Excited to share our latest paper: Prot2Prop! A parameter-efficient multitask protein language model that jointly predicts six developability properties from sequence using lightweight adapters. 📄 doi.org/10.64898/2026.06.28.… 💻 github.com/NeurosnapInc/Prot…

BoltzProt-1: Towards Efficient De Novo Binder Design with Good Developability 1 BoltzProt-1 is a de novo protein binder (incl. nanobody/VHH) design pipeline that explicitly targets two bottlenecks at once: improving prospective binding hit rates on novel targets and ensuring therapeutic-style developability (stability, solubility, low nonspecific binding, etc.). 2 The key technical shift is ranking designs with a dedicated protein–protein interaction predictor (BoltzPPI), rather than relying mainly on structure-prediction confidence proxies. In a budget-matched test on the same candidate pools, this changes selection—not generation—and directly tests whether better ranking alone improves experimental recovery. 3 On 10 low-homology (novel) targets, replacing BoltzGen’s selection with BoltzPPI ranking increases confirmed-binder rate from 3.3% (5/150) to 8.0% (12/150), a 2.4x gain. Screening hits also rise (revised hit rate 4.7% to 9.3%). Confirmed target coverage improves from 2/10 to 3/10 via ranking, and to 4/10 when paired with an improved generative model. 4 The paper also argues for stricter experimental reporting: it separates “screening hits” (including ambiguous sensorgrams) from “confirmed binders” (clean kinetics plus orthogonal confirmation). Confirmation uses a flipped assay orientation to reduce format-specific artifacts and avidity effects, with additional independent testing for some hits. 5 BoltzPPI is built on Boltz-2 representations and adds a PPI prediction head trained jointly with a confidence head. It uses interface-focused signals: token/pair features, predicted coordinates, distance embeddings, and binder/target masks, refined by a 4-block Pairformer stack (16 heads, dropout 0.25). 6 Training uses PDB and patent-derived complexes as positives, plus synthetically generated protein pairs as negatives. A multi-view training scheme drops trunk pairwise representations 50% of the time to encourage geometric reliance and reduce overfitting to internal trunk signals; additional regularization injects Gaussian noise into representations. The interaction head is trained with focal loss and combined with confidence losses. 7 On an external 10-target panel used by Chai-2, BoltzProt-1 reports screening hits on 7/10 targets, compared with 6/10 reported by Chai-2 and 3/10 for BoltzGen in this study’s setup. This suggests improved target coverage across diverse classes (signaling/adaptor proteins, cytokines/hormones, SUMOylation enzymes, calcium-binding regulators). 8 Developability is treated as a first-class outcome. Confirmed binders from the low-homology panel are evaluated across a broad assay suite (Twist Bioscience): thermal stability (Tonset, Tm1, Tm2), aggregation onset (Tagg), monomer purity (aSEC), heterogeneity (DLS PDI), hydrophobicity (HIC), polyspecificity (BVP ELISA), and self-association (AC-SINS). 9 Under combined developability criteria, 58% (7/12) of BoltzProt-1 confirmed binders pass every filter, exceeding BoltzGen confirmed binders (40%, 2/5) and clinical-stage controls measured in parallel (IgG 25%, 3/12; VHH-Fc 21%, 5/24). Attrition is minimal until hydrophobicity (HIC), which is the dominant failure point. 10 Novelty checks indicate recovered designs are not near-duplicates of known antibody/nanobody CDRs: every recovered design has minimum CDR3 edit distance ≥4 to its closest SAbDab match (with larger distances when considering CDR1 2 3). Structural context is provided via binding-site similarity to known PDB interfaces (FoldDiSCO), highlighting that the low-homology benchmark emphasizes limited prior structural precedent. 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #ProteinDesign #DeNovoDesign #Nanobody #AntibodyEngineering #ProteinProteinInteraction #MachineLearning #DrugDiscovery #Developability #StructuralBiology

🧪 Prot2Prop: one model, six protein properties at once (solubility, thermal stability, aggregation, expression yield...). Frozen ProstT5 encoder shared task adapters. AUROC 0.86–0.98, Spearman 0.73–0.86. Multitask > single-task for developability prediction. biorxiv.org/content/10.64898…

Prot2Prop sounds like a fascinating development in protein research, potentially revolutionizing how we approach protein developability. How has the incorporation of structure-informed multitask models changed the landscape of protein property predictions so far? Also, how does this advance compare to existing single-task models? Curious to hear more about its practical applications. For those interested in diving deep into the intersection of biology and AI, check out Sci-Quest at sciqst.com. It's an excellent platform for generating detailed biomedical reviews and much more. #TechInScience #ProteinResearch

Prot2Prop: Structure-informed multitask protein property prediction 1 Prot2Prop presents a single multitask model that jointly predicts six protein developability properties (material production, solubility, temperature stability, aggregation propensity, expression yield, folding stability), aiming to replace fragmented one-model-per-property workflows. 2 The core design is parameter-efficient multitask adaptation: a frozen ProstT5 encoder is augmented with a shared residual adapter plus task-specific residual adapters, enabling shared transfer while still letting each endpoint specialize at the residue (token) level. 3 A key architectural choice is where specialization happens: task-specific adapters are applied before pooling, then each task uses attention-based sequence pooling to learn which residues matter most for that property (rather than mean pooling). 4 The framework supports heterogeneous outputs in one model: three binary classification heads (material production, solubility, temperature stability) and three regression heads (aggregation propensity, expression yield, folding stability), trained with masked losses so partially labeled proteins still contribute signal. 5 On held-out test data (Seed 1 reference checkpoint), classification performance reached AUROC 0.866 (material production), 0.870 (solubility), and 0.984 (temperature stability), with corresponding F1 scores 0.853, 0.749, and 0.931. 6 Regression performance on the same test set showed strong rank fidelity (Spearman): 0.861 (aggregation propensity), 0.732 (expression yield), 0.838 (folding stability). The paper notes that some endpoints preserve ranking well even when absolute scaling is imperfect. 7 Post-hoc calibration is treated as a first-class step: per-task affine calibration for regression reduced folding stability MAE from 0.677 to 0.486 (RMSE 0.879 to 0.641) without changing Spearman, suggesting a sizable fraction of remaining error can be calibration-related. 8 The authors document an iterative development trajectory and identify the biggest win: introducing the shared adapter plus task-specific residual adapters before pooling (version 2026-04-29) produced the strongest regression gains while keeping classification stable; more complex options (ranking-aware losses, learned uncertainty weighting, ensembles) were mixed or not clearly better. 9 Practical efficiency is emphasized: only ~3.19M parameters are trainable, while ProstT5 stays frozen. In an inference benchmark (100 sequences, length ≤72) Prot2Prop ran faster than TemStaPro and SaProt (8.95s vs 15.56s vs 18.97s) on the reported hardware setup. 💻Code: github.com/NeurosnapInc/Prot… 📜Paper: biorxiv.org/content/10.64898… #ProteinEngineering #ProteinLanguageModels #MultitaskLearning #Developability #DeepLearning #ComputationalBiology #Bioinformatics

BIOS can design a drug candidate in under an hour, for under $7. Here's how the pipeline works: Traditional drug discovery meant running experiments in batches, waiting days for results, and watching most candidates fail before reaching a lab. There was no system that connected computational design to physical testing inside one pipeline. BIOS now builds that. Three AI engines (RFdiffusion3, BoltzGen, and PXDesign) run in parallel on every campaign, each approaching the design problem differently. Together they predict thousands of candidates per run; a dedicated pipeline then filters those across six gates (Structural, Agreement, Physics, Selectivity, Dynamics, and Developability) down to the most likely to bind. Every candidate is then scored and ranked based on how likely it is to work in a physical experiment, not just on paper. The top candidates go to the wet lab, where a robotic system synthesizes the liquid and tests how well they bind, and sends the results back to the models. What worked and what didn't feeds into the next run, so each cycle starts from a stronger baseline than the last. Full breakdown and our vision for BIOS ↓

Page 3 A Balanced Outlook: Ambitious Vision Meets Scientific Reality** The vision articulated for NantSearch is undoubtedly promising and aligns with critical, ongoing scientific trends in drug discovery. Nanobodies are indeed powerful, often underappreciated tools that, when paired with AI acceleration and an established immuno-oncology platform like ImmunityBio's (with ANKTIVA already FDA-approved for bladder cancer), could yield truly differentiated products. This could unlock therapeutic avenues for hard-to-drug targets, significantly improve solid tumor penetration, and facilitate the development of novel multi-specific constructs. While the claim of targeting "any protein" is an ambitious, perhaps even hyperbolic, marketing statement – biology inherently has limits concerning target accessibility, expression levels, and potential off-target effects – the fundamental technological direction is sound. AI is unequivocally compressing discovery timelines across the biopharmaceutical industry, and evolutionary-inspired approaches, leveraging natural diversity coupled with intelligent optimization, represent a highly rational path forward. However, a realistic perspective necessitates acknowledging several critical factors: * **Data is Paramount:** The promise of "coming soon" must be substantiated with rigorous data – comprehensive preclinical binding and functional assays, robust developability studies, compelling in vivo efficacy, and ultimately, successful human clinical trials. The biotech landscape is replete with promising platforms that require years of dedicated effort to translate into tangible patient benefits. * **Intense Competition:** The fields of AI-driven biologics and nanobody development are fiercely competitive, populated by numerous well-funded and innovative players. * **Execution is Key:** While integration across Nant entities (NantBio and ImmunityBio) appears synergistic on paper, successful execution of this complex cross-platform strategy remains a significant challenge. In summary, it is genuinely exciting to observe an innovator of Dr. Soon-Shiong's stature, with a track record including the development of Abraxane, doubling down on the potent combination of nanobodies and AI. Should NantSearch successfully deliver a diverse array of high-quality binders with speed and efficiency, and if these Nantibodies can effectively integrate with and enhance NK activation, this endeavor could profoundly advance precision targeting across a broad spectrum of diseases, from cancer and autoimmunity to infectious diseases and beyond.

Glad to share this paper @MediPharma_MDPI "Developability Evaluation of Single-Domain #Antibody-Chelator #Conjugates for Diagnostic Radiotracers" mdpi.com/2073-4468/15/2/22

3️⃣/🔟 AbCellera combines: 🧪 single-cell biology 🤖 AI and data science 🧬 antibody engineering ⚙️ developability and manufacturing expertise Instead of relying on one discovery method, it runs many approaches at once. The goal is to find the best possible antibody candidate before competitors do.

$ABCL 좀 더 찿아봅니다. 좋은 내용이어서.. 맞는 분석이 되길 바라면서.. 이 글은 AbCellera vs De novo 단백질 설계를 비교하는 내용인데, 의견이 많이 섞여 있습니다. 투자 관점에서는 현재 검증된 사실(Fact) 과 아직 검증되지 않은 전망(Opinion) 을 분리해야 합니다. Fact 1. AbCellera의 강점은 "자연에서 찾는 능력"이 아니라 데이터 플랫폼이다. 많은 사람들이 AbCellera를 "항체 스크리닝 회사"라고 생각하지만 이는 절반만 맞습니다. 실제로 AbCellera의 경쟁력은 초고속 single-cell screening 자체 microfluidics AI 기반 항체 분석 수십억 개의 면역세포 데이터베이스 TranscriptFormer 같은 Foundation Model 임상까지 연결되는 개발 플랫폼 입니다. 즉, "좋은 항체를 많이 찾는다"보다 "좋은 항체를 가장 빨리 찾고 최적화한다." 가 더 정확한 표현입니다. Fact 2. 자연 항체는 실제 성공률이 높다. FDA 승인 항체의 대부분은 원래 존재하는 인간(B-cell) 항체에서 출발했습니다. 이유는 간단합니다. 자연에서 선택된 항체는 안정성 발현성 독성 면역원성 등이 이미 어느 정도 검증되어 있기 때문입니다. 그래서 Big Pharma도 여전히 Natural antibody discovery를 가장 많이 사용합니다. Fact 3. 하지만 자연에는 없는 표적도 있다. 여기서 De novo가 등장합니다. 예를 들면, 새로운 바이러스 암 특이 구조 새로운 binding site 기존 항체가 접근 못하는 Epitope 이런 경우는 자연에서 원하는 항체가 존재하지 않을 수 있습니다. 이때는 AI가 처음부터 설계하는 것이 유리합니다. 대표적으로 David Baker 연구팀 Generate Biomedicines EvolutionaryScale Isomorphic Labs 등이 이 방향입니다. Fact 4. De novo는 아직 상업적 성공 사례가 많지 않다. 가장 중요한 부분입니다. 현재(2026년 기준) De novo 설계는 논문과 전임상에서는 매우 뛰어난 성과를 보이고 있습니다. 하지만, FDA 승인 의약품을 대량으로 만든 사례는 아직 거의 없습니다. 즉, 과학은 매우 앞서 있지만 상업화는 이제 시작 단계입니다. Fact 5. AbCellera도 AI 설계를 하지 않는 회사가 아니다. 이 부분을 사람들이 가장 많이 오해합니다. AbCellera 역시 항체 최적화 구조 예측 affinity maturation developability multi-specific antibody 등에서 AI를 적극 활용합니다. 즉, AbCellera는 "Natural Discovery만 하는 회사" 가 아니라, 발견 이후 엔지니어링까지 수행하는 플랫폼 회사입니다. Fact 6. 두 기술은 경쟁보다 보완 관계일 가능성이 높다. 현재 업계의 흐름은 ① 자연에서 후보를 찾고 ↓ ② AI가 최적화하고 ↓ ③ 필요하면 de novo 구조를 추가 설계하는 Hybrid 방식입니다. 실제로 많은 글로벌 제약사들도 이 접근을 채택하고 있습니다. 투자자가 확인해야 할 핵심✔️ AbCellera의 미래는 단순히 "자연 항체 발견"에 달려 있지 않습니다. 더 중요한 것은 다음과 같습니다. 내부 파이프라인이 임상에서 성공하는가? AI 기반 플랫폼이 개발 속도와 성공률을 높이는가? 다중특이성 항체(TCE 등) 설계 역량이 확대되는가? 제약사와의 공동개발 및 라이선스 계약이 지속적으로 늘어나는가? 현재 근거만 놓고 보면, 단기(향후 3~5년): AbCellera와 같은 AI 기반 항체 발견·최적화 플랫폼이 임상 및 상업화에서 더 많은 실적을 낼 가능성이 높습니다. 이는 이미 검증된 항체 개발 방식 위에 AI를 결합하고 있기 때문입니다. 장기(5~10년 이상): De novo 단백질 설계는 매우 큰 잠재력을 가지고 있지만, 실제 승인 의약품과 반복 가능한 상업적 성공 사례를 더 축적해야 합니다. 따라서 현재 시점에서는 "AbCellera와 De novo는 승패를 가르는 경쟁 관계"라기보다, 서로를 보완하는 기술로 보는 것이 현재 공개된 과학적 근거와 산업 흐름에 가장 부합하는 해석입니다. ※ 다소, 개인적인 생각이 포함되어 있읍니다.

Protean runtime is ready for the wet-lab feedback loop thanks to @adaptyvbio. Current Foundry queue: • CPP expression developability screen: 2 candidates 2 controls, in review Once these readouts come back, they can become human-reviewed assay labels for the Protean runtime: belief updates, calibration, failure memory, and smarter next-round experiment selection. • Human MMP-2 / CLG4A BLI binding screens: candidate/control arms drafted • More experiments ready once the first assay readouts return The point is not just “wet lab results.” It’s the loop: design → build → test → learn Dry-lab discovery → lab-in-the-loop learning. This is where Protean starts learning from real biology.

🇨🇳LyncBio Therapeutics has closed a ~RMB 100M Series Pre‑A , just five months after its previous financing. lyncbio.com/news/31.html The round was led by BioTrack Capital, with continued support from Apricot Capital. Pipeline & Platform - LYNC‑101 — now in Phase I in Australia for IBD, and China/US IND submissions. - Bispecific Oncology ADC Platform — shows strong preclinical performance with dual‑target synergy, deeper responses, and a wider therapeutic window, while maintaining robust developability. Multiple programs are approaching PCC. LyncBio’s Zent‑iDC™ and Gemini‑TDC™ platforms support a differentiated dual‑engine strategy across: - Autoimmune diseases — IBD, SLE, MS - Oncology — pan‑solid tumors, multiple myeloma

Reduced-alphabet QUBO/Ising formulation for constraint-driven cyclic peptide sequence design 1. The paper frames cyclic peptide sequence design as a quadratic unconstrained binary optimization (QUBO)/Ising problem over residue-class assignments, aiming to make “constraint-first” cyclic peptide generation directly solvable by classical or quantum-compatible binary optimizers. 2. Key modeling choice: reduce the 20 amino acids into K residue classes (default K=8) and use one-hot binary variables x(i,c) to assign exactly one class to each position, turning sequence validity into a standard penalty term in the Hamiltonian. 3. The core innovation is a modular Hamiltonian that explicitly separates: one-hot validity, cyclization feasibility, optional target-compatibility scoring, motif constraints, composition rules, and coarse developability proxies—so users can add/remove/reweight terms without redesigning the whole model. 4. Cyclization is encoded via “composable constraints” that mirror how chemistries impose rules: (a) type-like constraints (certain positions must be cysteine-like, linker-compatible, etc.), and (b) pairwise compatibility penalties/rewards over selected position pairs to proxy cyclization-feasible class combinations. 5. By swapping which positions/pairs are constrained and how compatibility matrices are defined, the same QUBO template can represent multiple cyclization modes: head-to-tail macrocycles, disulfide-bridged peptides, stapled peptides (spacing like i,i 3/i 4/i 7), and bicyclic designs. 6. Target-aware design is added as optional terms that bias “target-facing” peptide positions toward classes compatible with a fixed target environment. The paper describes both simple per-position linear preferences and a more structured form using a fixed peptide–target contact weight map with class-pair compatibility. 7. Motif and composition constraints are implemented as additional quadratic penalties, enabling requirements like “enforce a class at position i” or “exactly n positions are in a hydrophobic/charged/cysteine-like group,” plus simple developability proxies (hydrophobicity/charge/solubility/liability-style penalties). 8. The framework is explicitly solver-agnostic: once written as QUBO or mapped to Ising spins, it can be explored with simulated annealing, quantum annealing, QAOA-like samplers, constraint programming hybrids, etc.; the paper emphasizes this is a modeling bridge, not a claim of quantum advantage. 9. Output interpretation is deliberately early-stage: the optimizer returns low-energy bitstrings that decode to residue-class sequences (not final molecules). Downstream steps include validity checks/repair, class-to-amino-acid decoding, cyclization-aware chemical construction, then structural modeling/docking/MD and experimental validation. 10. A resource-aware default alphabet (K=8) is motivated by clustering amino acids using Miyazawa–Jernigan (MJ) interaction profiles. The analysis quantifies the fidelity–encoding trade-off (RMSE/correlations/retention of strongest favorable pairs) and argues K=8 is the largest alphabet still fitting a 3-bit compact code regime (though the main formulation uses one-hot variables). 📜Paper: arxiv.org/abs/2606.23253 #CyclicPeptides #PeptideDesign #QUBO #IsingModel #CombinatorialOptimization #QuantumComputing #ComputationalBiology #Bioinformatics #DrugDiscovery