Ash Jogalekar

Night Thoughts on the Promise of Agentic AI in Drug Discovery It's here, and it's real

Night Thoughts on the Promise of Agentic AI in Drug Discovery medchemash.substack.com/p/night-thou...

Night Thoughts on the Promise of Agentic AI in Drug Discovery It's here, and it's real

Night Thoughts on the Promise of Agentic AI in Drug Discovery medchemash.substack.com/p/night-thou...

Covalent inhibitors of the PI3Kα RAS binding domain impair tumor growth driven by RAS and HER2 Genetic disruption of the RAS binding domain (RBD) of Phosphoinositide 3-kinase alpha(PI3Kα) impairs the growth of tumors driven by the small guanosine triphosphatase RAS in mice and does not impact P...

This looks promising. www.science.org/doi/10.1126/...

😂

LLMs for Computational Drug Design The continuing saga

New post: LLMs for Computational Drug Design
The continuing saga medchemash.substack.com/p/llms-for-c...

Once again we find out that AI algorithms are based on memorization of training set cases rather than any kind of understanding. This makes them brittle and unable to adapt to new cases (of which there are many). I think a hybrid approach would probably work best.

In a nutshell: the authors perturb the binding site of known protein-ligand binding modes in a variety of ways, most of which should have abolished binding. But the AI algorithms (Boltz-1, Chai-1, AF3) still place the ligands in the exact same pose, ignoring clashes and forming spurious interactions

AI don't trump physics First principles remain important

New post about a very interesting new paper: When it comes to predicting protein-ligand binding, AI don't trump physics. medchemash.substack.com/p/ai-dont-tr...

RFK Jr. for "advances in vaccine innovation." Hits too close to home though.

Frustrated PROTACs: Exceptions or the rule? Discomfort may be productive

New post: Frustrated PROTACs: Exceptions or the rule? Aka discomfort may be productive. medchemash.substack.com/p/frustrated...

From descriptive to quantitative biocatalysis - Nature Catalysis The 1913 study ‘Die Kinetik der Invertinwirkung’, by Michaelis and Menten, marked a pivotal advancement in enzymology by illustrating the application of mechanistic models and quantitative kinetics to biocatalysis. The foundational framework described back then continues to have a strong impact on enzymology, with profound influences that range from undergraduate education to structure–function studies and the format and content of contemporary kinetic databases.

From descriptive to quantitative biocatalysis

“The 1913 study ‘Die Kinetik der Invertinwirkung’, by Michaelis and Menten, marked a pivotal advancement in enzymology by illustrating the application of mechanistic models and quantitative kinetics to biocatalysis.”

doi.org/10.1038/s419...

Will read the paper, but does "folding-docking-affinity prediction" indicate a co-folding method? Because those are new.

Code, not Drugs: The LLM Revolution in Computational Drug Design Process before product

Forget “AI discovering drugs”, as I say in this post, coding for computational drug design is where the AI/LLM revolution is. medchemash.substack.com/p/code-not-d...

The statistical is not the practical A guide to gauging the real world significance of computational methods

The statistical is not the practical:
A guide to gauging the real world significance of computational methods
medchemash.substack.com/p/the-statis...

Highly Position- and Enantioselective Catalytic Epoxidation of Polyolefins Mediated by a Chiral Mn Complex, Including a One-Step Conversion of Squalene to the (S)-2,3-Epoxide, a Precursor of Natural S... Reported herein is the synthesis of a novel chiral dicarboxylic ligand for Mn(II) and the application of the Mn complex to the highly enantio- and position-selective epoxidation of C═C under mild cond...

Amazing to see Corey still publishing at age 97! pubs.acs.org/doi/abs/10.1...

That was an unhinged, foul-mouthed performance even by Measles F. Kennedy Jr.’s standards.

"Since I take solace in the belief that God looks after drunks, little children, and the USA, I still think maybe, somehow, out of chaos will come something good." —Graham Allison

The Sarin shortcut: How AI lowers the bar for chemical weapons Generative AI and off-the-shelf computational tool can make the synthesis and use of nerve agents for nefarious purposes more probable.

Generative AI and off-the-shelf computational tools can make the synthesis and use of nerve agents for nefarious purposes more probable.

"The Sarin shortcut: How AI lowers the bar for chemical weapons," by @ashjogalekar.bsky.social. ⬇️

Things math and facts are for losers. Winners simply use their imagination.

We gave our infant son one literally the other day because the doctor showed us data on a COVID summer surge. I personally know at least two people who just got it last week from travel. We are living in the shadow of an administration who would deny the sky is blue.

Sheep who vote for a wolf because they think he will only eat other sheep.

10/ The key message is that we can use the same tools that are used to do harm to fight that harm. If AI is used by bad actors to design WMDs, AI can also be used to defeat or anticipate that misuse.

9/ Ultimately this is a classic dual-use dilemma: these tools are hugely beneficial for doing good, but as Feynman once put it, the same key that opens the doors of heaven can open the doors of hell.

8/ The fix: Shift to living lists updated in real time, integrate AI into threat prediction, and partner with AI developers that can anticipate and not just react to emerging threats. Organizations like the OPCW should adopt a "function first" approach.

7/ Policy problem: The Chemical Weapons Convention lists are static, structure-based, and slow. It took 27 months to list Novichok after its first known use. AI won’t give us that kind of lead time.

6/ LLMs make these workflows easier than ever. No coding, no special licenses, just prompt engineering to stitch together pathways, assays, and even equipment sourcing. RDKit is already part of chatGPT. More open-source tools will undoubtedly be embedded.

Identification of optimum computational protocols for modeling the aryl hydrocarbon receptor (AHR) and its interaction with ligands The aryl hydrocarbon receptor (AHR) is one of the principal xenobiotic receptors in living organisms and is responsible for interacting with several d…

5/ Computational chemistry can be inverted, maximizing instead of minimizing toxicity in drug discovery or agricultural chemistry programs. No repurposing of existing algorithms is needed. For instance in 2010 I compared docking algorithms for AHR binders www.sciencedirect.com/science/arti...

Dual use of artificial-intelligence-powered drug discovery - Nature Machine Intelligence An international security conference explored how artificial intelligence (AI) technologies for drug discovery could be misused for de novo design of biochemical weapons. A thought experiment evolved ...

4/ Machine learning toxicity models, meant for drug safety, can be inverted to maximize lethal effects. One 2022 study generated 40,000 VX analogs, many unlisted under current law www.nature.com/articles/s42...

3/ AI-guided retrosynthesis can provide alternative, legal, unscheduled chemicals for synthesizing dangerous agents, sidestepping the need for controlled precursors.

2/ Generative AI + modern chemistry software can now design unlisted, equally toxic nerve agent analogs in hours, bypassing the static lists in the Chemical Weapons Convention (CWC).