Ryohei Yamaguchi

@porunga00000

ケモインフォマティクスに興味があります。

Joined May 2020

198 Following

34 Followers

76 Posts

porunga00000 retweeted

Biology+AI Daily @BiologyAIDaily

15 days ago

Distribution-Constrained Optimization for Reliable ML-Guided 5′UTR Sequence Design 1. The paper studies a common failure mode in ML-guided 5′UTR optimization: genetic algorithms can “reward hack” a translation-efficiency predictor by drifting out of the model’s validated training distribution, producing sequences with high predicted mean ribosome load (MRL) but elevated risk of misprediction. 2. Their core proposal is a trust-region style constrained optimization: keep candidate 5′UTRs inside the predictor’s training distribution (where accuracy has been validated). “Reliable” is defined as in-distribution with respect to training data, not as a guarantee of wet-lab performance. 3. To define out-of-distribution (OOD) for nucleotide sequences, they compare two scores computed from UTR-Insight’s internals: pseudo-perplexity (PPPL) vs k-nearest-neighbor (KNN) distance in the model’s embedding space. PPPL fails to separate in- vs out-of-distribution in this setting (the 4-letter alphabet makes PPPL broadly permissive), while embedding KNN distance provides a usable OOD signal. 4. They show OOD drift is not hypothetical: under unconstrained GA optimization, 72–96% of final-generation candidates exceed the self-KNN p95 threshold (across seeds), and the median KNN distance increases ~4.7-fold from early generations—consistent with optimization pushing into extrapolative regions. 5. Using KNN distance as a hard GA constraint (feasibility boundary set by percentiles of the training data’s self-KNN distribution) keeps all candidates within the trust region while maintaining predicted MRL at essentially the unconstrained level (median predicted MRL ~9.8–10.1 across seeds under KNN-Pr). 6. A practical payoff: compared with post-hoc filtering of unconstrained outputs under the same compute budget, the hard KNN constraint yields ~4.3× more “selectable” low-risk candidates (i.e., high predicted MRL without leaving the trust region). 7. They examine which reference distribution to use for KNN: (Pr) unlabeled native 5′UTRs from pretraining corpora vs (Sv) the supervised MPRA library used for predictor training. Distances correlate strongly (Pearson ~0.94), but the native-reference trust region is effectively stricter: KNN-Pr constraint tends to satisfy both Pr and Sv trust regions, while KNN-Sv may not satisfy Pr consistently across seeds. 8. They compare other search-space controls: an output extrapolation guard (constraining predicted MRL to be within the supervised training MRL range, e.g., p95=8.28) strongly suppresses OOD as a side effect, but it also caps predicted MRL and concentrates solutions near the ceiling—trading performance headroom for conservatism. 9. They test additional objectives/constraints often used in 5′UTR design: adding RNA secondary-structure accessibility near the start codon (RNAplfold) as a secondary objective broadens exploration and creates a Pareto trade-off with predicted MRL, but does not suppress OOD risk; reference-sequence similarity suppresses OOD when used as a constraint (at the cost of lower MRL), while using similarity as an objective enables fine-grained “distance-from-reference” control but can still include many OOD candidates. 📜Paper: https://t.co/Drj63MuszW #ComputationalBiology #SyntheticBiology #mRNA #UTR #MachineLearning #OutOfDistribution #GeneticAlgorithms #SequenceDesign #TrustRegion #Bioinformatics

BiologyAIDaily's tweet photo. Distribution-Constrained Optimization for Reliable ML-Guided 5′UTR Sequence Design

1. The paper studies a common failure mode in ML-guided 5′UTR optimization: genetic algorithms can “reward hack” a translation-efficiency predictor by drifting out of the model’s validated training distribution, producing sequences with high predicted mean ribosome load (MRL) but elevated risk of misprediction.

2. Their core proposal is a trust-region style constrained optimization: keep candidate 5′UTRs inside the predictor’s training distribution (where accuracy has been validated). “Reliable” is defined as in-distribution with respect to training data, not as a guarantee of wet-lab performance.

3. To define out-of-distribution (OOD) for nucleotide sequences, they compare two scores computed from UTR-Insight’s internals: pseudo-perplexity (PPPL) vs k-nearest-neighbor (KNN) distance in the model’s embedding space. PPPL fails to separate in- vs out-of-distribution in this setting (the 4-letter alphabet makes PPPL broadly permissive), while embedding KNN distance provides a usable OOD signal.

4. They show OOD drift is not hypothetical: under unconstrained GA optimization, 72–96% of final-generation candidates exceed the self-KNN p95 threshold (across seeds), and the median KNN distance increases ~4.7-fold from early generations—consistent with optimization pushing into extrapolative regions.

5. Using KNN distance as a hard GA constraint (feasibility boundary set by percentiles of the training data’s self-KNN distribution) keeps all candidates within the trust region while maintaining predicted MRL at essentially the unconstrained level (median predicted MRL ~9.8–10.1 across seeds under KNN-Pr).

6. A practical payoff: compared with post-hoc filtering of unconstrained outputs under the same compute budget, the hard KNN constraint yields ~4.3× more “selectable” low-risk candidates (i.e., high predicted MRL without leaving the trust region).

7. They examine which reference distribution to use for KNN: (Pr) unlabeled native 5′UTRs from pretraining corpora vs (Sv) the supervised MPRA library used for predictor training. Distances correlate strongly (Pearson ~0.94), but the native-reference trust region is effectively stricter: KNN-Pr constraint tends to satisfy both Pr and Sv trust regions, while KNN-Sv may not satisfy Pr consistently across seeds.

8. They compare other search-space controls: an output extrapolation guard (constraining predicted MRL to be within the supervised training MRL range, e.g., p95=8.28) strongly suppresses OOD as a side effect, but it also caps predicted MRL and concentrates solutions near the ceiling—trading performance headroom for conservatism.

9. They test additional objectives/constraints often used in 5′UTR design: adding RNA secondary-structure accessibility near the start codon (RNAplfold) as a secondary objective broadens exploration and creates a Pareto trade-off with predicted MRL, but does not suppress OOD risk; reference-sequence similarity suppresses OOD when used as a constraint (at the cost of lower MRL), while using similarity as an objective enables fine-grained “distance-from-reference” control but can still include many OOD candidates.

📜Paper: https://t.co/Drj63MuszW
#ComputationalBiology #SyntheticBiology #mRNA #UTR #MachineLearning #OutOfDistribution #GeneticAlgorithms #SequenceDesign #TrustRegion #Bioinformatics

porunga00000 retweeted

Biology+AI Daily @BiologyAIDaily

10 months ago

BoltzGen: Toward Universal Binder Design 1. BoltzGen is introduced as a novel all-atom generative model for universal binder design, uniquely capable of creating proteins and peptides of all modalities, including nanobodies, cyclic peptides, and globular proteins, to target a vast array of biomolecules, such as proteins, small molecules, and nucleic acids. 2. This universal approach was rigorously validated in eight diverse wet-lab campaigns across 26 targets, demonstrating exceptional generalization capacity, especially against nine novel targets with no similar bound structures in the PDB. 3. For these highly novel targets, BoltzGen successfully designed nanomolar (nM) affinity nanobodies and protein binders for 66% of the tested targets, highlighting its ability to extrapolate design solutions beyond complexes observed during training. 4. A core innovation of the model is the unification of the design process with all-atom structure prediction, allowing it to build strong structural reasoning capabilities and simultaneously match the state-of-the-art performance of dedicated folding models. 5. Researchers can precisely control the design process using a flexible specification language to impose constraints such as covalent bonds, structural motifs, and specific binding sites, enabling complex tasks like designing disulfide-bonded cyclic peptides and targeting disordered protein regions. 6. The BoltzGen pipeline is presented as a comprehensive, end-to-end solution for real-world binder design problems, including candidate generation, filtering, ranking, and optimization for diversity, which is now open-sourced to the research community. 💻Code: https://t.co/qMrw9tJunB 📜Paper: https://t.co/1cz7MWtn3j #ComputationalBiology #ProteinDesign #DeNovoDesign #AIinScience #BiomolecularDesign #MachineLearning

BiologyAIDaily's tweet photo. BoltzGen: Toward Universal Binder Design

1. BoltzGen is introduced as a novel all-atom generative model for universal binder design, uniquely capable of creating proteins and peptides of all modalities, including nanobodies, cyclic peptides, and globular proteins, to target a vast array of biomolecules, such as proteins, small molecules, and nucleic acids.

2. This universal approach was rigorously validated in eight diverse wet-lab campaigns across 26 targets, demonstrating exceptional generalization capacity, especially against nine novel targets with no similar bound structures in the PDB.

3. For these highly novel targets, BoltzGen successfully designed nanomolar (nM) affinity nanobodies and protein binders for 66% of the tested targets, highlighting its ability to extrapolate design solutions beyond complexes observed during training.

4. A core innovation of the model is the unification of the design process with all-atom structure prediction, allowing it to build strong structural reasoning capabilities and simultaneously match the state-of-the-art performance of dedicated folding models.

5. Researchers can precisely control the design process using a flexible specification language to impose constraints such as covalent bonds, structural motifs, and specific binding sites, enabling complex tasks like designing disulfide-bonded cyclic peptides and targeting disordered protein regions.

6. The BoltzGen pipeline is presented as a comprehensive, end-to-end solution for real-world binder design problems, including candidate generation, filtering, ranking, and optimization for diversity, which is now open-sourced to the research community.

💻Code: https://t.co/qMrw9tJunB
📜Paper: https://t.co/1cz7MWtn3j
#ComputationalBiology #ProteinDesign #DeNovoDesign #AIinScience #BiomolecularDesign #MachineLearning

123

porunga00000 retweeted

Bioinformatics Advances

@BioinfoAdv

10 months ago

🧠 Now published in Bioinformatics Advances: “PSO-FeatureFusion: A general framework for fusing heterogeneous features via particle swarm optimization.” Read the full paper here: https://t.co/Mc55Y1ZJ1E

BioinfoAdv's tweet photo. 🧠 Now published in Bioinformatics Advances: “PSO-FeatureFusion: A general framework for fusing heterogeneous features via particle swarm optimization.”

Read the full paper here: https://t.co/Mc55Y1ZJ1E https://t.co/Yv1kBvB8NY

156

porunga00000 retweeted

Biology+AI Daily @BiologyAIDaily

9 months ago

High-accuracy structure modeling for antibody-antigen complexes 1. A new computational protocol called DeepAAAssembly has been introduced to enhance the accuracy of antibody-antigen complex modeling. This method integrates deep learning-based inter-chain residue distance predictions with a two-stage conformational sampling strategy to capture both global orientation and local flexibility. 2. DeepAAAssembly leverages predicted inter-chain residue distances to construct a flexibility-aware energy function, which drives a two-stage sampling process. The first stage explores global orientations using a multi-objective optimization algorithm, while the second stage refines local structures, particularly focusing on the highly flexible CDR loops. 3. Tested on a benchmark set of 67 antibody-antigen complexes, DeepAAAssembly outperformed AlphaFold3, achieving a 12.9% higher average DockQ score and producing more medium- and high-quality models. This demonstrates its ability to reliably elevate challenging cases from incorrect to acceptable accuracy. 4. The method incorporates a built-in confidence selection mechanism that identifies the most structurally optimal and energetically favorable models. This feature ensures that DeepAAAssembly not only generates high-quality models but also reliably selects the best ones. 5. DeepAAAssembly’s success lies in its ability to capture conserved interaction motifs and conformational flexibility. By integrating multi-scale contextual information and considering relative positional relationships, it provides a robust framework for high-accuracy antibody-antigen modeling. 6. The study highlights the potential of combining deep learning with traditional docking approaches to address the complexity of antibody-antigen interactions. This hybrid strategy effectively bridges the gap between end-to-end AI models and physics-based docking methods. 📜Paper: https://t.co/cU433D8xUj #AntibodyAntigen #ProteinModeling #DeepLearning #StructuralBiology #Bioinformatics

BiologyAIDaily's tweet photo. High-accuracy structure modeling for antibody-antigen complexes

1. A new computational protocol called DeepAAAssembly has been introduced to enhance the accuracy of antibody-antigen complex modeling. This method integrates deep learning-based inter-chain residue distance predictions with a two-stage conformational sampling strategy to capture both global orientation and local flexibility.

2. DeepAAAssembly leverages predicted inter-chain residue distances to construct a flexibility-aware energy function, which drives a two-stage sampling process. The first stage explores global orientations using a multi-objective optimization algorithm, while the second stage refines local structures, particularly focusing on the highly flexible CDR loops.

3. Tested on a benchmark set of 67 antibody-antigen complexes, DeepAAAssembly outperformed AlphaFold3, achieving a 12.9% higher average DockQ score and producing more medium- and high-quality models. This demonstrates its ability to reliably elevate challenging cases from incorrect to acceptable accuracy.

4. The method incorporates a built-in confidence selection mechanism that identifies the most structurally optimal and energetically favorable models. This feature ensures that DeepAAAssembly not only generates high-quality models but also reliably selects the best ones.

5. DeepAAAssembly’s success lies in its ability to capture conserved interaction motifs and conformational flexibility. By integrating multi-scale contextual information and considering relative positional relationships, it provides a robust framework for high-accuracy antibody-antigen modeling.

6. The study highlights the potential of combining deep learning with traditional docking approaches to address the complexity of antibody-antigen interactions. This hybrid strategy effectively bridges the gap between end-to-end AI models and physics-based docking methods.

📜Paper: https://t.co/cU433D8xUj
#AntibodyAntigen #ProteinModeling #DeepLearning #StructuralBiology #Bioinformatics

Who to follow

かるべん

@carbene_

しがないデータサイエンティスト仕事でMI(Materials Informatics)に携わっています

porunga00000 retweeted

LangChain

@LangChain

over 2 years ago

1/ 🌲Language Agent Tree Search (LATS) LATS is a unified algorithm that uses self-reflection (and additional external feedback) to self-optimize and efficiently adapt to new tasks and environments. It performs better than similar algorithms like Tree of Thoughts, ReAct, and Reflexion. LATS unifies a few important trends in AI: 1. Search (algorithmic) 2. Planning & "Reasoning" 3. Executing / Acting It does this without any updates to the model (unlike complicated RL). We created a simple implementation of LATS in LangGraph to demystify things, so you can appropriately apply it in your applications. Check out the technical walkthrough and other results in the thread: Paper by @andyz245, @aiatillinois,@HaohanWang, et. al.

LangChain's tweet photo. 1/ 🌲Language Agent Tree Search (LATS)

LATS is a unified algorithm that uses self-reflection (and additional external feedback) to self-optimize and efficiently adapt to new tasks and environments. It performs better than similar algorithms like Tree of Thoughts, ReAct, and Reflexion.

LATS unifies a few important trends in AI:

1. Search (algorithmic)
2. Planning & "Reasoning"
3. Executing / Acting

It does this without any updates to the model (unlike complicated RL).

We created a simple implementation of LATS in LangGraph to demystify things, so you can appropriately apply it in your applications.

Check out the technical walkthrough and other results in the thread:

Paper by @andyz245, @aiatillinois,@HaohanWang, et. al.

379

308

39K

Ryohei Yamaguchi

@porunga00000

Who to follow

Last Seen Users on Sotwe

Trends for you

Most Popular Users