Bio-PrecisionAI Health

10/29/2025

By Bo Wang, PhD

10/29/2025

10/29/2025

BoltzGen Democratizes AI Therapeutic Design, Expands Druggable Universe

The latest Boltz open-source model designs therapeutics across "any" modality with experimental validation for diverse real-world challenging targets

Link: https://www.genengnews.com/topics/artificial-intelligence/boltzgen-democratizes-ai-therapeutic-design-expands-druggable-universe/

Copied

The open-source model designs therapeutics across "any" modality with experimental validation for diverse real-world challenging targets.

10/29/2025

Boltz-2 Released to Democratize AI Molecular Modeling for Drug Discovery

The AlphaFold 3 competitor from MIT predicts molecular binding affinity at unprecedented speed and accuracy, offering a powerful open-source tool for commercial drug discovery.

Link: https://www.genengnews.com/topics/artificial-intelligence/boltz-2-released-to-democratize-ai-molecular-modeling-for-drug-discovery/

Copied

The MIT model predicts molecular binding affinity at newfound speed and accuracy, offering a powerful tool for commercial drug discovery.

10/29/2025

The genetic code is full of synonymous codons that, for decades, were assumed to be interchangeable.

Today, in collaboration with NVIDIA and led by Hani Goodarzi and Laksshman Sundaram, we announce CodonFM––a family of open-source AI models that reveal the intricate grammar underlying codon choice.

There is a natural redundancy in the genetic code. 64 codons encode 20 amino acids, resulting in multiple "synonymous" codons each encoding for the same amino acid. Some of these, however, appear far more often than others, following consistent, non-random patterns across genes.

Trained on 130 million coding sequences from more than 20,000 species, CodonFM uses large-scale language modeling to uncover the patterns behind codon choice and reveal the regulatory logic that links sequence variation to gene expression and protein abundance.

Two complementary architectures power the CodonFM family: Encodon, released today, interprets codon context to understand regulatory effects and mutations, and Decodon, to be released later this year, generates optimized sequences for design applications.

As the models scaled, a clear grammar began to emerge. Smaller versions hinted at weak patterns, but at the billion-parameter scale, CodonFM could accurately predict which codons cells would choose in context—uncovering long-range dependencies that link codon choice to translation and expression.

By modeling codon usage directly, CodonFM achieves multi-fold improvements over prior approaches, revealing how synonymous variation influences expression and opening doors for advances in clinical genetics and therapeutic design.

- Read the preprint: https://lnkd.in/exFMss-y
- Read a technical blog about the CodonFM model: https://lnkd.in/e5MsxMwG
- Read a conversation with the project leads: https://lnkd.in/e6wuGudd

Copied

This link will take you to a page that’s not on LinkedIn

10/27/2025

A GENETIC ALGORITHM FOR NAVIGATING
SYNTHESIZABLE MOLECULAR SPACES

The SynGA genetic algorithm model is a specialized algorithm designed to generate novel molecules that are explicitly constrained to be synthesizable. Unlike traditional genetic algorithms (GAs) that optimize for properties alone, SynGA evolves synthesis routes directly, ensuring that the generated molecules can actually be manufactured.

Developed by researchers at MIT, SynGA is used for molecular design tasks such as analog searching and property optimization, and can also be integrated into more advanced workflows.

How SynGA works

SynGA's process is similar to a standard GA but with key modifications to enforce synthesizability.
Initialization: The algorithm creates an initial population of synthesis routes rather than generating random molecules. These routes are constructed from a predefined set of commercially available molecular building blocks and known reaction templates.

Fitness evaluation: Each synthesis route is evaluated by a fitness function based on the desired molecular properties (e.g., solubility, toxicity, or a 2D/3D objective).

Selection: Better-performing routes are selected to "reproduce" for the next generation.
Custom genetic operators: SynGA uses special crossover and mutation operations that mix and alter the synthesis routes while maintaining their validity and ensuring the resulting "offspring" molecules are still synthesizable.

Enhanced variants: SynGA can be augmented with a machine learning-based filter that dynamically restricts the building block set to those most relevant for the optimization task. A model-based variant called SynGBO uses this filter within a Bayesian optimization framework to achieve state-of-the-art performance.

Key features and advantages

Synthesis-aware design: By operating on synthesis routes, SynGA guarantees that all generated molecules are synthetically accessible, which is a major bottleneck in traditional molecular design methods.

Lightweight and modular: The core SynGA algorithm is simple and machine-learning-free, allowing it to be used as a standalone tool or as a modular component in more complex systems.

Competitive performance: Benchmarks show that SynGA and its augmented versions achieve state-of-the-art performance on a variety of molecular design optimization tasks.

Efficient for inference: Although it can be slower than amortized models during inference, SynGA is more lightweight and doesn't require extensive training on large datasets. This makes it easier to adopt with new building blocks and reaction sets.

Link: https://arxiv.org/abs/2509.20719

Copied

Inspired by the effectiveness of genetic algorithms and the importance of synthesizability in molecular design, we present SynGA, a simple genetic algorithm that operates directly over synthesis routes. Our method features custom crossover and mutation operators that explicitly constrain it to synth...

10/19/2025

Exome Sequencing (ES) vs. Genome Sequencing (GS): Understanding the Difference ‼️🧬

🧩 Genome Sequencing (GS):
Genome Sequencing covers the entire human genome — around 3 billion base pairs — including both coding and non-coding regions.

Data Analysis – Sequences are aligned to the reference genome to identify variants across the entire DNA sequence.

GS provides a complete view of the genome, useful for studying:
- Structural and regulatory variants
- Non-coding mutations
- Complex diseases and rare genetic disorders

🎯 Exome Sequencing (ES):
Exome Sequencing focuses only on the coding regions (exons) of genes — the parts that are translated into proteins — representing about 1–2% of the genome (30–60 million base pairs).

Data Analysis – Reads are aligned to the reference genome, and variants are identified in coding regions.

Since about 85% of known disease-causing mutations occur in coding sequences, ES is highly effective for:

- Identifying pathogenic variants in genes
- Diagnosing Mendelian or monogenic disorders
- Cost-effective clinical diagnostics



Copied

10/14/2025

Diffusion Transformers with Representation Auto-encoders

Link: https://arxiv.org/abs/2510.11690v1

Copied

Latent generative modeling, where a pretrained autoencoder maps pixels into a latent space for the diffusion process, has become the standard strategy for Diffusion Transformers (DiT); however, the autoencoder component has barely evolved. Most DiTs continue to rely on the original VAE encoder, whic...

10/14/2025

🧠 What Is In Silico Drug Design?

“In silico” refers to computational or computer-based methods used to simulate, model, and predict biological and chemical phenomena. In the context of drug discovery, in silico drug design (also part of CADD — Computer-Aided Drug Design) is the use of computational tools to:

Predict how small molecules interact with targets
Filter large chemical libraries
Estimate pharmacokinetics and toxicity early on
Reduce wet-lab costs and failures

Because of advances in computing, machine learning, AI, and bioinformatics, in silico methods are now central to modern drug pipelines, reducing dependence on brute-force screening and animal testing.

🔍 Types & Methods of In Silico Drug Design

1. Structure-Based Drug Design (SBDD)
Uses the 3D structure of the protein target (X-ray, NMR, Cryo-EM, or predicted) and docks ligands into the binding site. Methods include molecular docking, induced fit docking, binding free energy calculations (MM-GBSA, MM-PBSA).

2. Ligand-Based Drug Design (LBDD)
Used when target structure is unknown. Techniques include:
QSAR / 2D & 3D QSAR
Pharmacophore modeling
Similarity searching

3. De Novo & Generative Design
Algorithms generate novel molecules from scratch, guided by scoring functions (affinity, synthetic accessibility, ADMET). E.g. diffusion models, reinforcement learning.

Methods:
1. Molecular Dynamics (MD) & Enhanced Sampling
Simulates the time-dependent behavior of molecules to explore conformational flexibility, binding/unbinding events, stability. Methods include classical MD, accelerated MD, metadynamics.

2. Free Energy Calculations
MM-GBSA, MM-PBSA, alchemical free energy methods (FEP) to more accurately predict binding energy differences.

3. ADMET / Toxicity Prediction
Predict absorption, distribution, metabolism, excretion, and toxicity (hepatotoxicity, cardiotoxicity, hERG, neurotoxicity) using machine learning models, neural nets, graph-based methods.

4. Drug Repurposing / Virtual Screening
Screen existing drugs or public libraries for new targets. Use docking, similarity, or ML-based models.

5. In Silico Clinical Trials / Model-Informed Drug Development (MIDD)
Simulate populations, disease progression, PK/PD models to predict clinical outcomes or guide trials.

🛠️ Key Tools & Platforms (2025 Era)

Some widely used tools, databases, and platforms include:
Docking / Virtual Screening Tools: AutoDock Vina, PyRx, Glide, GOLD, SwissDock
Structure tools: AlphaFold / ColabFold for predicting protein structure
Cheminformatics / ML Libraries: RDKit, DeepChem, scikit-learn
QSAR / ML Platforms: MOE, KNIME, Weka, PaDEL
Toxicity / ADMET Prediction: pkCSM, ProTox, DeepTox, AI models for hepatotoxicity / cardiotoxicity
Generative AI Platforms: Diffusion models, IDOLpro (multi-objective generative AI)
Clinical Simulation Tools: software for in silico trial simulation / model-informed design
Databases: ChEMBL, PubChem, DrugBank, ZINC, BindingDB
Open Platforms / Portals:

Copied

10/13/2025

Skepticism About COVID-19 Vaccines Among Christians!

I wrote this on a Christian group I belong to, and feel you may find it useful as well:

No credible evidence supports the fear surrounding COVID-19 vaccines. Are the vaccines 100%, NO! but most of them show acceptable baseline in clinical trials with more than 90% safety, just like other vaccines. I’m a child of God, I believe in God, I also believe in science, I believe it’s God’s calling for me. There is a lot of disinformation and ignorance in the responses here being demonstrated as strength, maturity, etc. I read Biochemistry, have two Masters degrees in Bioinformatics and doing PhD in Biomedical Engineering. If you believe in laws such as “whatever goes up must come down,” and you know that a vehicle will not remain stationary if you put it in drive and press the gas, and if you live in 50-story buildings trusting their engineering principles to keep them standing—then why do you find it difficult to believe in vaccines, which you have always been a beneficiary of in the past, present, and future?

Multiple pre-clinical studies found no effect on fertility or reproductive organs. Real-world data on adenoviral and mRNA vaccines show no increased risk of birth defects or infertility. Because the vaccine does not enter reproductive cells or integrate into DNA, there’s no pathway for it to alter s***m, eggs, or developing embryos. However, pregnant women were initially excluded from early trials, so later recommendations relied on observational safety data. As of current scientific consensus (WHO, EMA, CDC), adenoviral and mRNA COVID vaccines are considered safe before, during, and after pregnancy. The only thing that stays in your immune system is the immune memory (B cells, T cells). If you are rejecting some vaccines, do you know that the same technologies were used in manufacturing other vaccines so dear to you, such as cancer vaccines, common cold, etc. The pressing questions you should be asking, which comes from critical thinking should be: Can immune memory cause new diseases? No, it’s antigen specific and dormant. Does the spike protein stay forever? No, it’s broken down within weeks. Can immune memory be inherited? No, only temporary antibodies transfer from mother to baby. Does it change your DNA? No integration or heritable effect. The vaccine trains your immune system temporarily using a small piece of viral information. After that, all that remains are memory cells — like your body’s “notes” on how to fight the virus next time. They don’t rewrite genes, cause other illnesses, or pass into future generations.

So:
• Immune memory = personal and temporary.
• Genetic inheritance = permanent and germline-based.
The two do not overlap.

~ Joseph Luper Tsenum