or improve RNA function while exploring new sequences and structural arrangements. 🔬 The study reports several experimental demonstrations: → Diverse, active ribozymes with only 40–70% sequence similarity to their wild-type counterparts. → Hovlinc cleavage efficiency reaching up to sixfold that of the expert-truncated comparator. → Fluorogenic aptamer scaffolds and aptamer-based FRET scaffolds that spatially organize two functional cores while preserving both components’ binding activities. The contribution is the combination: constrained RNA generation, experimental evaluation and feedback-driven refinement, extending to scaffolds containing multiple functional components. ⚠️ The sixfold result is specific to hovlinc and its expert-truncated comparator—not a general improvement over wild-type RNA. This summary draws on the official abstract and public repository documentation; the full paper’s experimental details and component-level comparisons have not been reviewed. RhoAgent and RDiff code are publicly available. 📄 Paper: nature.com/articles/s41587-0… (nature.com/articles/s41587-0…) 💻 RhoAgent: github.com/ml4bio/RhoAgent (github.com/ml4bio/RhoAgent?u…) 💻 RDiff: github.com/ml4bio/RDiff (github.com/ml4bio/RDiff?utm_…) #RNADesign #SyntheticBiology #AIforScience #AIAgents #Nature Biotechnology
2
7
53
2,420
🧪 Science: a de novo bi-zinc enzyme designed without a natural template or pre-specified catalytic motif achieves a reported rate enhancement of up to 3 × 10¹³ relative to the uncatalyzed reaction—in the range of characterized natural phosphatases. Stanford, EPFL and collaborators developed dEVA, a multi-objective protein-design method, and experimentally characterized an enzyme that hydrolyzes both phosphomonoesters and phosphodiesters. 🧩 The innovation is how the catalytic site is designed. Many enzyme-design approaches begin with a predefined catalytic motif and build a protein around it. dEVA instead searches for sequences and metal environments that work together, balancing sequence–structure compatibility with predictions of a catalytically suitable zinc environment. Its evolutionary algorithm iteratively proposes changes and selects candidates representing favorable trade-offs between these objectives. 🔬 Binding a metal is not enough. The surrounding residues must position the metal and create an environment that supports chemistry. Here, the characterized design delivered hydrolytic activity, with X-ray crystallography and absorption spectroscopy supporting its de novo active site. The 3 × 10¹³ figure is defined as (kcat/KM)/kw, relative to the uncatalyzed reaction. It is not a speed advantage over natural enzymes. The contribution is a design route that reduces reliance on pre-specified active-site motifs while producing experimentally demonstrated catalysis. ⚠️ These results concern one characterized bi-zinc enzyme and the tested substrates, not a general demonstration that de novo enzymes match natural catalysts. This post draws on the Science abstract, earlier preprint and public code; the final Science full text was not accessible. Earlier versions were already public—the news is formal publication. 📄 Science: doi.org/10.1126/science.aej3… 📑 Earlier preprint: biorxiv.org/content/10.64898… (biorxiv.org/content/10.64898…) 💻 Code: github.com/gelnesr/dEVA (github.com/gelnesr/dEVA?utm_…) #ProteinDesign #EnzymeDesign #Biocatalysis #AIforScience #Science
3
16
719
🧬 In Arc Institute’s 2026 Virtual Cell Challenge, the winning predictions covered 44% of the distance between the cell context mean and an independent replicate. The organizers note that this number does not represent accuracy and that individual metrics vary widely around this average. The official wrap-up, dated November 19, 2026, reports that CellularAI Algorithm from Illumina scored 0.440 on three held-out lines. The official results table lists 0.4399, 0.3348 and 0.3263 for first through third. On this scale, 0 is the cell context mean — the average response observed across knockdowns — and 1 is an independent replicate. The overall score is the unweighted average of six metrics and three cell lines. The winners are BioInsight AI Lab from Illumina (first, $100,000), FlowState (second, $50,000) and X-team from Xaira Therapeutics (third, $25,000). Each prize is paid half in cash and half in NVIDIA Brev credits. The wrap-up lists FlowState as Independent Researchers, trained exclusively on public data, and names X-team’s model X-Cell Transfer. Zero-shot here means the Challenge provided no perturbation data from the six evaluation lines. It does not mean the models had never seen the perturbed genes. All but one of the 300 test targets already had public perturbation data in some other cell line. The final ranking used AC16 (cardiomyocyte), MOLM13 (acute myeloid leukemia) and U-87MG (glioblastoma). Validation used Cal-33, HeLa and Jurkat; the two scores are not directly comparable. All three winning teams built on reference transfer of measured responses from other cell types. They differed mainly in how they weighed the evidence, adapted it to a new line using only its control cells, and converted the predicted effects into cells: BioInsight AI Lab used a random-effects meta-analysis with weights for context similarity, noise and knockdown efficiency; FlowState learned shared response programs from a broad public corpus; X-team used a large pretrained foundation model to decide which measured effects should carry into a new context and how strongly. The organizers write that the pattern suggests that, for now, reference data and how carefully it is transferred to a new context matter at least as much as model architecture. On the validation lines, they found that the top ten entries recovered roughly what a cross-line average would predict, while the line-specific part of the response remains largely missing. ⚠️ Open weights and code were not required this year; the method descriptions are based on information provided by each team. These are computational scores on the six Challenge cell lines, not a test in primary human tissue or the clinic. 📄 Official wrap-up: arcinstitute.org/news/virtua… 🔗 Challenge site: virtualcellchallenge.org/ 📐 Scoring: virtualcellchallenge.org/eva… #VirtualCellChallenge #GenePerturbation #VirtualCell #AIforScience
3
1
6
574
🧬 AlphaGenome and Borzoi improved after adaptation to personal genomes—but still trailed a conventional elastic-net model in this study of individual-level blood gene expression. In 5,120 Hispanic/Latino participants, the authors reported mean per-gene Spearman correlations of 0.33 for adapted Borzoi, 0.29 for adapted AlphaGenome and 0.46 for elastic net. Importantly, elastic net produced no predictions for 15 of the 150 evaluated genes, so these averages do not cover identical gene sets. 🔬 The question was specific: can a model predict how expression of the same gene differs between people? Researchers from Harvard T.H. Chan School of Public Health, Beth Israel Deaconess Medical Center and collaborators paired personal genomes with whole-blood RNA-seq from HCHS/SOL. They selected 150 genes spanning low, medium and high cis-heritability—how much expression variation is attributable to nearby genetic variants—and reserved 768 participants for testing. 🧩 Adaptation helped substantially. → Borzoi: 0.11 pretrained → 0.33 after adaptation. → AlphaGenome: 0.15 → 0.29. Borzoi improved over its pretrained version for 129 genes, but exceeded elastic net for only 21. However, this was not a comparison under identical input conditions. Elastic net used variants within ±1 Mb of each gene’s transcription start site; adapted sequence models used 49,152-bp inputs. Borzoi used LoRA adaptation. AlphaGenome’s final configuration froze its backbone and trained only a regression head, after LoRA provided little additional benefit. 🌍 Ancestry analyses found broadly similar pretrained performance across European-, African- and Amerindian-enriched groups. Adapted models and elastic net generally performed slightly better when training and test ancestry groups matched. The takeaway: learning regulatory patterns from reference sequences does not automatically translate into accurate prediction of expression differences between individuals. ⚠️ Preprint; limited to bulk whole blood and 150 selected genes. This evaluates AlphaGenome, not AlphaGenome 2, and does not establish performance across all genomic tasks. The authors also considered only single-nucleotide variants when reconstructing personal sequences. 📄 Paper: doi.org/10.64898/2026.11.17.… #AlphaGenome #Genomics #GeneExpression #AIforScience
2
7
306
⚡ Rosetta physical scoring on GPUs—with up to 362× higher throughput for batched scoring and gradient calculations. tmol makes these calculations available as PyTorch components that AI protein-design workflows can call. tmol is a software library for evaluating and refining molecular structures. It implements the RosettaGenFF energy function alongside side-chain packing, energy minimization and FastRelax, helping developers incorporate physical scoring into generation and optimization. 🧩 Why does that matter? A generated structure may contain unfavorable atomic contacts or poorly satisfied interactions. tmol lets developers evaluate these features across batches of candidates. Its coordinate gradients can also indicate how moving atoms changes the energy. 🔬 The speedup has specific conditions. On an H200 NVL, batched scoring plus gradients ran 233–362× faster than single-threaded PyRosetta; torsion-space relax ran 38–102× faster. These are component-level throughput gains, not whole-workflow speedups. Across 1,000 high-resolution protein chains, total scores correlated with PyRosetta at Pearson r = 0.997. 🧪 The authors also tested physical scoring during generation. Across 24 protein–protein targets, tmol-guided RFD4-Proteina search increased the share of designs with at least five hydrogen-bonded interface residue pairs from 16% to 66%. Selecting the best 32 of 128 unguided samples already reached 53%. OpenFold3 refolding success improved from 14.5% to 22.9%, below that selection control’s 25.5%. The practical advance is making Rosetta-style energetics easier to use during AI design, with demonstrated control over specific interface features. ⚠️ Preprint; no wet-lab validation. The generation demonstration used scoring without gradients and produced larger interfaces with slightly more clashes. Benchmarks used tmol 0.2.0; v1.0.0 is available. Code is open under BSD-3-Clause. 📄 Paper: arxiv.org/abs/2611.21857 💻 Code: github.com/RosettaCommons/tm… (github.com/RosettaCommons/tm…) 📚 Documentation: rosettacommons.github.io/tmo… #ProteinDesign #Rosetta #GenerativeAI #AIforScience
5
14
934
🧬 A 36-amino-acid AI-designed protein disrupted a motor–regulator interaction and reorganized mitochondria inside cultured cells. Sidetrak, designed by University of Bristol researchers using BoltzGen, targets a regulatory surface on kinesin-1—the molecular motor that transports cargo along microtubules. The advance is a functional design application: a compact binder can interfere with a specific regulatory interaction and produce a measurable cellular phenotype. 🎯 Targeting the regulatory stalk. Sidetrak binds the TPR docking site, or TDS, on kinesin-1’s coiled-coil stalk. This surface participates in motor regulation and overlaps the binding site for MAP7, which promotes kinesin-1 recruitment to microtubules. Starting from an AlphaFold3 model, the team generated 10,000 designs. Six passed the design filters; cellular screening identified Sidetrak as the strongest candidate for further characterization. 🔬 Binding and competition were tested experimentally. → Fluorescence polarization measured an apparent KD of 110 ± 13 nM against a kinesin-1 construct containing the target site. → Disrupting either the target site or Sidetrak’s predicted interface disrupted binding. → In cell-lysate competition experiments, Sidetrak reduced MAP7-associated kinesin-1 recovery by up to approximately 85%, with half-maximal displacement at 86 nM for the heavy-chain readout. 🧫 The effect extended into cells. Cells expressing GFP-Sidetrak showed reduced MAP7-dependent motor recruitment to microtubules and a pronounced redistribution of mitochondria toward the nucleus. The binding-defective control did not reproduce the mitochondrial phenotype. The main mitochondrial analysis covered 60 cells across three independent experiments. Similar redistribution occurred in two additional cell lines, while gross microtubule architecture remained unaffected. Quantitative proteomics also supported selective engagement of endogenous kinesin-1 complexes, although off-target associations could not be excluded. 💡 Sidetrak demonstrates how designed proteins can probe regulatory interfaces and connect molecular binding to cellular function. Its interface-dependent binding also provides experimental support for the local structural model used for design. ⚠️ November 18 preprint, not peer-reviewed. Sidetrak modulates a regulatory mechanism; it is not established as a general kinesin-1 inhibitor. Cellular experiments used intracellularly expressed GFP fusions—not externally administered protein—and no therapeutic benefit was demonstrated. 🔗 Paper: biorxiv.org/content/10.64898… #ProteinDesign #Miniproteins #CellBiology #AIforScience
5
21
1,146
🧬 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
1
11
632
🧬 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
2
15
955
🧬 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
1
3
17
829
💰 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
3
7
491
🧬 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
1
8
26
1,210
🧬 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
4
12
615