
Founding Scientist @b12Labs (YC 25) | https://xitter.cf/t.co/HfFWBmmziQ Reimagining complex molecule synthesis ⚗️🤖💻 Prev PhD @EPFL | @gradcscunistra | @unirioja
Lausanne, Switzerland
Joined September 2012
- Tweets2.7K
- Following498
- Followers568
- Likes3.9K
Pinned Tweet
Making molecules is hard. How can we simplify the predicted synthesis routes of generated, property‑optimized small molecules? Our last work presents a framework that lets you "mix-and-match" multiple reaction constraints in the synthesis of your generated molecules.
Víctor Sabanza Gil retweeted
This is a great development. Now if only we had labs autonomous enough to execute these kind of plans and make the molecules 🤔
Replying to @claudeai
Planning and executing the reactions to make a complex molecule (a “route”) is one of the hardest jobs in organic chemistry.
From one image of each of four molecules, with no internet and with code as its only tool, Claude wrote a complete route in about half an hour.
Víctor Sabanza Gil retweeted
Very excited about this one: Today we're launching the b12 Foundry in partnership with the High-Throughput Experimentation Lab at the University of Zurich. AI can now design a molecule in seconds, but making it can still take a chemist several weeks. b12 Foundry closes that gap. b12-labs.com/foundry
Víctor Sabanza Gil retweeted
📢💃It takes two to TANGO! @JeffGuo & @pschwllr present TANGO, a reward function that augments molecular generative models to directly optimize for constrained synthesizability. nature.com/articles/s43588-0…
🔓rdcu.be/e6C69
Víctor Sabanza Gil retweeted
DISEÑO MOLECULAR
Nuestro antiguo alumno, Víctor Sabanza, ha impartido en @unirioja una conferencia sobre IA y Diseño Molecular.
@VictorSabanza estudió Química en la UR y desarrolla su Tesis Doctoral en la EPF de Lausana
@vaya_elementos @epfl_en @hebusto
iescosmegarcia.larioja.edu.e…
Víctor Sabanza Gil retweeted
🔊| Víctor Sabanza Gil @VictorSabanza, alumni del Grado en #QuímicaUR y fundador del grupo de divulgación @vaya_elementos regresa a la UR para hablar de inteligencia artificial y de diseño molecular
📆15Dic
🕐13 horas
📍CCT
🎫Entrada libre
#InstitutoQuímicaUR #IQUR #PolíticaCientíficaUR
➕Info👇
f.mtr.cool/rgybqznyle
Víctor Sabanza Gil retweeted
Out now! @pschwllr, @SchwallerGroup, @loic_roch, @VictorSabanza and colleagues provide guidelines and recommendations for when to use multi-fidelity Bayesian optimization over their single-fidelity counterparts
nature.com/articles/s43588-0…
🔓rdcu.be/exyUh
Using a generalist molecular generative model and reinforcement learning, we can address synthesizability multi-parameter optimization objectives without additional inductive biases.
Code: github.com/schwallergroup/sa…
Preprint: arxiv.org/abs/2505.08774
1
4
148
Coauthors: @JeffGuo__ , @ZJoncev, @JLuterbacher, @pschwllr
Thanks to @NCCR_Catalysis and @EPFL_ReO GlobaLeaders for the support!
3
169
Finally out! We have been working on a generative molecular design framework to allow steerable and granular control over synthetic routes for property-optimized molecules. ⚗️ 🖥️ Take a look below! 👇
Generate property-optimized small molecules with 𝘴𝘵𝘦𝘦𝘳𝘢𝘣𝘭𝘦 𝘢𝘯𝘥 𝘨𝘳𝘢𝘯𝘶𝘭𝘢𝘳 synthesizability control - allowing complete user-flexibility to impose various reaction constraints!
Pre-print: arxiv.org/abs/2505.08774
Code: github.com/schwallergroup/sa…
(1/4)
Víctor Sabanza Gil retweeted
Hello from Singapore 🇸🇬! Thrilled to be at #ICLR2025 presenting our work on fragment-based drug discovery 🧩. We go beyond virtual screening with a generative, structure-aware approach.
📃 openreview.net/forum?id=bZW1…
🔗 github.com/rneeser/LatentFra…
A thread 🧵👇
Víctor Sabanza Gil retweeted
Check out the updated published version of our pre-print in @ChemicalScience!
(1) *General-purpose* generative model + retrosynthesis model = design molecules with optimized properties with an *explicit* predicted synthesis pathway
Paper: pubs.rsc.org/en/content/arti…
(1/2)
Molecular generative models can *directly* optimize for synthesizability using retrosynthesis models!
Check out initial results which can be an alternative to synthesizability-constrained generation
Pre-print: arxiv.org/abs/2407.12186v1
Code: github.com/schwallergroup/sa…
(1/2)
Víctor Sabanza Gil retweeted
🎉Our BoLudo paper is in @J_A_C_S! Bayesian Optimization for nanocrystaL strUcture Design Optimization (jk, it's BOjana & LUDO😂)
We show you can optimize ANYTHING when put on the right scale - even nanocrystals with surface energy!
Also we discovered a new Cu shape! See how👇
A Holistic Data-Driven Approach to Synthesis Predictions of Colloidal Nanocrystal Shapes | Journal of the American Chemical Society @lnce_epfl @SchwallerGroup @EPFL_CHEM_Tweet @EPFL_en
@NCCR_Catalysis @pschwllr
@6ojaHa pubs.acs.org/doi/10.1021/jac…
Víctor Sabanza Gil retweeted
🧵 Excellent showing from @SchwallerGroup at #NeurIPS2024!
Our team received multiple spotlight presentations and acceptances across key AI+Science workshops, showcasing cutting-edge work in AI for chemistry and materials science.
Víctor Sabanza Gil retweeted
Atinary @ #NeurIPS in Vancouver this week🍁
Connect with our #AI #ML researchers @VictorSabanza & @shreyaspadhy.
Our research paper on Multi-Fidelity Bayesian Optimization (MFBO) will be @ AIDrugX workshop on Dec 15!
Full article: arxiv.org/abs/2410.00544
@AtinaryTech #SDLabs
Víctor Sabanza Gil retweeted
1/ Starting material constrained synthesis planning is now possible using a general retrosynthesis algorithm *without* training a dedicated value network!
Check-out our new paper, TangoStar.
Preprint : arxiv.org/abs/2412.03424
Víctor Sabanza Gil retweeted
📜 two very interesting, fundamental papers that shed light on when multi-fidelity (MF) Bayesian optimization (BayesOpt) is fruitful, compared to single-fidelity (SF) BayesOpt.
arxiv.org/pdf/2410.00544 (led by @loic_roch, @pschwllr)
arxiv.org/abs/2409.07190 (led by @JelfsChem)
particularly, the papers investigate the degree to which the low-fidelity experiment must:
- have a lower cost than the high-fidelity experiment
- give an outcome correlated with the outcome of the analogous high-fidelity experiment (for low-fidelity experiments to inform what we ultimately wish to optimize, the outcome of the high-fidelity experiment)
[fantastic ideas to use synthetic functions where both can be tuned and to hypothetically adjust the cost ratio in real data sets.]
MFBO was successful in our problem to search for COFs with high simulated Xe/Kr selectivity, thanks to the low cost ratio 0.065 and high correlation R^2=0.98.
in our paper in Digital Discovery, we [within a computer simulation] demonstrate multi-fidelity Bayesian optimization (MFBO) for reducing the cost of materials discovery and orchestrating "self-driving" labs.
pubs.rsc.org/en/content/arti…
🧪 problem setup
we
(1) wish to search a candidate set of materials for the one with the optimal property and
(2) have access to two experimental protocols to measure this property: one cheap but low-fidelity, the other high-fidelity but expensive.
what is the sequence of experiments (ie., material-fidelity pairs) that allows us to find the optimal material (according to the high-fidelity measurement) early in the search, incurring the least cost?
🔨 MFBO
multi-fidelity Bayesian optimization (MFBO) combines:
- supervised machine learning
- uncertainty quantification
- automated decision-making under uncertainty
to iteratively pick experiments (material, fidelity pairs) towards finding the optimal material while expending the least cost.
MFBO leverages the cheap, low-fidelity measurements to prune unpromising regions of materials space and focus the expensive, high-fidelity measurements on materials that lie in more promising regions.
MFBO constitutes an experiment-update-plan feedback loop.
1⃣ experiment phase: we conduct a low- or high-fidelity measurement of a material.
2⃣ update phase: we construct a surrogate model, trained on all data gathered thus far, that inexpensively predicts, with quantified uncertainty, both low- and high-fidelity measurements of material properties based on structural and chemical features of the materials.
3⃣ plan phase: MFBO picks the material and fidelity for the next experiment by maximizing an acquisition function that scores each material-fidelity pair according to its appeal for the next experiment, balancing exploration, exploitation, and cost.
🖥️ MFBO of COFs for Xe/Kr separations
we demonstrated MFBO by searching a database of ~600 covalent organic framework (COF) materials for the one with the highest adsorptive xenon/krypton selectivity, armed with (i) a high-fidelity binary grand-canonical Monte Carlo simulation of adsorption and (ii) a low-fidelity molecular simulation in the dilute regime to calculate the Henry coefficients of Xe and Kr in the COF. we found that MFBO acquires the optimal COF using only 30 low-fidelity and seven high-fidelity simulations, incurring only 2%, 4% on average, and 20% on average of the computational runtime of a single-[high-]fidelity exhaustive, random, and BO search, respectively.
🧠 the team
a close collaboration with Nick Gantzler, @deshwal_aryan, and @janadoppa.



























