BioLM

We loved working with our client on this - we’re always excited to contribute toward a more sustainable future. Results in our first round impressed both us and our client, leaving us happy and awaiting another challenge! Looking forward to taking our enzyme design work further. We’re fast approaching a world in which we could get the desired molecular properties in just a few rounds! And are thrilled to be a part of it. Reach out if wondering whether your enzyme development could benefit from bio-AI! #proteinengineering #enzymes #techbio #ai #syntheticbiology #pllms #synbio #biotech #proteins

Our own Nikhil Haas, Ahmad Qamar, and Chance Challacombe will be hacking on AI-based enzyme and nanobody design Oct. 10-20th! So far, along with Vishwas Prabhu, Joe Davis, and Ryan Cloke as Team Silica, they will seek to demonstrate that state-of-the-art models, when used correctly and in the right workflows, can generate viable, stable, and diverse enzymes and therapeutics in just a few rounds! #techbio #bioai #enzymes #nanobodies #hackathon #synbio

We enjoyed participating in Adaptyv Bio's competition so much that we ordered and characterized an additional 36 nanobodies and will be releasing the data - stay tuned later this week. We used a different selection criteria than the competition, but for anyone wishing to model the EGFR results our dataset contains additional measurements and sequences that you can include. Kudos to Adaptyv Bio for creating such an impassioned response and competition! #proteinengineering #nanobodies #bioai #techbio

We're excited to bring our #biosimilars search to light, for rapidly and accurately discovering lead candidates with similar structure and biophysical properties compared to an existing protein. Whether you're looking for #antibodies in natural repertoires, or wild-type #enzymes or #proteins that share characteristics with an existing lead, embeddings-based based search can be a faster and more relevant alternative to BLAST and other sequence-identity based methods. The scale associated with biomolecular data has made this kind of method cost prohibitive: billions-scale datasets require significant computational resources. Together with Lance Dofflemyer of Tech42, we've tackled this challenge and pioneered TWO first-of-its-kind embeddings databases: * UniRef * OAS (Observed Antibody Space) For those unfamiliar with this type of similarity search, here are several open publications to help get you up to speed: https://lnkd.in/dKkpThbr https://lnkd.in/dkBz9dvq https://lnkd.in/dRfgc5F7 #synbio #bioai #llms #biotech #techbio

We are pleased to receive our second grant from NVIDIA Inception! Thanks to the team there for assistance and support! This funding will support our development of rapid semantic searches of #proteins and #DNA with structural and functional LLM #embeddings. Say goodbye to BLAST! https://lnkd.in/ePm5f7Pu #bioinformatics #bioml #nlp #bioai #techbio #gpu #nvidia #startups #biotech

We set out to build a generic predictor of variant effects on enzymes. ∆∆G and Tm data from dozens of publications and thousands of molecules were used as training data, along with LLM embeddings we computed from several models. Even without encoding experimental conditions into our model, we see a strong correlation between the predicted Group that a variant falls into and whether or not the mutations are stabilizing. Results from our validation dataset, below, which held out full sets of WT and variants, ensuring validation only on unseen enzymes and studies. #enzymeengineering #datascience #bioml #machinelearning #enzymes #proteinengineering #llms #biollm #techbio #ml

We're excited to be executing this pipeline for early clients: Tm improvements of up to 10ºC while maintaining activity with generative protein model finetuning. For the first time, we augment the training data with the top synthetically generated variants. Are the best candidates going to the lab? You betcha! Equally as exciting: we can now do this in as little as one week. #proteinengineering #compbio #bioai #bioml #genai #enzymeengineering #enzymes #synbio #techbio #biotech #llms #proteins

Ranking a generated protein with one LLM is not enough! Here we show perplexity - often used as a measure of 'naturalness' - versus ESM2 log probability (closer to zero is better). In a third dimension, you can see that proteins with the greatest predicted Tm span the same region as low Tm proteins. Meaning, both low and high Tm proteins are generally stable and natural - of course! It often takes dedicated models, like our Tm predictor, to extract the pertinent information that pretrained models are aware of, but not optimized for. #antibodyengineering #proteinengineering #qsar #llm #bioai #techbio

We're excited to preview natural-language interaction with bio-AI models! From ⌨️ to 🧬 ! Our agents return info from our API documentation, and are capable of making API requests themselves. This gives our GPT the power to generate new molecules using BioLM APIs. Here we say, "hi, give me some example proteins from all three progen2 models, starting with EVQL" and in a few seconds we see generated sequences from three models. You can chat with dozens of bio-LLMs in our BioLM Playground during the coming weeks. This feature is great toying around with new models before diving into more serious work with them. No need to spin up your own GPU, or futz with buggy code - trying out new bio-AI models should be easier! #bionlp #techbio #bioai #biochat #gpt #generativeai #biotech

At last year's SynBioBeta, co-founders Nikhil Haas and Ahmad Qamar presented the future of AI for SynBio Workflows - bio-chat agents, finetuning, and embeddings search for molecules. If you missed it, check out the interactive presentation and demos here: https://lnkd.in/gY8pt-Fd #synbiobeta #techbio #bionlp