RapWords was a hip-hop "word of the day" blog I started in 2015 to highlight awesome vocabulary in rap lyrics. For fun, I brought it back with lyric discovery and synced video generation using Claude. Subscribe to hear your favourite rappers drop knowledge! www.instagram.com/rapwordstv/

Refreshing new startup idea: "Every app now has a chat interface, so to simplify things, we built a nice point-and-click GUI so we don’t have to chat with anything."

Love "Noun Project" but... have any of these designers seen a molecule before?

Searching for a DOI in 2026

"Yeah, I’ve got a multi-agent stack doing end-to-end development. One handles planning and orchestration, another does test and verification, and a third just vibes and boosts morale." "Oh, why am I needed? Taste, man."

I'm surprised, but not surprised, that ChatGPT citations often point to sites like ResearchGate. RG sucks but they seem to genuinely be helping piece together scientific literature "behind the paywall" that's enabling LLMs to deliver more value...

You have to be opportunistic with new technologies—mostly as a result of institutional paranoia. My surge in AI coding assistant use is primarily a race against the enshittification timeline.

Conventional wisdom: "Don't let your pipes freeze, be careful!!" Refreshing advice: "Almost let your pipes freeze, your tap water will be ice cold and delicious!"

As far as I can tell, the real reason why antimicrobial peptides aren't in clinical development is that models are too good. Shame it's a solved problem =(

Okay, now try a protein LLM trained on only ancestral proteins to reduce modern bias (a simpler time without so many protein fitness influencers) github.com/haykgrigo3/T...

"Taken together, these results demonstrate that [method X] is a general-purpose model capable of aiding in all stages of designing and optimizing [Y] in real-world drug discovery campaigns." (...Well, except for ligand-receptor structure prediction)

Patsnap is really putting the "landscape" in IP landscape with this one. Looking forward to the minecraft server extension to properly explore the IP ecosystem (watch out for patent trolls...)

Never thought I'd get a chance to smell a supernova in my lifetime. Thankful.

Toronto property tax portal is like: you must enter this random 21-digit number! Then on another page their like, screw it, just put in your address we'll auto-populate it for you.

You'd be inclined to believe this innocent helper function from ChatGPT is not flawed/wrong ("atom_number" does not exist in biotite, and it's not safe to assume hydrogen atom names start with "H", i.e. if they were output from Rosetta).

2023: "de novo antibody design is solved" -AbSci 2024: "de novo antibody design is solved, we raised $1B on this" -Xaira 2025: "de novo antibody design is solved, the last time was jokes" -Chai, Nabla 20XX: "de novo antibody design is solved, can anyone hear me... anyone out there?" -Startup X

Trick-or-treaters weren't the only ones wearing a mask this Halloween, doi.org/10.1093/nar/...

Critics: "Boo, -latest/greatest AI drug discovery method- only works if you have a similar training example!" Biotech Industry: "Hurry, let's find the smallest possible modification to this approved ligand which hits the identical target and epitope"

Real data tracking my head nodding at the Autechre tour stop in Toronto last night

“What fascinated me was the fact that you could potentially predict which molecule could have this effect using thermodynamic principles in silica” www.fiercebiotech.com/special-repo...

"We discovered a TfR1-targeted brain shuttle, check out the brain uptake" (Crook et al. 2020, doi.org/10.1016/j.jm...). "This radioligand does not efficiently cross the blood-brain barrier" (Lin et al. 2025, doi.org/10.1007/s002...)

When will a generative model pass the "Internet Turing Test"? That is, a test subject cannot distinguish an AI-generated internet from the real internet.

Scientists: answering the questions we were dying to know, but too afraid to ask

There's a form of "Personal Computer" Stockholm syndrome where an accumulation of weird tech issues makes computer use border on torture, with a surprising user tolerance threshold. Usually it's only identified by a 3rd party using your machine and remarking... "How the heck do you use this thing?"

"The Illiad" is kinda like "Three-Body Problem" in that their both about vengeful omnipresent beings fucking around with physics on earth.

Sequence logo's are so underutilized, even in protein design papers. Mainly because all the libraries to create them have zero imagination for applications. Represent ensembles of any residue-level property with symbol/color, or use pos/neg enrichment to show differential bias between any two sets!

Grant me the confidence of a scientist asking for help installing software in the Boltz2 webinar Zoom chat

Increasingly convinced there's collusion in "big spinach" to sell washed leaves "wet" in grocery store tubs to accelerate rot and increase sales 🤫🥬💧🤑

leetspeak still alive in the bioinformatics community, academic.oup.com/nar/advance-...

Kinda like mining bitcoin

Love and respect the four-digit PDB code convention, and will miss it. I like how you can use it to roughly "date" structures. Like, 3J5P as a marker for the start of the cryo-EM membrane protein era in 2013.

As if the big scientific publishers weren't bad enough, they sometimes just take journals offline. "Current Opinion in Drug Discovery & Development" ran from 1998-2010. Owned by Thomson but almost all links are broken with no other hosts (that I can tell), pubmed.ncbi.nlm.nih.gov?term=%22Curr...

Biology is hard example #103342: "Our in vivo screening efforts confirmed the requirement for high-affinity binding of the ligand for the cognate receptor to achieve robust activity" (*whisper* R^2 of 0.14) academic.oup.com/nar/article/...

Pharma: "Drug discovery is so hard, we have great binders but they fail in preclinical and clinical research" AI Drug Discovery companies: "So... if we make better binders all diseases will be cured?"

Oral-B Pro 1000 and Philips Sonicare 4100 are the canonical "best electric toothbrushes", but they are really different UX. Philips brush size and vibration is so small it's basically regular brushing with a buzz, but with the Oral-B its truly elevated brushing on easy mode (not a paid advert!)

Academic in Bio ML: "Oh rats, my inverse design sequence recovery metric is really poor on this complex... better lower the temperature." Biotech patent: "In some embodiments, our claim is the endogenous ligand and anything over 20% identity. Trust me look at all these sequences we generated."

In the zone. Completely in sync with the universe. Measured the same weight on my digital scale 3 days in a row. That's even harder than losing or gaining weight 👏👏

How many yogurt lid related cuts go unreported each year due to shame/disgrace? Destigmatize yogurt lid related injury and hold the diary industry accountable.

It's probably been said before, but random Reddit comments with like 1 upvote have a lot of power for specific recommendations. For example, trip planning: city A vs B, or hotel A vs B. Is there any better source of "truth" than these anecdotes? They can rank high on search, pretty empowering...

"I'll be candid: by now you know I care about you personally, but I need to challenge you directly so my behavior doesn't come across as ruinously empathetic or obnoxiously aggressive. Really, it's the only way I can give you feedback so I can sincerely manipulate. Do you have any feedback on that?"

Okay, so it's almost a year after AF3 reported ~75% accuracy on structural prediction of glycosylation. What's the status update on predicting the glycosylated structural proteome?

"Our founding team has combined experience of 20K A100 GPU years"

Excellent study on peptide discovery and characterization, but I was not expecting the computational heavy-lifting of their prediction methodology to be... regex 😜 github.com/Svensson-Lab...

The art of writing a high impact manuscript, github.com/RosettaCommo...

In the VantAI webinar on Boltz-1 the authors describe challenges of arbitrary depositor-defined protein assemblies. Headaches incoming when they discover assembly is also dependent on experimental conditions 😅 E.g. www.biorxiv.org/content/10.1...

Academics may not appreciate the impact of licensing ML models/weights non-commercial use. I've *not* experimentally tested thousands of data points from predictive models because of it. Probably impossible, but there should be a license that requires anonymized validation data to flow to authors.

Ajinomoto Co doing flavour R&D on a whole other level...

How do you scramble a peptide binder? Not by randomly scrambling the sequence! Must account for binding motifs (can have gaps), physicochemical similarity, structure, sequence complexity, chemical identity (like NCAA anchor residues). Chirality-swapped retro-inversions are not a silver bullet...

The field needs a benchmark for benchmarking datasets that ensures collections of data have high scores for data relevance, diversity, quality, and standardization, balanced splits, data annotation, reproducibility, and evaluation metrics; otherwise we're in for "yet another"...

ChatGPT: "Conformational monism" denotes a belief system or viewpoint that a given protein sequence has only one definitive, biologically relevant 3D conformation. Ex. The researcher remained a conformational monist, insisting that their AlphaFold2 model was sufficient to explain structure-function.